Skip on CPU

slow slow slow
add tiny test_data
2020-07-22 12:19:28 -04:00 · 2020-07-22 12:00:12 -04:00 · 2020-07-22 11:35:08 -04:00
223 changed files with 5999 additions and 13787 deletions
@@ -10,8 +10,7 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
-            - run: sudo pip install .[sklearn,tf-cpu,torch,testing] --no-cache-dir
+            - run: sudo pip install .[sklearn,tf-cpu,torch,testing]
            - run: sudo pip install codecov pytest-cov
            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ --cov  | tee output.txt
            - run: codecov
@@ -28,8 +27,7 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
-            - run: sudo pip install .[sklearn,torch,testing] --no-cache-dir
+            - run: sudo pip install .[sklearn,torch,testing]
            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ | tee output.txt
            - store_artifacts:
                  path: ~/transformers/output.txt
@@ -45,8 +43,7 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
-            - run: sudo pip install .[sklearn,tf-cpu,testing] --no-cache-dir
+            - run: sudo pip install .[sklearn,tf-cpu,testing]
            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ | tee output.txt
            - store_artifacts:
               path: ~/transformers/output.txt
@@ -59,8 +56,7 @@ jobs:
            RUN_CUSTOM_TOKENIZERS: yes
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
-            - run: sudo pip install .[ja,testing]
+            - run: sudo pip install .[mecab,testing]
            - run: python -m pytest -s ./tests/test_tokenization_bert_japanese.py | tee output.txt
            - store_artifacts:
                path: ~/transformers/output.txt
@@ -75,10 +71,9 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
-            - run: sudo pip install .[sklearn,torch,testing] --no-cache-dir
+            - run: sudo pip install .[sklearn,torch,testing]
            - run: sudo pip install -r examples/requirements.txt
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./examples/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -s ./examples/ | tee output.txt
            - store_artifacts:
                  path: ~/transformers/output.txt
                  destination: test_output.txt
@@ -88,8 +83,7 @@ jobs:
            - image: circleci/python:3.6
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
-            - run: sudo pip install .[tf,torch,docs] --no-cache-dir
+            - run: sudo pip install .[tf,torch,docs]
            - run: cd docs && make html SPHINXOPTS="-W"
            - store_artifacts:
                path: ./docs/_build
@@ -102,7 +96,7 @@ jobs:
                fingerprints:
                    - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
            - checkout
-            - run: sudo pip install .[tf,torch,docs] --no-cache-dir
+            - run: sudo pip install .[tf,torch,docs]
            - run: ./.circleci/deploy.sh
    check_code_quality:
        working_directory: ~/transformers
@@ -112,10 +106,9 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install --upgrade pip
            # we need a version of isort with https://github.com/timothycrosley/isort/pull/1000
            - run: sudo pip install git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort
-            - run: sudo pip install .[tf,torch,quality] --no-cache-dir
+            - run: sudo pip install .[tf,torch,quality]
            - run: black --check --line-length 119 --target-version py35 examples templates tests src utils
            - run: isort --check-only --recursive examples templates tests src utils
            - run: flake8 examples templates tests src utils
@@ -47,4 +47,4 @@ deploy_doc "e7cfc1a" v2.9.0
 deploy_doc "7cb203f" v2.9.1
 deploy_doc "10d7239" v2.10.0 
 deploy_doc "b42586e" v2.11.0
-deploy_doc "7fb8bdf" #v3.0.2 Latest stable release
+deploy_doc "1158e56" #v3.0.2 Latest stable release
@@ -7,51 +7,14 @@ assignees: ''

 ---

-
-## Environment info
-<!-- You can run the command `transformers-cli env` and copy-and-paste its output below.
-     Don't forget to fill out the missing fields in that output! -->
-     
- `transformers` version:
- Platform:
- Python version:
- PyTorch version (GPU?):
- Tensorflow version (GPU?):
- Using GPU in script?:
- Using distributed or parallel set-up in script?:
-
-### Who can help
-<!-- Your issue will be replied to more quickly if you can figure out the right person to tag with @
- If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**.
- Please tag fewer than 3 people.
- 
- albert, bert, GPT2, XLM: @LysandreJik 
- tokenizers: @mfuntowicz
- Trainer: @sgugger
- Speed and Memory Benchmarks: @patrickvonplaten
- Model Cards: @julien-c
- Translation: @sshleifer
- Summarization: @sshleifer
- TextGeneration: @TevenLeScao 
- examples/distillation: @VictorSanh
- nlp datasets: [different repo](https://github.com/huggingface/nlp)
- rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
- Text Generation: @TevenLeScao
- blenderbot: @mariamabarham
- Bart: @sshleifer
- Marian: @sshleifer
- T5: @patrickvonplaten
- Longformer/Reformer: @patrickvonplaten
- TransfoXL/XLNet: @TevenLeScao 
- examples/seq2seq: @sshleifer
- tensorflow: @jplu 
-documentation: @sgugger
- -->
+# 🐛 Bug

 ## Information

 Model I am using (Bert, XLNet ...):

+Language I am using the model on (English, Chinese ...):
+
 The problem arises when using:
 * [ ] the official example scripts: (give details below)
 * [ ] my own modified scripts: (give details below)
@@ -75,3 +38,15 @@ Steps to reproduce the behavior:
 ## Expected behavior

 <!-- A clear and concise description of what you would expect to happen. -->
+
+## Environment info
+<!-- You can run the command `transformers-cli env` and copy-and-paste its output below.
+     Don't forget to fill out the missing fields in that output! -->
+     
+- `transformers` version:
+- Platform:
+- Python version:
+- PyTorch version (GPU?):
+- Tensorflow version (GPU?):
+- Using GPU in script?:
+- Using distributed or parallel set-up in script?:
@@ -24,4 +24,4 @@ assignees: ''

 <!-- You should first ask your question on the forum or SO, and only if
     you didn't get an answer ask it here on GitHub. -->
-**A link to original question on the forum/Stack Overflow**:
+**A link to original question on Stack Overflow**:
@@ -20,8 +20,7 @@ jobs:
        python-version: 3.7
    - name: Install dependencies
      run: |
-        pip install --upgrade pip
-        pip install torch --no-cache-dir
+        pip install torch
        pip install numpy tokenizers filelock requests tqdm regex sentencepiece sacremoses packaging

    - name: Torch hub list
@@ -35,8 +35,7 @@ jobs:
    - name: Install dependencies
      run: |
        source .env/bin/activate
-        pip install --upgrade pip
-        pip install torch --no-cache-dir
+        pip install torch
        pip install .[sklearn,testing]

    - name: Are GPUs recognized by our DL frameworks
@@ -31,8 +31,7 @@ jobs:
    - name: Install dependencies
      run: |
        source .env/bin/activate
-        pip install --upgrade pip
-        pip install .[sklearn,torch,testing] --no-cache-dir
+        pip install .[sklearn,torch,testing]

    - name: Are GPUs recognized by our DL frameworks
      run: |
@@ -215,7 +215,7 @@ Follow these steps to start contributing:
     `RUN_SLOW=1 python -m pytest tests/test_my_new_model.py`. 
   - If you are adding a new tokenizer, write tests, and make sure 
     `RUN_SLOW=1 python -m pytest tests/test_tokenization_{your_model_name}.py` passes.
-   CircleCI does not run the slow tests, but github actions does every night!
+   CircleCI does not run the slow tests. 
 6. All public methods must have informative docstrings that work nicely with sphinx. See `modeling_ctrl.py` for an 
   example.

@@ -239,16 +239,6 @@ $ pip install -r examples/requirements.txt  # only needed the first time
 $ python -m pytest -n auto --dist=loadfile -s -v ./examples/
 ```

-and for the slow tests:
-
-```bash
-RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/
-```
-or 
-```python
-RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/
-```
-
 In fact, that's how `make test` and `make test-examples` are implemented!

 You can specify a smaller set of tests in order to test only the feature
@@ -174,19 +174,12 @@ These implementations have been tested on several datasets (see the example scri

 ## Online demo

-You can test our inference API on most model pages from the model hub: https://huggingface.co/models
-
-For example: 
- [Masked word completion with BERT](https://huggingface.co/bert-base-uncased?text=Paris+is+the+%5BMASK%5D+of+France)
- [NER with Electra](https://huggingface.co/dbmdz/electra-large-discriminator-finetuned-conll03-english?text=My+name+is+Sarah+and+I+live+in+London+city)
- [Text generation with GPT-2](https://huggingface.co/gpt2?text=A+long+time+ago%2C+)
- [NLI with RoBERTa](https://huggingface.co/roberta-large-mnli?text=The+dog+was+lost.+Nobody+lost+any+animal)
- [Summarization with BART](https://huggingface.co/facebook/bart-large-cnn?text=The+tower+is+324+metres+%281%2C063+ft%29+tall%2C+about+the+same+height+as+an+81-storey+building%2C+and+the+tallest+structure+in+Paris.+Its+base+is+square%2C+measuring+125+metres+%28410+ft%29+on+each+side.+During+its+construction%2C+the+Eiffel+Tower+surpassed+the+Washington+Monument+to+become+the+tallest+man-made+structure+in+the+world%2C+a+title+it+held+for+41+years+until+the+Chrysler+Building+in+New+York+City+was+finished+in+1930.+It+was+the+first+structure+to+reach+a+height+of+300+metres.+Due+to+the+addition+of+a+broadcasting+aerial+at+the+top+of+the+tower+in+1957%2C+it+is+now+taller+than+the+Chrysler+Building+by+5.2+metres+%2817+ft%29.+Excluding+transmitters%2C+the+Eiffel+Tower+is+the+second+tallest+free-standing+structure+in+France+after+the+Millau+Viaduct)
- [Question answering with DistilBERT](https://huggingface.co/distilbert-base-uncased-distilled-squad?text=Which+name+is+also+used+to+describe+the+Amazon+rainforest+in+English%3F&context=The+Amazon+rainforest+%28Portuguese%3A+Floresta+Amaz%C3%B4nica+or+Amaz%C3%B4nia%3B+Spanish%3A+Selva+Amaz%C3%B3nica%2C+Amazon%C3%ADa+or+usually+Amazonia%3B+French%3A+For%C3%AAt+amazonienne%3B+Dutch%3A+Amazoneregenwoud%29%2C+also+known+in+English+as+Amazonia+or+the+Amazon+Jungle%2C+is+a+moist+broadleaf+forest+that+covers+most+of+the+Amazon+basin+of+South+America.+This+basin+encompasses+7%2C000%2C000+square+kilometres+%282%2C700%2C000+sq+mi%29%2C+of+which+5%2C500%2C000+square+kilometres+%282%2C100%2C000+sq+mi%29+are+covered+by+the+rainforest.+This+region+includes+territory+belonging+to+nine+nations.+The+majority+of+the+forest+is+contained+within+Brazil%2C+with+60%25+of+the+rainforest%2C+followed+by+Peru+with+13%25%2C+Colombia+with+10%25%2C+and+with+minor+amounts+in+Venezuela%2C+Ecuador%2C+Bolivia%2C+Guyana%2C+Suriname+and+French+Guiana.+States+or+departments+in+four+nations+contain+%22Amazonas%22+in+their+names.+The+Amazon+represents+over+half+of+the+planet%27s+remaining+rainforests%2C+and+comprises+the+largest+and+most+biodiverse+tract+of+tropical+rainforest+in+the+world%2C+with+an+estimated+390+billion+individual+trees+divided+into+16%2C000+species)
- [Translation with T5](https://huggingface.co/t5-base?text=My+name+is+Wolfgang+and+I+live+in+Berlin)
-
-
 **[Write With Transformer](https://transformer.huggingface.co)**, built by the Hugging Face team at transformer.huggingface.co, is the official demo of this repo’s text generation capabilities.
+You can use it to experiment with completions generated by `GPT2Model`, `TransfoXLModel`, and `XLNetModel`.
+
+> “🦄 Write with transformer is to writing what calculators are to calculus.”
+
+![write_with_transformer](https://transformer.huggingface.co/front/assets/thumbnail-large.png)

 ## Quick tour

@@ -0,0 +1,23 @@
+cd docs
+
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ] 
+	then
+		echo "Pushing version" $2
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+
+deploy_doc "master" 
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
+deploy_doc "f2f3294" v2.2.0
+deploy_doc "d0f8b9a" v2.3.0
@@ -157,8 +157,8 @@ conversion utilities for the following models:
    notebooks
    converting_tensorflow_models
    migration
+    torchscript
    contributing
-    serialization

 .. toctree::
    :maxdepth: 2
@@ -177,9 +177,9 @@ conversion utilities for the following models:
    main_classes/model
    main_classes/tokenizer
    main_classes/pipelines
-    main_classes/trainer
    main_classes/optimizer_schedules
    main_classes/processors
+    main_classes/trainer
    model_doc/auto
    model_doc/encoderdecoder
    model_doc/bert
@@ -205,6 +205,3 @@ conversion utilities for the following models:
    model_doc/retribert
    model_doc/mobilebert
    model_doc/dpr
-    internal/modeling_utils
-    internal/tokenization_utils
-    internal/pipelines_utils
@@ -1,88 +0,0 @@
-Custom Layers and Utilities
---------------------------
-
-This page lists all the custom layers used by the library, as well as the utility functions it provides for modeling.
-
-Most of those are only useful if you are studying the code of the models in the library.
-
-
-``Pytorch custom modules``
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_utils.Conv1D
-
-.. autoclass:: transformers.modeling_utils.PoolerStartLogits
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.PoolerEndLogits
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.PoolerAnswerClass
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.SquadHeadOutput
-
-.. autoclass:: transformers.modeling_utils.SQuADHead
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.SequenceSummary
-    :members: forward
-
-
-``PyTorch Helper Functions``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autofunction:: transformers.apply_chunking_to_forward
-
-.. autofunction:: transformers.modeling_utils.find_pruneable_heads_and_indices
-
-.. autofunction:: transformers.modeling_utils.prune_layer
-
-.. autofunction:: transformers.modeling_utils.prune_conv1d_layer
-
-.. autofunction:: transformers.modeling_utils.prune_linear_layer
-
-``TensorFlow custom layers``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_tf_utils.TFConv1D
-
-.. autoclass:: transformers.modeling_tf_utils.TFSharedEmbeddings
-    :members: call
-
-.. autoclass:: transformers.modeling_tf_utils.TFSequenceSummary
-    :members: call
-
-
-``TensorFlow loss functions``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_tf_utils.TFCausalLanguageModelingLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFMaskedLanguageModelingLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFMultipleChoiceLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFQuestionAnsweringLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFSequenceClassificationLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFTokenClassificationLoss
-    :members:
-
-
-``TensorFlow Helper Functions``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autofunction:: transformers.modeling_tf_utils.cast_bool_to_primitive
-
-.. autofunction:: transformers.modeling_tf_utils.get_initializer
-
-.. autofunction:: transformers.modeling_tf_utils.keras_serializable
-
-.. autofunction:: transformers.modeling_tf_utils.shape_list
@@ -1,40 +0,0 @@
-Utilities for pipelines
-----------------------
-
-This page lists all the utility functions the library provides for pipelines.
-
-Most of those are only useful if you are studying the code of the models in the library.
-
-
-Argument handling
-~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.pipelines.ArgumentHandler
-
-.. autoclass:: transformers.pipelines.ZeroShotClassificationArgumentHandler
-
-.. autoclass:: transformers.pipelines.QuestionAnsweringArgumentHandler
-
-
-Data format
-~~~~~~~~~~~
-
-.. autoclass:: transformers.pipelines.PipelineDataFormat
-    :members:
-
-.. autoclass:: transformers.pipelines.CsvPipelineDataFormat
-    :members:
-
-.. autoclass:: transformers.pipelines.JsonPipelineDataFormat
-    :members:
-
-.. autoclass:: transformers.pipelines.PipedPipelineDataFormat
-    :members:
-
-
-Utilities
-~~~~~~~~~
-
-.. autofunction:: transformers.pipelines.get_framework
-
-.. autoclass:: transformers.pipelines.PipelineException
@@ -1,38 +0,0 @@
-Utilities for Tokenizers
------------------------
-
-This page lists all the utility functions used by the tokenizers, mainly the class
-:class:`~transformers.tokenization_utils_base.PreTrainedTokenizerBase` that implements the common methods between
-:class:`~transformers.PreTrainedTokenizer` and :class:`~transformers.PreTrainedTokenizerFast` and the mixin
-:class:`~transformers.tokenization_utils_base.SpecialTokensMixin`.
-
-Most of those are only useful if you are studying the code of the tokenizers in the library.
-
-``PreTrainedTokenizerBase``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.tokenization_utils_base.PreTrainedTokenizerBase
-    :special-members: __call__
-    :members:
-
-
-``SpecialTokensMixin``
-~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.tokenization_utils_base.SpecialTokensMixin
-    :members:
-
-
-Enums and namedtuples
-~~~~~~~~~~~~~~~~~~~~~
-.. autoclass:: transformers.tokenization_utils_base.ExplicitEnum
-
-.. autoclass:: transformers.tokenization_utils_base.PaddingStrategy
-
-.. autoclass:: transformers.tokenization_utils_base.TensorType
-
-.. autoclass:: transformers.tokenization_utils_base.TruncationStrategy
-
-.. autoclass:: transformers.tokenization_utils_base.CharSpan
-
-.. autoclass:: transformers.tokenization_utils_base.TokenSpan
@@ -1,32 +1,23 @@
 Models
 ----------------------------------------------------

-The base classes :class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` implement the
-common methods for loading/saving a model either from a local file or directory, or from a pretrained model
-configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).
+The base class ``PreTrainedModel`` implements the common methods for loading/saving a model either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).

-:class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` also implement a few methods which
-are common among all the models to:
+``PreTrainedModel`` also implements a few methods which are common among all the models to:

 - resize the input token embeddings when new tokens are added to the vocabulary
 - prune the attention heads of the model.

-The other methods that are common to each model are defined in :class:`~transformers.modeling_utils.ModuleUtilsMixin`
-(for the PyTorch models) and :class:`~transformers.modeling_tf_utils.TFModuleUtilsMixin` (for the TensorFlow models).
-
-
 ``PreTrainedModel``
 ~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.PreTrainedModel
    :members:

+``Helper Functions``
+~~~~~~~~~~~~~~~~~~~~~

-``ModuleUtilsMixin``
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_utils.ModuleUtilsMixin
-    :members:
+.. autofunction:: transformers.apply_chunking_to_forward


 ``TFPreTrainedModel``
@@ -34,16 +25,3 @@ The other methods that are common to each model are defined in :class:`~transfor

 .. autoclass:: transformers.TFPreTrainedModel
    :members:
-
-
-``TFModelUtilsMixin``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_tf_utils.TFModelUtilsMixin
-    :members:
-
-
-Generative models
-~~~~~~~~~~~~~~~~~
-
-Coming soon
@@ -3,24 +3,13 @@ Pipelines

 The pipelines are a great and easy way to use models for inference. These pipelines are objects that abstract most
 of the complex code from the library, offering a simple API dedicated to several tasks, including Named Entity
-Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering. See the
-:doc:`task summary <../task_summary>` for examples of use.
+Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering.

 There are two categories of pipeline abstractions to be aware about:

- The :func:`~transformers.pipeline` which is the most powerful object encapsulating all other pipelines.
- The other task-specific pipelines:
-
-    - :class:`~transformers.ConversationalPipeline`
-    - :class:`~transformers.FeatureExtractionPipeline`
-    - :class:`~transformers.FillMaskPipeline`
-    - :class:`~transformers.QuestionAnsweringPipeline`
-    - :class:`~transformers.SummarizationPipeline`
-    - :class:`~transformers.TextClassificationPipeline`
-    - :class:`~transformers.TextGenerationPipeline`
-    - :class:`~transformers.TokenClassificationPipeline`
-    - :class:`~transformers.TranslationPipeline`
-    - :class:`~transformers.ZeroShotClassificationPipeline`
+- The :func:`~transformers.pipeline` which is the most powerful object encapsulating all other pipelines
+- The other task-specific pipelines, such as :class:`~transformers.TokenClassificationPipeline`
+  or :class:`~transformers.QuestionAnsweringPipeline`

 The pipeline abstraction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -32,82 +21,53 @@ other pipeline but requires an additional argument which is the `task`.


 The task specific pipelines
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-ConversationalPipeline
-==========================================
+Parent class: Pipeline
+=========================================

-.. autoclass:: transformers.Conversation
-
-.. autoclass:: transformers.ConversationalPipeline
-    :special-members: __call__
-    :members:
-
-FeatureExtractionPipeline
-==========================================
-
-.. autoclass:: transformers.FeatureExtractionPipeline
-    :special-members: __call__
-    :members:
-
-FillMaskPipeline
-==========================================
-
-.. autoclass:: transformers.FillMaskPipeline
-    :special-members: __call__
-    :members:
-
-NerPipeline
-==========================================
-
-This class is an alias of the :class:`~transformers.TokenClassificationPipeline` defined below. Please refer to that
-pipeline for documentation and usage examples.
-
-QuestionAnsweringPipeline
-==========================================
-
-.. autoclass:: transformers.QuestionAnsweringPipeline
-    :special-members: __call__
-    :members:
-
-SummarizationPipeline
-==========================================
-
-.. autoclass:: transformers.SummarizationPipeline
-    :special-members: __call__
-    :members:
-
-TextClassificationPipeline
-==========================================
-
-.. autoclass:: transformers.TextClassificationPipeline
-    :special-members: __call__
-    :members:
-
-TextGenerationPipeline
-==========================================
-
-.. autoclass:: transformers.TextGenerationPipeline
-    :special-members: __call__
-    :members:
+.. autoclass:: transformers.Pipeline
+    :members: predict, transform, save_pretrained

 TokenClassificationPipeline
 ==========================================

 .. autoclass:: transformers.TokenClassificationPipeline
-    :special-members: __call__
-    :members:

-ZeroShotClassificationPipeline
+NerPipeline
 ==========================================

-.. autoclass:: transformers.ZeroShotClassificationPipeline
-    :special-members: __call__
-    :members:
+This class is an alias of the :class:`~transformers.TokenClassificationPipeline` defined above. Please refer to that pipeline for
+documentation and usage examples.
+
+FillMaskPipeline
+==========================================
+
+.. autoclass:: transformers.FillMaskPipeline
+
+FeatureExtractionPipeline
+==========================================
+
+.. autoclass:: transformers.FeatureExtractionPipeline
+
+TextClassificationPipeline
+==========================================
+
+.. autoclass:: transformers.TextClassificationPipeline
+
+QuestionAnsweringPipeline
+==========================================
+
+.. autoclass:: transformers.QuestionAnsweringPipeline


-Parent class: :obj:`Pipeline`
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+SummarizationPipeline
+==========================================

-.. autoclass:: transformers.Pipeline
-    :members:
+.. autoclass:: transformers.SummarizationPipeline
+
+
+TextGenerationPipeline
+==========================================
+
+.. autoclass:: transformers.TextGenerationPipeline
@@ -1,40 +1,17 @@
 Tokenizer
 ----------------------------------------------------

-A tokenizer is in charge of preparing the inputs for a model. The library contains tokenizers for all the models. Most
-of the tokenizers are available in two flavors: a full python implementation and a "Fast" implementation based on the
-Rust library `tokenizers <https://github.com/huggingface/tokenizers>`__. The "Fast" implementations allows:
+A tokenizer is in charge of preparing the inputs for a model. The library comprise tokenizers for all the models. Most of the tokenizers are available in two flavors: a full python implementation and a "Fast" implementation based on the Rust library `tokenizers`. The "Fast" implementations allows (1) a significant speed-up in particular when doing batched tokenization and (2) additional methods to map between the original string (character and words) and the token space (e.g. getting the index of the token comprising a given character or the span of characters corresponding to a given token). Currently no "Fast" implementation is available for the SentencePiece-based tokenizers (for T5, ALBERT, CamemBERT, XLMRoBERTa and XLNet models).

-1. a significant speed-up in particular when doing batched tokenization and
-2. additional methods to map between the original string (character and words) and the token space (e.g. getting the
-   index of the token comprising a given character or the span of characters corresponding to a given token). Currently
-   no "Fast" implementation is available for the SentencePiece-based tokenizers (for T5, ALBERT, CamemBERT, XLMRoBERTa
-   and XLNet models).
+The base classes ``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` implements the common methods for encoding string inputs in model inputs (see below) and instantiating/saving python and "Fast" tokenizers either from a local file or directory or from a pretrained tokenizer provided by the library (downloaded from HuggingFace's AWS S3 repository).

-The base classes :class:`~transformers.PreTrainedTokenizer` and :class:`~transformers.PreTrainedTokenizerFast`
-implement the common methods for encoding string inputs in model inputs (see below) and instantiating/saving python and
-"Fast" tokenizers either from a local file or directory or from a pretrained tokenizer provided by the library
-(downloaded from HuggingFace's AWS S3 repository). They both rely on
-:class:`~transformers.tokenization_utils_base.PreTrainedTokenizerBase` that contains the common methods, and
-:class:`~transformers.tokenization_utils_base.SpecialTokensMixin`.
+``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` thus implements the main methods for using all the tokenizers:

-:class:`~transformers.PreTrainedTokenizer` and :class:`~transformers.PreTrainedTokenizerFast` thus implement the main
-methods for using all the tokenizers:
-
- Tokenizing (splitting strings in sub-word token strings), converting tokens strings to ids and back, and
-  encoding/decoding (i.e., tokenizing and converting to integers).
- Adding new tokens to the vocabulary in a way that is independent of the underlying structure (BPE, SentencePiece...).
- Managing special tokens (like mask, beginning-of-sentence, etc.): adding them, assigning them to attributes in the
-  tokenizer for easy access and making sure they are not split during tokenization.
-
-:class:`~transformers.BatchEncoding` holds the output of the tokenizer's encoding methods (``__call__``,
-``encode_plus`` and ``batch_encode_plus``) and is derived from a Python dictionary. When the tokenizer is a pure python
-tokenizer, this class behaves just like a standard python dictionary and holds the various model inputs computed by these
-methods (``input_ids``, ``attention_mask``...). When the tokenizer is a "Fast" tokenizer (i.e., backed by HuggingFace
-`tokenizers library <https://github.com/huggingface/tokenizers>`__), this class provides in addition several advanced
-alignment methods which can be used to map between the original string (character and words) and the token space (e.g.,
-getting the index of the token comprising a given character or the span of characters corresponding to a given token).
+- tokenizing (spliting strings in sub-word token strings), converting tokens strings to ids and back, and encoding/decoding (i.e. tokenizing + convert to integers),
+- adding new tokens to the vocabulary in a way that is independant of the underlying structure (BPE, SentencePiece...),
+- managing special tokens like mask, beginning-of-sentence, etc tokens (adding them, assigning them to attributes in the tokenizer for easy access and making sure they are not split during tokenization)

+``BatchEncoding`` holds the output of the tokenizer's encoding methods (``__call__``, ``encode_plus`` and ``batch_encode_plus``) and is derived from a Python dictionary. When the tokenizer is a pure python tokenizer, this class behave just like a standard python dictionary and hold the various model inputs computed by these methodes (``input_ids``, ``attention_mask``...). When the tokenizer is a "Fast" tokenizer (i.e. backed by HuggingFace tokenizers library), this class provides in addition several advanced alignement methods which can be used to map between the original string (character and words) and the token space (e.g. getting the index of the token comprising a given character or the span of characters corresponding to a given token).

 ``PreTrainedTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~
@@ -43,7 +20,6 @@ getting the index of the token comprising a given character or the span of chara
    :special-members: __call__
    :members:

-
 ``PreTrainedTokenizerFast``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -51,9 +27,14 @@ getting the index of the token comprising a given character or the span of chara
    :special-members: __call__
    :members:

-
 ``BatchEncoding``
 ~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.BatchEncoding
    :members:
+
+``SpecialTokensMixin``
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SpecialTokensMixin
+    :members:
@@ -11,23 +11,6 @@ customization during training.
 The API supports distributed training on multiple GPUs/TPUs, mixed precision through `NVIDIA Apex
 <https://github.com/NVIDIA/apex>`__ for PyTorch and :obj:`tf.keras.mixed_precision` for TensorFlow.

-Both :class:`~transformers.Trainer` and :class:`~transformers.TFTrainer` contain the basic training loop supporting the
-previous features. To inject custom behavior you can subclass them and override the following methods:
-
- **get_train_dataloader**/**get_train_tfdataset** -- Creates the training DataLoader (PyTorch) or TF Dataset.
- **get_eval_dataloader**/**get_eval_tfdataset** -- Creates the evaulation DataLoader (PyTorch) or TF Dataset.
- **get_test_dataloader**/**get_test_tfdataset** -- Creates the test DataLoader (PyTorch) or TF Dataset.
- **log** -- Logs information on the various objects watching training.
- **setup_wandb** -- Setups wandb (see `here <https://docs.wandb.com/huggingface>`__ for more information).
- **create_optimizer_and_scheduler** -- Setups the optimizer and learning rate scheduler if they were not passed at
-  init.
- **training_step** -- Performs a training step.
- **prediction_step** -- Performs an evaluation/test step.
- **run_model** (TensorFlow only) -- Basic pass through the model.
- **evaluate** -- Runs an evaluation loop and returns metrics.
- **predict** -- Returns predictions (with metrics if labels are available) on a test set.
-
-
 ``Trainer`` 
 ~~~~~~~~~~~

@@ -27,8 +27,13 @@ Tips:

 - BERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
  the right rather than the left.
- BERT was trained with the masked language modeling (MLM) and next sentence prediction (NSP) objectives. It is efficient at predicting masked
-  tokens and at NLU in general, but is not optimal for text generation.
+- BERT was trained with a masked language modeling (MLM) objective. It is therefore efficient at predicting masked
+  tokens and at NLU in general, but is not optimal for text generation. Models trained with a causal language
+  modeling (CLM) objective are better in that regard.
+- Alongside MLM, BERT was trained using a next sentence prediction (NSP) objective using the [CLS] token as a sequence
+  approximate. The user may use this token (the first token in a sequence built with special tokens) to get a sequence
+  prediction rather than a token prediction. However, averaging over the sequence may yield better results than using
+  the [CLS] token.

 The original code can be found `here <https://github.com/google-research/bert>`_.

@@ -16,7 +16,7 @@ Longformer Self Attention
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 Longformer self attention employs self attention on both a "local" context and a "global" context.
 Most tokens only attend "locally" to each other meaning that each token attends to its :math:`\frac{1}{2} w` previous tokens and :math:`\frac{1}{2} w` succeding tokens with :math:`w` being the window length as defined in `config.attention_window`. Note that `config.attention_window` can be of type ``list`` to define a different :math:`w` for each layer. 
-A selected few tokens attend "globally" to all other tokens, as it is conventionally done for all tokens in *e.g.* `BertSelfAttention`.
+A selecetd few tokens attend "globally" to all other tokens, as it is conventionally done for all tokens in *e.g.* `BertSelfAttention`.

 Note that "locally" and "globally" attending tokens are projected by different query, key and value matrices.
 Also note that every "locally" attending token not only attends to tokens within its window :math:`w`, but also to all "globally" attending tokens so that global attention is *symmetric*.
@@ -102,25 +102,3 @@ LongformerForQuestionAnswering

 .. autoclass:: transformers.LongformerForQuestionAnswering
    :members:
-
-
-TFLongformerModel
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFLongformerModel
-    :members:
-
-
-TFLongformerForMaskedLM
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFLongformerForMaskedLM
-    :members:
-
-
-TFLongformerForQuestionAnswering
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFLongformerForQuestionAnswering
-    :members:
-
@@ -38,13 +38,13 @@ T5 can be trained / fine-tuned both in a supervised and unsupervised fashion.

  In this setup spans of the input sequence are masked by so-called sentinel tokens (*a.k.a* unique mask tokens) 
  and the output sequence is formed as a concatenation of the same sentinel tokens and the *real* masked tokens. 
-  Each sentinel token represents a unique mask token for this sentence and should start with ``<extra_id_0>``, ``<extra_id_1>``, ... up to ``<extra_id_99>``. As a default 100 sentinel tokens are available in ``T5Tokenizer``.
+  Each sentinel token represents a unique mask token for this sentence and should start with ``<extra_id_1>``, ``<extra_id_2>``, ... up to ``<extra_id_100>``. As a default 100 sentinel tokens are available in ``T5Tokenizer``.
  *E.g.* the sentence "The cute dog walks in the park" with the masks put on "cute dog" and "the" should be processed as follows: 

 ::

-  input_ids = tokenizer.encode('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt')
-  labels = tokenizer.encode('<extra_id_0> cute dog <extra_id_1> the <extra_id_2> </s>', return_tensors='pt')
+  input_ids = tokenizer.encode('The <extra_id_1> walks in <extra_id_2> park', return_tensors='pt')
+  labels = tokenizer.encode('<extra_id_1> cute dog <extra_id_2> the <extra_id_3> </s>', return_tensors='pt')
  # the forward function automatically creates the correct decoder_input_ids
  model(input_ids=input_ids, labels=labels)

@@ -195,7 +195,7 @@ tokens in the sentence, then allows the model to use the last n tokens to predic
 with a mask, the sentence is actually fed in the model in the right order, but instead of masking the first n tokens 
 for n+1, XLNet uses a mask that hides the previous tokens in some given permutation of 1,...,sequence length.

-XLNet also uses the same recurrence mechanism as Transformer-XL to build long-term dependencies. 
+XLNet also uses the same recurrence mechanism as TransformerXL to build long-term dependencies. 

 The library provides a version of the model for language modeling, token classification, sentence classification, 
 multiple choice classification and question answering.
@@ -74,16 +74,14 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``cl-tohoku/bert-base-japanese``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
-|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
-|                   |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
-|                   |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece.                                                               |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``cl-tohoku/bert-base-japanese-whole-word-masking``        | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
-|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
-|                   |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
-|                   |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
+|                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece.                                      |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -230,16 +230,19 @@ final activations of the model.

    >>> ## PYTORCH CODE
    >>> print(pt_outputs)
-    (tensor([[-4.0833,  4.3364],
-            [ 0.0818, -0.0418]], grad_fn=<AddmmBackward>),)
+    SequenceClassifierOutput(loss=None, logits=tensor([[-4.0833,  4.3364],
+            [ 0.0818, -0.0418]], grad_fn=<AddmmBackward>), hidden_states=None, attentions=None)
    >>> ## TENSORFLOW CODE
    >>> print(tf_outputs)
    (<tf.Tensor: shape=(2, 2), dtype=float32, numpy=
    array([[-4.0832963 ,  4.336414  ],
           [ 0.08181786, -0.04179301]], dtype=float32)>,)

-The model can return more than just the final activations, which is why the output is a tuple. Here we only asked for
-the final activations, so we get a tuple with one element.
+The model can return more than just the final activations, which is why the PyTorch output is a special class and the
+TensorFlow output is a tuple. Here we only asked for the final activations, so we get a tuple with one element on the
+TensorFlow side and a :class:`~transformers.modeling_outputs.SequenceClassifierOutput` with just the ``logits`` field
+filled on the PyTorch side.
+
 .. note::

    All 🤗 Transformers models (PyTorch or TensorFlow) return the activations of the model *before* the final
@@ -251,7 +254,7 @@ Let's apply the SoftMax activation to get predictions.

    >>> ## PYTORCH CODE
    >>> import torch.nn.functional as F
-    >>> pt_predictions = F.softmax(pt_outputs[0], dim=-1)
+    >>> pt_predictions = F.softmax(pt_outputs.logits, dim=-1)
    >>> ## TENSORFLOW CODE
    >>> import tensorflow as tf
    >>> tf_predictions = tf.nn.softmax(tf_outputs[0], axis=-1)
@@ -338,8 +341,8 @@ code is easy to access and tweak if you need to.

 In our previous example, the model was called "distilbert-base-uncased-finetuned-sst-2-english", which means it's
 using the :doc:`DistilBERT </model_doc/distilbert>` architecture. As
-:class:`~transformers.AutoModelForSequenceClassification` (or :class:`~transformers.TFAutoModelForSequenceClassification`
-if you are using TensorFlow) was used, the model automatically created is then a
+:class:`~transformers.AutoModelForSequenceClassification` (or  :class:`~transformers.TFAutoModelForSequenceClassification`
+if you are using TensorFlow)` was used, the model automatically created is then a
 :class:`~transformers.DistilBertForSequenceClassification`. You can look at its documentation for all details relevant
 to that specific model, or browse the source code. This is how you would directly instantiate model and tokenizer
 without the auto magic:
@@ -1,252 +0,0 @@
-**********************************************
-Exporting transformers models
-**********************************************
-
-ONNX / ONNXRuntime
-==============================================
-
-Projects `ONNX (Open Neural Network eXchange) <http://onnx.ai>`_ and `ONNXRuntime (ORT) <https://microsoft.github.io/onnxruntime/>`_ are part of an effort from leading industries in the AI field
-to provide a unified and community-driven format to store and, by extension, efficiently execute neural network leveraging a variety
-of hardware and dedicated optimizations.
-
-Starting from transformers v2.10.0 we partnered with ONNX Runtime to provide an easy export of transformers models to
-the ONNX format. You can have a look at the effort by looking at our joint blog post `Accelerate your NLP pipelines using
-Hugging Face Transformers and ONNX Runtime <https://medium.com/microsoftazure/accelerate-your-nlp-pipelines-using-hugging-face-transformers-and-onnx-runtime-2443578f4333>`_.
-
-Exporting a model is done through the script `convert_graph_to_onnx.py` at the root of the transformers sources.
-The following command shows how easy it is to export a BERT model from the library, simply run:
-
-.. code-block:: bash
-
-    python convert_graph_to_onnx.py --framework <pt, tf> --model bert-base-cased bert-base-cased.onnx
-
-The conversion tool works for both PyTorch and Tensorflow models and ensures:
-
-* The model and its weights are correctly initialized from the Hugging Face model hub or a local checkpoint.
-* The inputs and outputs are correctly generated to their ONNX counterpart.
-* The generated model can be correctly loaded through onnxruntime.
-
-.. note::
-    Currently, inputs and outputs are always exported with dynamic sequence axes preventing some optimizations
-    on the ONNX Runtime. If you would like to see such support for fixed-length inputs/outputs, please
-    open up an issue on transformers.
-
-
-Also, the conversion tool supports different options which let you tune the behavior of the generated model:
-
-* **Change the target opset version of the generated model.**  (More recent opset generally supports more operators and enables faster inference)
-
-* **Export pipeline-specific prediction heads.**  (Allow to export model along with its task-specific prediction head(s))
-
-* **Use the external data format (PyTorch only).**  (Lets you export model which size is above 2Gb (`More info <https://github.com/pytorch/pytorch/pull/33062>`_))
-
-
-Optimizations
------------------------------------------------
-
-ONNXRuntime includes some transformers-specific transformations to leverage optimized operations in the graph.
-Below are some of the operators which can be enabled to speed up inference through ONNXRuntime (*see note below*):
-
-* Constant folding
-* Attention Layer fusing
-* Skip connection LayerNormalization fusing
-* FastGeLU approximation
-
-
-Fortunately, you can let ONNXRuntime find all the possible optimized operators for you. Simply add ``--optimize``
-when exporting your model through ``convert_graph_to_onnx.py``.
-
-Example:
-
-.. code-block:: bash
-
-    python convert_graph_to_onnx.py --framework <pt, tf> --model bert-base-cased --optimize bert-base-cased.onnx
-
-.. note::
-    For more information about the optimizations enabled by ONNXRuntime, please have a look at the (`ONNXRuntime Github <https://github.com/microsoft/onnxruntime/tree/master/onnxruntime/python/tools/transformers>`_)
-
-Quantization
------------------------------------------------
-
-ONNX exporter supports generating a quantized version of the model to allow efficient inference.
-
-Quantization works by converting the memory representation of the parameters in the neural network
-to a compact integer format. By default, weights of a neural network are stored as single-precision float (`float32`)
-which can express a wide-range of floating-point numbers with decent precision.
-These properties are especially interesting at training where you want fine-grained representation.
-
-On the other hand, after the training phase, it has been shown one can greatly reduce the range and the precision of `float32` numbers
-without changing the performances of the neural network.
-
-More technically, `float32` parameters are converted to a type requiring fewer bits to represent each number, thus reducing
-the overall size of the model. Here, we are enabling `float32` mapping to `int8` values (a non-floating, single byte, number representation)
-according to the following formula:
-
-.. math::
-    y_{float32} = scale * x_{int8} - zero\_point
-
-.. note::
-    The quantization process will infer the parameter `scale` and `zero_point` from the neural network parameters
-
-Leveraging tiny-integers has numerous advantages when it comes to inference:
-
-* Storing fewer bits instead of 32 bits for the `float32` reduces the size of the model and makes it load faster.
-* Integer operations execute a magnitude faster on modern hardware
-* Integer operations require less power to do the computations
-
-In order to convert a transformers model to ONNX IR with quantized weights you just need to specify ``--quantize``
-when using ``convert_graph_to_onnx.py``. Also, you can have a look at the ``quantize()`` utility-method in this
-same script file.
-
-Example of quantized BERT model export:
-
-.. code-block:: bash
-
-    python convert_graph_to_onnx.py --framework <pt, tf> --model bert-base-cased --quantize bert-base-cased.onnx
-
-.. note::
-    Quantization support requires ONNX Runtime >= 1.4.0
-
-.. note::
-    When exporting quantized model you will end up with two different ONNX files. The one specified at the end of the
-    above command will contain the original ONNX model storing `float32` weights.
-    The second one, with ``-quantized`` suffix, will hold the quantized parameters.
-
-.. note::
-    The quantization export gives the best performances when used in combination with ``--optimize``.
-
-TorchScript
-=======================================
-
-.. note::
-    This is the very beginning of our experiments with TorchScript and we are still exploring its capabilities
-    with variable-input-size models. It is a focus of interest to us and we will deepen our analysis in upcoming
-    releases, with more code examples, a more flexible implementation, and benchmarks comparing python-based codes
-    with compiled TorchScript.
-
-
-According to Pytorch's documentation: "TorchScript is a way to create serializable and optimizable models from PyTorch code".
-Pytorch's two modules `JIT and TRACE <https://pytorch.org/docs/stable/jit.html>`_ allow the developer to export
-their model to be re-used in other programs, such as efficiency-oriented C++ programs.
-
-We have provided an interface that allows the export of 🤗 Transformers models to TorchScript so that they can
-be reused in a different environment than a Pytorch-based python program. Here we explain how to use our models so that
-they can be exported, and what to be mindful of when using these models with TorchScript.
-
-Exporting a model needs two things:
-
-* dummy inputs to execute a model forward pass.
-* the model needs to be instantiated with the ``torchscript`` flag.
-
-These necessities imply several things developers should be careful about. These are detailed below.
-
-
-Implications
------------------------------------------------
-
-TorchScript flag and tied weights
------------------------------------------------
-This flag is necessary because most of the language models in this repository have tied weights between their
-``Embedding`` layer and their ``Decoding`` layer. TorchScript does not allow the export of models that have tied weights,
-it is therefore necessary to untie the weights beforehand.
-
-This implies that models instantiated with the ``torchscript`` flag have their ``Embedding`` layer and ``Decoding`` layer
-separate, which means that they should not be trained down the line. Training would de-synchronize the two layers,
-leading to unexpected results.
-
-This is not the case for models that do not have a Language Model head, as those do not have tied weights. These models
-can be safely exported without the ``torchscript`` flag.
-
-Dummy inputs and standard lengths
------------------------------------------------
-
-The dummy inputs are used to do a model forward pass. While the inputs' values are propagating through the layers,
-Pytorch keeps track of the different operations executed on each tensor. These recorded operations are then used
-to create the "trace" of the model.
-
-The trace is created relatively to the inputs' dimensions. It is therefore constrained by the dimensions of the dummy
-input, and will not work for any other sequence length or batch size. When trying with a different size, an error such
-as:
-
-``The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2``
-
-will be raised. It is therefore recommended to trace the model with a dummy input size at least as large as the largest
-input that will be fed to the model during inference. Padding can be performed to fill the missing values. As the model
-will have been traced with a large input size however, the dimensions of the different matrix will be large as well,
-resulting in more calculations.
-
-It is recommended to be careful of the total number of operations done on each input and to follow performance closely
-when exporting varying sequence-length models.
-
-Using TorchScript in Python
-------------------------------------------------
-
-Below are examples of using the Python to save, load models as well as how to use the trace for inference.
-
-Saving a model
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-This snippet shows how to use TorchScript to export a ``BertModel``. Here the ``BertModel`` is instantiated
-according to a ``BertConfig`` class and then saved to disk under the filename ``traced_bert.pt``
-
-.. code-block:: python
-
-    from transformers import BertModel, BertTokenizer, BertConfig
-    import torch
-
-    enc = BertTokenizer.from_pretrained("bert-base-uncased")
-
-    # Tokenizing input text
-    text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
-    tokenized_text = enc.tokenize(text)
-
-    # Masking one of the input tokens
-    masked_index = 8
-    tokenized_text[masked_index] = '[MASK]'
-    indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
-    segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
-
-    # Creating a dummy input
-    tokens_tensor = torch.tensor([indexed_tokens])
-    segments_tensors = torch.tensor([segments_ids])
-    dummy_input = [tokens_tensor, segments_tensors]
-
-    # Initializing the model with the torchscript flag
-    # Flag set to True even though it is not necessary as this model does not have an LM Head.
-    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
-        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
-
-    # Instantiating the model
-    model = BertModel(config)
-
-    # The model needs to be in evaluation mode
-    model.eval()
-
-    # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
-    model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
-
-    # Creating the trace
-    traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
-    torch.jit.save(traced_model, "traced_bert.pt")
-
-Loading a model
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-This snippet shows how to load the ``BertModel`` that was previously saved to disk under the name ``traced_bert.pt``.
-We are re-using the previously initialised ``dummy_input``.
-
-.. code-block:: python
-
-    loaded_model = torch.jit.load("traced_model.pt")
-    loaded_model.eval()
-
-    all_encoder_layers, pooled_output = loaded_model(dummy_input)
-
-Using a traced model for inference
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Using the traced model for inference is as simple as using its ``__call__`` dunder method:
-
-.. code-block:: python
-
-    traced_model(tokens_tensor, segments_tensors)
@@ -0,0 +1,135 @@
+TorchScript
+================================================
+
+.. note::
+    This is the very beginning of our experiments with TorchScript and we are still exploring its capabilities
+    with variable-input-size models. It is a focus of interest to us and we will deepen our analysis in upcoming
+    releases, with more code examples, a more flexible implementation, and benchmarks comparing python-based codes
+    with compiled TorchScript.
+
+
+According to Pytorch's documentation: "TorchScript is a way to create serializable and optimizable models from PyTorch code".
+Pytorch's two modules `JIT and TRACE <https://pytorch.org/docs/stable/jit.html>`_ allow the developer to export
+their model to be re-used in other programs, such as efficiency-oriented C++ programs.
+
+We have provided an interface that allows the export of 🤗 Transformers models to TorchScript so that they can
+be reused in a different environment than a Pytorch-based python program. Here we explain how to use our models so that
+they can be exported, and what to be mindful of when using these models with TorchScript.
+
+Exporting a model needs two things:
+
+* dummy inputs to execute a model forward pass.
+* the model needs to be instantiated with the ``torchscript`` flag.
+
+These necessities imply several things developers should be careful about. These are detailed below.
+
+
+Implications
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+TorchScript flag and tied weights
+------------------------------------------------
+This flag is necessary because most of the language models in this repository have tied weights between their
+``Embedding`` layer and their ``Decoding`` layer. TorchScript does not allow the export of models that have tied weights,
+it is therefore necessary to untie the weights beforehand.
+
+This implies that models instantiated with the ``torchscript`` flag have their ``Embedding`` layer and ``Decoding`` layer
+separate, which means that they should not be trained down the line. Training would de-synchronize the two layers,
+leading to unexpected results.
+
+This is not the case for models that do not have a Language Model head, as those do not have tied weights. These models
+can be safely exported without the ``torchscript`` flag.
+
+Dummy inputs and standard lengths
+------------------------------------------------
+
+The dummy inputs are used to do a model forward pass. While the inputs' values are propagating through the layers,
+Pytorch keeps track of the different operations executed on each tensor. These recorded operations are then used
+to create the "trace" of the model.
+
+The trace is created relatively to the inputs' dimensions. It is therefore constrained by the dimensions of the dummy
+input, and will not work for any other sequence length or batch size. When trying with a different size, an error such
+as:
+
+``The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2``
+
+will be raised. It is therefore recommended to trace the model with a dummy input size at least as large as the largest
+input that will be fed to the model during inference. Padding can be performed to fill the missing values. As the model
+will have been traced with a large input size however, the dimensions of the different matrix will be large as well,
+resulting in more calculations.
+
+It is recommended to be careful of the total number of operations done on each input and to follow performance closely
+when exporting varying sequence-length models.
+
+Using TorchScript in Python
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Below are examples of using the Python to save, load models as well as how to use the trace for inference.
+
+Saving a model
+------------------------------------------------
+
+This snippet shows how to use TorchScript to export a ``BertModel``. Here the ``BertModel`` is instantiated
+according to a ``BertConfig`` class and then saved to disk under the filename ``traced_bert.pt``
+
+.. code-block:: python
+
+    from transformers import BertModel, BertTokenizer, BertConfig
+    import torch
+
+    enc = BertTokenizer.from_pretrained("bert-base-uncased")
+
+    # Tokenizing input text
+    text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+    tokenized_text = enc.tokenize(text)
+
+    # Masking one of the input tokens
+    masked_index = 8
+    tokenized_text[masked_index] = '[MASK]'
+    indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
+    segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+    # Creating a dummy input
+    tokens_tensor = torch.tensor([indexed_tokens])
+    segments_tensors = torch.tensor([segments_ids])
+    dummy_input = [tokens_tensor, segments_tensors]
+
+    # Initializing the model with the torchscript flag
+    # Flag set to True even though it is not necessary as this model does not have an LM Head.
+    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
+        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
+
+    # Instantiating the model
+    model = BertModel(config)
+
+    # The model needs to be in evaluation mode
+    model.eval()
+
+    # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
+    model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
+
+    # Creating the trace
+    traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
+    torch.jit.save(traced_model, "traced_bert.pt")
+
+Loading a model
+------------------------------------------------
+
+This snippet shows how to load the ``BertModel`` that was previously saved to disk under the name ``traced_bert.pt``.
+We are re-using the previously initialised ``dummy_input``.
+
+.. code-block:: python
+
+    loaded_model = torch.jit.load("traced_model.pt")
+    loaded_model.eval()
+
+    all_encoder_layers, pooled_output = loaded_model(dummy_input)
+
+Using a traced model for inference
+------------------------------------------------
+
+Using the traced model for inference is as simple as using its ``__call__`` dunder method:
+
+.. code-block:: python
+
+    traced_model(tokens_tensor, segments_tensors)
@@ -49,7 +49,7 @@ put it in train mode.
 .. code-block:: python

    from transformers import BertForSequenceClassification
-    model = BertForSequenceClassification.from_pretrained('bert-base-uncased', return_dict=True)
+    model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
    model.train()

 This is useful because it allows us to make use of the pre-trained BERT
@@ -1,7 +1,7 @@
 # Examples

 Version 2.9 of 🤗 Transformers introduces a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
-Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.
+Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.1+.

 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
@@ -255,11 +255,7 @@ class HansProcessor(DataProcessor):
        return self._create_examples(self._read_tsv(os.path.join(data_dir, "heuristics_evaluation_set.txt")), "dev")

    def get_labels(self):
-        """See base class.
-        Note that we follow the standard three labels for MNLI
-        (see :class:`~transformers.data.processors.utils.MnliProcessor`)
-        but the HANS evaluation groups `contradiction` and `neutral` into `non-entailment` (label 0) while
-        `entailment` is label 1."""
+        """See base class."""
        return ["contradiction", "entailment", "neutral"]

    def _create_examples(self, lines, set_type):
@@ -272,7 +268,7 @@ class HansProcessor(DataProcessor):
            text_a = line[5]
            text_b = line[6]
            pairID = line[7][2:] if line[7].startswith("ex") else line[7]
-            label = line[0]
+            label = line[-1]
            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label, pairID=pairID))
        return examples

@@ -1,11 +0,0 @@
-# tests directory-specific settings - this file is run automatically
-# by pytest before any tests are run
-
-import sys
-from os.path import abspath, dirname, join
-
-
-# allow having multiple repository checkouts and not needing to remember to rerun
-# 'pip install -e .[dev]' when switching between checkouts and running tests.
-git_repo_path = abspath(join(dirname(dirname(__file__)), "src"))
-sys.path.insert(1, git_repo_path)
@@ -5,7 +5,6 @@ import unittest
 from unittest.mock import patch

 import run_glue_deebert
-from transformers.testing_utils import slow


 logging.basicConfig(level=logging.DEBUG)
@@ -21,13 +20,10 @@ def get_setup_file():


 class DeeBertTests(unittest.TestCase):
-    def setup(self) -> None:
+    def test_glue_deebert(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

-    @slow
-    def test_glue_deebert_train(self):
-
        train_args = """
            run_glue_deebert.py
            --model_type roberta
@@ -50,10 +46,6 @@ class DeeBertTests(unittest.TestCase):
            --overwrite_cache
            --eval_after_first_stage
            """.split()
-        with patch.object(sys, "argv", train_args):
-            result = run_glue_deebert.main()
-            for value in result.values():
-                self.assertGreaterEqual(value, 0.666)

        eval_args = """
            run_glue_deebert.py
@@ -71,10 +63,6 @@ class DeeBertTests(unittest.TestCase):
            --overwrite_cache
            --per_gpu_eval_batch_size=1
            """.split()
-        with patch.object(sys, "argv", eval_args):
-            result = run_glue_deebert.main()
-            for value in result.values():
-                self.assertGreaterEqual(value, 0.666)

        entropy_eval_args = """
            run_glue_deebert.py
@@ -92,7 +80,18 @@ class DeeBertTests(unittest.TestCase):
            --overwrite_cache
            --per_gpu_eval_batch_size=1
            """.split()
+
+        with patch.object(sys, "argv", train_args):
+            result = run_glue_deebert.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)
+
+        with patch.object(sys, "argv", eval_args):
+            result = run_glue_deebert.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)
+
        with patch.object(sys, "argv", entropy_eval_args):
            result = run_glue_deebert.main()
            for value in result.values():
-                self.assertGreaterEqual(value, 0.666)
+                self.assertGreaterEqual(value, 0.75)
@@ -60,27 +60,3 @@ python run_language_modeling.py \
    --mlm
 ```

-### XLNet and permutation language modeling
-
-XLNet uses a different training objective, which is permutation language modeling. It is an autoregressive method 
-to learn bidirectional contexts by maximizing the expected likelihood over all permutations of the input 
-sequence factorization order.
-
-We use the `--plm_probability` flag to define the ratio of length of a span of masked tokens to surrounding 
-context length for permutation language modeling.
-
-The `--max_span_length` flag may also be used to limit the length of a span of masked tokens used 
-for permutation language modeling.
-
-```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_name_or_path=xlnet-base-cased \
-    --do_train \
-    --train_data_file=$TRAIN_FILE \
-    --do_eval \
-    --eval_data_file=$TEST_FILE \
-```
@@ -20,10 +20,6 @@ from transformers import (
    AutoTokenizer,
    PretrainedConfig,
    PreTrainedTokenizer,
-)
-from transformers.optimization import (
-    get_cosine_schedule_with_warmup,
-    get_cosine_with_hard_restarts_schedule_with_warmup,
    get_linear_schedule_with_warmup,
 )

@@ -43,19 +39,6 @@ MODEL_MODES = {
 }


-# update this and the import above to support new schedulers from transformers.optimization
-arg_to_scheduler = {
-    "linear": get_linear_schedule_with_warmup,
-    "cosine": get_cosine_schedule_with_warmup,
-    "cosine_w_restarts": get_cosine_with_hard_restarts_schedule_with_warmup,
-    # polynomial': '',                       # TODO
-    # '': get_constant_schedule,             # not supported for now
-    # '': get_constant_schedule_with_warmup, # not supported for now
-}
-arg_to_scheduler_choices = sorted(arg_to_scheduler.keys())
-arg_to_scheduler_metavar = "{" + ", ".join(arg_to_scheduler_choices) + "}"
-
-
 class BaseTransformer(pl.LightningModule):
    def __init__(
        self,
@@ -75,6 +58,7 @@ class BaseTransformer(pl.LightningModule):

        self.hparams = hparams
        self.step_count = 0
+        self.tfmr_ckpts = {}
        self.output_dir = Path(self.hparams.output_dir)
        cache_dir = self.hparams.cache_dir if self.hparams.cache_dir else None
        if config is None:
@@ -86,13 +70,6 @@ class BaseTransformer(pl.LightningModule):
            )
        else:
            self.config: PretrainedConfig = config
-
-        extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout")
-        for p in extra_model_params:
-            if getattr(self.hparams, p, None):
-                assert hasattr(self.config, p), f"model config doesn't have a `{p}` attribute"
-                setattr(self.config, p, getattr(self.hparams, p))
-
        if tokenizer is None:
            self.tokenizer = AutoTokenizer.from_pretrained(
                self.hparams.tokenizer_name if self.hparams.tokenizer_name else self.hparams.model_name_or_path,
@@ -114,16 +91,8 @@ class BaseTransformer(pl.LightningModule):
    def load_hf_checkpoint(self, *args, **kwargs):
        self.model = self.model_type.from_pretrained(*args, **kwargs)

-    def get_lr_scheduler(self):
-        get_schedule_func = arg_to_scheduler[self.hparams.lr_scheduler]
-        scheduler = get_schedule_func(
-            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=self.total_steps
-        )
-        scheduler = {"scheduler": scheduler, "interval": "step", "frequency": 1}
-        return scheduler
-
    def configure_optimizers(self):
-        """Prepare optimizer and schedule (linear warmup and decay)"""
+        "Prepare optimizer and schedule (linear warmup and decay)"
        model = self.model
        no_decay = ["bias", "LayerNorm.weight"]
        optimizer_grouped_parameters = [
@@ -139,8 +108,10 @@ class BaseTransformer(pl.LightningModule):
        optimizer = AdamW(optimizer_grouped_parameters, lr=self.hparams.learning_rate, eps=self.hparams.adam_epsilon)
        self.opt = optimizer

-        scheduler = self.get_lr_scheduler()
-
+        scheduler = get_linear_schedule_with_warmup(
+            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=self.total_steps
+        )
+        scheduler = {"scheduler": scheduler, "interval": "step", "frequency": 1}
        return [optimizer], [scheduler]

    def test_step(self, batch, batch_nb):
@@ -181,9 +152,11 @@ class BaseTransformer(pl.LightningModule):
    @pl.utilities.rank_zero_only
    def on_save_checkpoint(self, checkpoint: Dict[str, Any]) -> None:
        save_path = self.output_dir.joinpath("best_tfmr")
+        save_path.mkdir(exist_ok=True)
        self.model.config.save_step = self.step_count
        self.model.save_pretrained(save_path)
        self.tokenizer.save_pretrained(save_path)
+        self.tfmr_ckpts[self.step_count] = save_path

    @staticmethod
    def add_model_specific_args(parser, root_dir):
@@ -209,31 +182,7 @@ class BaseTransformer(pl.LightningModule):
            type=str,
            help="Where do you want to store the pre-trained models downloaded from s3",
        )
-        parser.add_argument(
-            "--encoder_layerdrop",
-            type=float,
-            help="Encoder layer dropout probability (Optional). Goes into model.config",
-        )
-        parser.add_argument(
-            "--decoder_layerdrop",
-            type=float,
-            help="Decoder layer dropout probability (Optional). Goes into model.config",
-        )
-        parser.add_argument(
-            "--dropout", type=float, help="Dropout probability (Optional). Goes into model.config",
-        )
-        parser.add_argument(
-            "--attention_dropout", type=float, help="Attention dropout probability (Optional). Goes into model.config",
-        )
        parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
-        parser.add_argument(
-            "--lr_scheduler",
-            default="linear",
-            choices=arg_to_scheduler_choices,
-            metavar=arg_to_scheduler_metavar,
-            type=str,
-            help="Learning rate scheduler",
-        )
        parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
        parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
        parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
@@ -302,6 +251,7 @@ def add_generic_args(parser, root_dir) -> None:
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
+
    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")


@@ -204,8 +204,6 @@ if is_tf_available():
            )

        def get_dataset(self):
-            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
-
            return self.dataset

        def __len__(self):
@@ -199,6 +199,9 @@ def train(args, train_dataset, model, tokenizer):
                        {"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
                    )

+            if isinstance(model, torch.nn.DataParallel):
+                inputs["return_tuple"] = True
+
            outputs = model(**inputs)
            # model outputs are always tuple in transformers (see doc)
            loss = outputs[0]
@@ -313,6 +316,8 @@ def evaluate(args, model, tokenizer, prefix=""):
                    inputs.update(
                        {"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
                    )
+            if isinstance(model, torch.nn.DataParallel):
+                inputs["return_tuple"] = True
            outputs = model(**inputs)

        for i, feature_index in enumerate(feature_indices):
@@ -21,8 +21,6 @@ import os
 from dataclasses import dataclass, field
 from typing import Optional

-import tensorflow as tf
-
 from transformers import (
    AutoConfig,
    AutoTokenizer,
@@ -70,7 +68,6 @@ class DataTrainingArguments:
    data_dir: Optional[str] = field(
        default=None, metadata={"help": "The input data dir. Should contain the .json files for the SQuAD task."}
    )
-    use_tfds: Optional[bool] = field(default=True, metadata={"help": "If TFDS should be used or not."})
    max_seq_length: int = field(
        default=128,
        metadata={
@@ -173,7 +170,7 @@ def main():
        )

    # Get datasets
-    if data_args.use_tfds:
+    if not data_args.data_dir:
        if data_args.version_2_with_negative:
            logger.warn("tensorflow_datasets does not handle version 2 of SQuAD. Switch to version 1 automatically")

@@ -182,7 +179,7 @@ def main():
        except ImportError:
            raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")

-        tfds_examples = tfds.load("squad", data_dir=data_args.data_dir)
+        tfds_examples = tfds.load("squad")
        train_examples = (
            SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=False)
            if training_args.do_train
@@ -212,8 +209,6 @@ def main():
        else None
    )

-    train_dataset = train_dataset.apply(tf.data.experimental.assert_cardinality(len(train_examples)))
-
    eval_dataset = (
        squad_convert_examples_to_features(
            examples=eval_examples,
@@ -228,8 +223,6 @@ def main():
        else None
    )

-    eval_dataset = eval_dataset.apply(tf.data.experimental.assert_cardinality(len(eval_examples)))
-
    # Initialize our Trainer
    trainer = TFTrainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)

@@ -13,4 +13,4 @@ streamlit
 elasticsearch
 pandas
 nlp
-fire
+wget
@@ -3,7 +3,7 @@
 This directory contains examples for finetuning and evaluating transformers on summarization and translation tasks.
 Summarization support is more mature than translation support.
 Please tag @sshleifer with any issues/unexpected behaviors, or send a PR!
-For `bertabs` instructions, see [`bertabs/README.md`](bertabs/README.md).
+For `bertabs` instructions, see `bertabs/README.md`. 


 ### Data
@@ -27,7 +27,17 @@ this should make a directory called `cnn_dm/` with files like `test.source`.
 ```

 WMT16 English-Romanian Translation Data:
-download with this command:
+
+This dataset comes in two formats. The "packed" version merges short training examples into examples of <200 tokens to increase GPU utilization (and also improves validation performance).
+
+```bash
+cd examples/seq2seq
+wget https://s3.amazonaws.com/datasets.huggingface.co/translation/wmt_en_ro_packed_train_200.tgz
+tar -xzvf wmt_en_ro_packed_200.tgz
+export ENRO_DIR=wmt_en_ro_packed_train_200
+```
+ 
+The original data can also be downloaded with this command:
 ```bash
 wget https://s3.amazonaws.com/datasets.huggingface.co/translation/wmt_en_ro.tar.gz
 tar -xzvf wmt_en_ro.tar.gz
@@ -35,23 +45,23 @@ export ENRO_DIR=${PWD}/wmt_en_ro
 this should make a directory called `wmt_en_ro/` with files like `test.source`.
 ```

-If you are using your own data, it must be formatted as one directory with 6 files: train.source, train.target, val.source, val.target, test.source, test.target.
+If you are using your own data, it must be formatted as one directory with 6 files: train.source, train.target, val.source, val.target, test.source, test.target.  
 The `.source` files are the input, the `.target` files are the desired output.

-
+ 
 ### Tips and Tricks

 General Tips:
- since you need to run from `examples/seq2seq`, and likely need to modify code, the easiest workflow is fork transformers, clone your fork, and run `pip install -e .` before you get started.
+- since you need to run from `examples/seq2seq`, and likely need to modify code, the easiest workflow is fork transformers, clone your fork, and run `pip install -e .` before you get started.   
 - try `--freeze_encoder` or `--freeze_embeds` for faster training/larger batch size.  (3hr per epoch with bs=8, see the "xsum_shared_task" command below)
 - `fp16_opt_level=O1` (the default works best).
 - In addition to the pytorch-lightning .ckpt checkpoint, a transformers checkpoint will be saved.
 Load it with `BartForConditionalGeneration.from_pretrained(f'{output_dir}/best_tfmr)`.
 - At the moment, `--do_predict` does not work in a multi-gpu setting. You need to use `evaluate_checkpoint` or the `run_eval.py` code.
- This warning can be safely ignored:
+- This warning can be safely ignored: 
    > "Some weights of BartForConditionalGeneration were not initialized from the model checkpoint at facebook/bart-large-xsum and are newly initialized: ['final_logits_bias']"
 - Both finetuning and eval are 30% faster with `--fp16`. For that you need to [install apex](https://github.com/NVIDIA/apex#quick-start).
- Read scripts before you run them!
+- Read scripts before you run them! 

 Summarization Tips:
 - (summ) 1 epoch at batch size 1 for bart-large takes 24 hours and requires 13GB GPU RAM with fp16 on an NVIDIA-V100.
@@ -60,34 +70,12 @@ Summarization Tips:
 - `--max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 ` is a reasonable setting for XSUM.
 - `wandb` can be used by specifying `--logger_name wandb`. It is useful for reproducibility. Specify the environment variable `WANDB_PROJECT='hf_xsum'` to do the XSUM shared task.
 - If you are finetuning on your own dataset, start from `distilbart-cnn-12-6` if you want long summaries and `distilbart-xsum-12-6` if you want short summaries.
-(It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods).
+(It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods). 

 **Update 2018-07-18**
 Datasets: Seq2SeqDataset will be used for all models besides MBart, for which MBartDataset will be used.**
 A new dataset is needed to support multilingual tasks.

-
-### Command Line Options
-
-To see all the possible command line options, run:
-
-```bash
-./finetune.sh --help  # this calls python finetune.py --help
-```
-
-### Finetuning Training Params
-
-To override the pretrained model's training params, you can pass them to `./finetune.sh`:
-
-```bash
-./finetune.sh \
-    [...]
-    --encoder_layerdrop 0.1 \
-    --decoder_layerdrop 0.1 \
-    --dropout 0.1 \
-    --attention_dropout 0.1 \
-```
-
 ### Summarization Finetuning
 Run/modify `finetune.sh`

@@ -111,26 +99,28 @@ Then you can finetune mbart_cc25 on english-romanian with the following command.
 Best performing command:
 ```bash
 # optionally
-export ENRO_DIR='wmt_en_ro' # Download instructions above
-# export WANDB_PROJECT="MT" # optional
-export MAX_LEN=128
+export ENRO_DIR='wmt_en_ro_packed_train_200' # Download instructions above
+# export WANDB_PROJECT="MT" # optional 
+export MAX_LEN=200
 export BS=4
+export GAS=8 # gradient accumulation steps
 ./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --label_smoothing 0.1 --fp16_opt_level=O1 --logger_name wandb --sortish_sampler
 ```
-This should take < 6h/epoch on a 16GB v100 and achieve test BLEU above 26
-To get results in line with fairseq, you need to do some postprocessing. (see `romanian_postprocessing.md`)
+This should take < 2h/epoch on a 16GB v100 and achieve val_avg_ BLEU score above 25. (you can see in wandb or metrics.json).
+To get results in line with fairseq, you need to do some postprocessing.

 MultiGPU command
 (using 8 GPUS as an example)
 ```bash
-export ENRO_DIR='wmt_en_ro' # Download instructions above
+export ENRO_DIR='wmt_en_ro_packed_train_200' # Download instructions above
 # export WANDB_PROJECT="MT" # optional
-export MAX_LEN=128
+export MAX_LEN=200
 export BS=4
+export GAS=1  # gradient accumulation steps
 ./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --gpus 8 --logger_name wandb
 ```
-### Finetuning Outputs
-As you train, `output_dir` will be filled with files, that look kind of like this (comments are mine).
+### Finetuning Outputs 
+As you train, `output_dir` will be filled with files, that look kind of like this (comments are mine). 
 Some of them are metrics, some of them are checkpoints, some of them are metadata. Here is a quick tour:

 ```bash
@@ -148,8 +138,8 @@ output_dir
 ├── student  # this is a huggingface checkpoint generated by SummarizationDistiller. It is the student before it gets finetuned.
 │   ├── config.json
 │   └── pytorch_model.bin
-├── test_generations.txt
-# ^^ are the summaries or translations produced by your best checkpoint on the test data. Populated when training is done
+├── test_generations.txt 
+# ^^ are the summaries or translations produced by your best checkpoint on the test data. Populated when training is done 
 ├── test_results.txt  # a convenience file with the test set metrics. This data is also in metrics.json['test']
 ├── hparams.pkl  # the command line args passed after some light preprocessing. Should be saved fairly quickly.
 ```
@@ -200,8 +190,6 @@ python run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_
    --task summarization \
    --n_obs 100 \
    --device cuda \
-    --max_source_length 1024 \
-    --max_target_length 56 \
    --fp16 \
    --bs 32
 ```
@@ -211,7 +199,7 @@ python run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_
 ![DBART](https://huggingface.co/front/thumbnails/distilbart_large.png)

 For the CNN/DailyMail dataset, (relatively longer, more extractive summaries), we found a simple technique that works:
-you just copy alternating layers from `bart-large-cnn` and finetune more on the same data.
+you just copy alternating layers from `bart-large-cnn` and finetune more on the same data. 

 For the XSUM dataset, that didn’t work as well so we used that same initialization strategy followed by a combination of Distillbert’s ce_loss and the hidden states MSE loss used in the tinybert paper.

@@ -227,7 +215,7 @@ They are initialized by copying layers from the associated `bart-large-{cnn|xsum
 The command that produced `sshleifer/distilbart-cnn-12-6` is
 ```bash
 ./train_distilbart_cnn.sh
-```
+```  
 runtime: 6H on NVIDIA RTX 24GB GPU

 *Note*: You can get the same simple distillation logic by using `./run_distiller.sh --no_teacher` followed by identical arguments as the ones in `train_distilbart_cnn.sh`.
@@ -243,15 +231,15 @@ This is how `sshleifer/distilbart-xsum*` checkpoints were produced.
 The command that produced `sshleifer/distilbart-xsum-12-6` is:

 ```bash
-./train_distilbart_xsum.sh
+./train_distilbart_xsum.sh  
 ```

-runtime: 13H on V-100 16GB GPU.
+runtime: 13H on V-100 16GB GPU. 

 ### Contributing
 - follow the standard contributing guidelines and code of conduct.
 - add tests to `test_seq2seq_examples.py`
 - To run only the seq2seq tests, you must be in the root of the repository and run:
 ```bash
-pytest examples/seq2seq/
+pytest examples/seq2seq/  
 ```
@@ -19,10 +19,6 @@ logger = logging.getLogger(__name__)


 class Seq2SeqLoggingCallback(pl.Callback):
-    def on_batch_end(self, trainer, pl_module):
-        lrs = {f"lr_group_{i}": param["lr"] for i, param in enumerate(pl_module.trainer.optimizers[0].param_groups)}
-        pl_module.logger.log_metrics(lrs)
-
    @rank_zero_only
    def _write_logs(
        self, trainer: pl.Trainer, pl_module: pl.LightningModule, type_path: str, save_generations=True
@@ -1,21 +0,0 @@
-from typing import Union
-
-import fire
-import torch
-from tqdm import tqdm
-
-
-def convert(src_path: str, map_location: str = "cpu", save_path: Union[str, None] = None) -> None:
-    """Convert a pytorch_model.bin or model.pt file to torch.float16 for faster downloads, less disk space."""
-    state_dict = torch.load(src_path, map_location=map_location)
-    for k, v in tqdm(state_dict.items()):
-        if not isinstance(v, torch.Tensor):
-            raise TypeError("FP16 conversion only works on paths that are saved state dics, like pytorch_model.bin")
-        state_dict[k] = v.half()
-    if save_path is None:  # overwrite src_path
-        save_path = src_path
-    torch.save(state_dict, save_path)
-
-
-if __name__ == "__main__":
-    fire.Fire(convert)
@@ -14,7 +14,7 @@ import torch
 from torch.utils.data import DataLoader

 from lightning_base import BaseTransformer, add_generic_args, generic_train
-from transformers import MBartTokenizer, T5ForConditionalGeneration, get_linear_schedule_with_warmup
+from transformers import MBartTokenizer, get_linear_schedule_with_warmup


 try:
@@ -105,13 +105,7 @@ class SummarizationModule(BaseTransformer):
        self.hparams.git_sha = get_git_info()["repo_sha"]
        self.num_workers = hparams.num_workers
        self.decoder_start_token_id = None
-        if self.model.config.decoder_start_token_id is None and isinstance(self.tokenizer, MBartTokenizer):
-            self.decoder_start_token_id = self.tokenizer.lang_code_to_id[hparams.tgt_lang]
-            self.model.config.decoder_start_token_id = self.decoder_start_token_id
-        if isinstance(self.tokenizer, MBartTokenizer):
-            self.dataset_class = MBartDataset
-        else:
-            self.dataset_class = Seq2SeqDataset
+        self.dataset_class = Seq2SeqDataset

    def freeze_embeds(self):
        """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
@@ -137,14 +131,8 @@ class SummarizationModule(BaseTransformer):
    def _step(self, batch: dict) -> Tuple:
        pad_token_id = self.tokenizer.pad_token_id
        source_ids, source_mask, target_ids = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
-
-        if isinstance(self.model, T5ForConditionalGeneration):
-            decoder_input_ids = self.model._shift_right(target_ids)
-            lm_labels = target_ids
-        else:
-            decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
-            lm_labels = target_ids[:, 1:].clone()  # why clone?
-
+        decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
+        lm_labels = target_ids[:, 1:].clone()  # why clone?
        outputs = self(source_ids, attention_mask=source_mask, decoder_input_ids=decoder_input_ids, use_cache=False)

        if self.hparams.label_smoothing == 0:
@@ -160,16 +148,9 @@ class SummarizationModule(BaseTransformer):
            )
        return (loss,)

-    @property
-    def pad(self) -> int:
-        return self.tokenizer.pad_token_id
-
    def training_step(self, batch, batch_idx) -> Dict:
        loss_tensors = self._step(batch)
-
        logs = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
-        # tokens per batch
-        logs["tpb"] = batch["input_ids"].ne(self.pad).sum() + batch["decoder_input_ids"].ne(self.pad).sum()
        return {"loss": loss_tensors[0], "log": logs}

    def validation_step(self, batch, batch_idx) -> Dict:
@@ -179,7 +160,7 @@ class SummarizationModule(BaseTransformer):
        self.step_count += 1
        losses = {k: torch.stack([x[k] for x in outputs]).mean() for k in self.loss_names}
        loss = losses["loss"]
-        rouges = {k: np.array([x[k] for x in outputs]).mean() for k in self.metric_names + ["gen_time", "gen_len"]}
+        rouges = {k: np.array([x[k] for x in outputs]).mean() for k in self.metric_names + ["gen_time", "summ_len"]}
        rouge_tensor: torch.FloatTensor = torch.tensor(rouges[self.val_metric]).type_as(loss)
        rouges.update({k: v.item() for k, v in losses.items()})
        losses.update(rouges)
@@ -197,21 +178,23 @@ class SummarizationModule(BaseTransformer):
        return calculate_rouge(preds, target)

    def _generative_step(self, batch: dict) -> dict:
+        pad_token_id = self.tokenizer.pad_token_id
+        source_ids, source_mask, y = Seq2SeqDataset.trim_seq2seq_batch(batch, pad_token_id)
        t0 = time.time()
        generated_ids = self.model.generate(
-            batch["input_ids"],
-            attention_mask=batch["attention_mask"],
+            input_ids=source_ids,
+            attention_mask=source_mask,
            use_cache=True,
            decoder_start_token_id=self.decoder_start_token_id,
        )
-        gen_time = (time.time() - t0) / batch["input_ids"].shape[0]
-        preds: List[str] = self.ids_to_clean_text(generated_ids)
-        target: List[str] = self.ids_to_clean_text(batch["decoder_input_ids"])
+        gen_time = (time.time() - t0) / source_ids.shape[0]
+        preds = self.ids_to_clean_text(generated_ids)
+        target = self.ids_to_clean_text(y)
        loss_tensors = self._step(batch)
        base_metrics = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
        rouge: Dict = self.calc_generative_metrics(preds, target)
        summ_len = np.mean(lmap(len, generated_ids))
-        base_metrics.update(gen_time=gen_time, gen_len=summ_len, preds=preds, target=target, **rouge)
+        base_metrics.update(gen_time=gen_time, summ_len=summ_len, preds=preds, target=target, **rouge)
        return base_metrics

    def test_step(self, batch, batch_idx):
@@ -342,6 +325,11 @@ class TranslationModule(SummarizationModule):
        super().__init__(hparams, **kwargs)
        self.dataset_kwargs["src_lang"] = hparams.src_lang
        self.dataset_kwargs["tgt_lang"] = hparams.tgt_lang
+        if self.model.config.decoder_start_token_id is None and isinstance(self.tokenizer, MBartTokenizer):
+            self.decoder_start_token_id = self.tokenizer.lang_code_to_id[hparams.tgt_lang]
+            self.model.config.decoder_start_token_id = self.decoder_start_token_id
+        if isinstance(self.tokenizer, MBartTokenizer):
+            self.dataset_class = MBartDataset

    def calc_generative_metrics(self, preds, target) -> dict:
        return calculate_bleu_score(preds, target)
@@ -2,7 +2,6 @@
 export PYTHONPATH="../":"${PYTHONPATH}"

 # the proper usage is documented in the README, you need to specify data_dir, output_dir and model_name_or_path
-# run ./finetune.sh --help to see all the possible options
 python finetune.py \
    --learning_rate=3e-5 \
    --fp16 \
@@ -8,7 +8,6 @@ python finetune.py \
 --eval_batch_size=$BS \
 --output_dir=$OUTPUT_DIR \
 --max_source_length=512 \
--max_target_length=56 \
 --val_check_interval=0.1 --n_val=200 \
 --do_train --do_predict \
 $@
@@ -1,19 +0,0 @@
-from pathlib import Path
-
-import fire
-
-
-def minify(src_dir: str, dest_dir: str, n: int):
-    """Write first n lines of each file f in src_dir to dest_dir/f """
-    src_dir = Path(src_dir)
-    dest_dir = Path(dest_dir)
-    dest_dir.mkdir(exist_ok=True)
-    for path in src_dir.iterdir():
-        new = [x.rstrip() for x in list(path.open().readlines())][:n]
-        dest_path = dest_dir.joinpath(path.name)
-        print(dest_path)
-        dest_path.open("w").write("\n".join(new))
-
-
-if __name__ == "__main__":
-    fire.Fire(minify)
@@ -6,7 +6,6 @@
 """

 import argparse
-import shutil
 from pathlib import Path

 from tqdm import tqdm
@@ -18,7 +17,7 @@ def pack_examples(tok, src_examples, tgt_examples, max_tokens=1024):

    finished_src, finished_tgt = [], []

-    sorted_examples = list(zip(src_examples, tgt_examples))
+    sorted_examples = list(sorted(zip(src_examples, tgt_examples), key=lambda x: len(x[0])))
    new_src, new_tgt = sorted_examples[0]

    def is_too_big(strang):
@@ -43,10 +42,20 @@ def pack_examples(tok, src_examples, tgt_examples, max_tokens=1024):
    return finished_src, finished_tgt


+def minify(src_dir: Path, dest_dir: Path, n: int):
+    """Write first n lines of each file f in src_dir to dest_dir/f"""
+    dest_dir.mkdir(exist_ok=True)
+    for path in src_dir.iterdir():
+        new = [x.rstrip() for x in list(path.open().readlines())][:n]
+        dest_path = dest_dir.joinpath(path.name)
+        print(dest_path)
+        dest_path.open("w").write("\n".join(new))
+
+
 def pack_data_dir(tok, data_dir: Path, max_tokens, save_path):
    save_path = Path(save_path)
    save_path.mkdir(exist_ok=True)
-    for split in ["train"]:
+    for split in ["val", "test", "train"]:
        src_path, tgt_path = data_dir / f"{split}.source", data_dir / f"{split}.target"
        src_docs = [x.rstrip() for x in Path(src_path).open().readlines()]
        tgt_docs = [x.rstrip() for x in Path(tgt_path).open().readlines()]
@@ -54,10 +63,6 @@ def pack_data_dir(tok, data_dir: Path, max_tokens, save_path):
        print(f"packed {split} split from {len(src_docs)} examples -> {len(packed_src)}.")
        Path(save_path / f"{split}.source").open("w").write("\n".join(packed_src))
        Path(save_path / f"{split}.target").open("w").write("\n".join(packed_tgt))
-    for split in ["val", "test"]:
-        src_path, tgt_path = data_dir / f"{split}.source", data_dir / f"{split}.target"
-        shutil.copyfile(src_path, save_path / f"{split}.source")
-        shutil.copyfile(tgt_path, save_path / f"{split}.target")


 def packer_cli():
@@ -1,16 +1,15 @@
 import argparse
 import os
 import sys
+import tarfile
 import tempfile
 from pathlib import Path
 from unittest.mock import patch

 import pytest
 import pytorch_lightning as pl
-import timeout_decorator
-import torch
+import wget

-from transformers import BartForConditionalGeneration
 from transformers.testing_utils import slow

 from .finetune import SummarizationModule, main
@@ -18,18 +17,20 @@ from .test_seq2seq_examples import CUDA_AVAILABLE, MBART_TINY
 from .utils import load_json


-MODEL_NAME = MBART_TINY
-# TODO(SS): MODEL_NAME = "sshleifer/student_mbart_en_ro_1_1"
+def fetch_and_save_wmt_100():
+    # TODO(SS): DELETEME
+    DATA_URL = "https://s3.amazonaws.com/datasets.huggingface.co/translation/wmt_en_ro_test_data.tgz"
+
+    dest_dir = "wmt_100_dir"
+    if os.path.isdir(dest_dir):
+        return dest_dir
+    filename = wget.download(DATA_URL)
+    tarball = tarfile.TarFile(filename)
+    tarball.extractall(path=dest_dir)
+    os.remove(filename)
+    return dest_dir


-@slow
-@pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
-def test_model_download():
-    """This warms up the cache so that we can time the next test without including download time, which varies between machines."""
-    BartForConditionalGeneration.from_pretrained(MODEL_NAME)
-
-
-@timeout_decorator.timeout(120)
@slow
@pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
 def test_train_mbart_cc25_enro_script():
@@ -39,10 +40,7 @@ def test_train_mbart_cc25_enro_script():
        "$BS": 4,
        "$GAS": 1,
        "$ENRO_DIR": data_dir,
-        "facebook/mbart-large-cc25": MODEL_NAME,
-        # 1 encoder and 1 decoder layer from finetuned mbart en-ro. Should be able to start >0 and improve quickly.
-        # Download is 600MB in previous test.
-        "val_check_interval=0.25": "val_check_interval=1.0",
+        "facebook/mbart-large-cc25": MBART_TINY,
    }

    # Clean up bash script
@@ -55,54 +53,40 @@ def test_train_mbart_cc25_enro_script():
    if CUDA_AVAILABLE:
        gpus = 1  # torch.cuda.device_count()
    else:
+        bash_script = bash_script.replace("--fp16", "")
        gpus = 0
-    bash_script = bash_script.replace("--fp16", "")
+
    testargs = (
        ["finetune.py"]
        + bash_script.split()
-        + [
-            f"--output_dir={output_dir}",
-            f"--gpus={gpus}",
-            "--learning_rate=3e-1",
-            "--warmup_steps=0",
-            "--val_check_interval=1.0",
-            "--tokenizer_name=facebook/mbart-large-en-ro",
-        ]
+        + [f"--output_dir={output_dir}", f"--gpus={gpus}", "--learning_rate=3e-1"]
    )
    with patch.object(sys, "argv", testargs):
        parser = argparse.ArgumentParser()
        parser = pl.Trainer.add_argparse_args(parser)
        parser = SummarizationModule.add_model_specific_args(parser, os.getcwd())
        args = parser.parse_args()
-        args.do_predict = False
-        # assert args.gpus == gpus THIS BREAKS for multigpu
+        # assert args.gpus == gpus THIS BREAKS
+        # args.gpus = gpus
        model = main(args)
+    contents = os.listdir(output_dir)
+    # ckpt_name = "val_avg_rouge2=0.0000-step_count=2.ckpt"  # "val_avg_rouge2=0.0000-epoch=1.ckpt"  #
+    # "epoch=1-val_avg_rouge2=0.0000.ckpt"
+    contents = {os.path.basename(p) for p in contents}
+
+    # self.assertIn(ckpt_name, contents)
+
+    assert "test_generations.txt" in contents
+    assert "test_results.txt" in contents

-    # Check metrics
    metrics = load_json(model.metrics_save_path)
    first_step_stats = metrics["val"][0]
    last_step_stats = metrics["val"][-1]
-    assert len(metrics["val"]) == (args.max_epochs / args.val_check_interval)  # +1 accounts for val_sanity_check
-
    assert last_step_stats["val_avg_gen_time"] >= 0.01
+    assert 1.0 >= last_step_stats["val_avg_gen_time"]
+    assert first_step_stats["val_avg_bleu"] < last_step_stats["val_avg_bleu"]

-    assert first_step_stats["val_avg_bleu"] < last_step_stats["val_avg_bleu"]  # model learned nothing
-    assert 1.0 >= last_step_stats["val_avg_gen_time"]  # model hanging on generate. Maybe bad config was saved.
    assert isinstance(last_step_stats[f"val_avg_{model.val_metric}"], float)
-
-    # check lightning ckpt can be loaded and has a reasonable statedict
-    contents = os.listdir(output_dir)
-    ckpt_path = [x for x in contents if x.endswith(".ckpt")][0]
-    full_path = os.path.join(args.output_dir, ckpt_path)
-    ckpt = torch.load(full_path, map_location="cpu")
-    expected_key = "model.model.decoder.layers.0.encoder_attn_layer_norm.weight"
-    assert expected_key in ckpt["state_dict"]
-    assert ckpt["state_dict"]["model.model.decoder.layers.0.encoder_attn_layer_norm.weight"].dtype == torch.float32
-
-    # TODO(SS): turn on args.do_predict when PL bug fixed.
-    if args.do_predict:
-        contents = {os.path.basename(p) for p in contents}
-        assert "test_generations.txt" in contents
-        assert "test_results.txt" in contents
-        # assert len(metrics["val"]) ==  desired_n_evals
-        assert len(metrics["test"]) == 1
+    # desired_n_evals = int(args_d["max_epochs"] * (1 / args_d["val_check_interval"]) + 1)
+    # assert len(metrics["val"]) ==  desired_n_evals
+    assert len(metrics["test"]) == 1
@@ -1,20 +1,100 @@
-UN Chief Says There Is No Military Solution in Syria Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people. The U.N. chief again urged all parties, including the divided U.N. Security Council, to unite and support inclusive negotiations to find a political solution. Ban told a news conference Wednesday that he plans to meet with foreign ministers of the five permanent council nations - the U.S., Russia, China, Britain and France - on the sidelines of the General Assembly's ministerial session later this month to discuss Syria.
-He expressed regret that divisions in the council and among the Syrian people and regional powers "made this situation unsolvable." Ban urged the five permanent members to show the solidarity and unity they did in achieving an Iran nuclear deal in addressing the Syria crisis. 8 Poll Numbers That Show Donald Trump Is For Real Some have tried to label him a flip-flopper. Others have dismissed him as a joke. And some are holding out for an implosion. But no matter how some Republicans are trying to drag Donald Trump down from atop the polls, it hasn't worked (yet).
-Ten of the last 11 national polls have shown Donald Trump's lead at double digits, and some are starting to ask seriously what it means for the real estate mogul's nomination chances. Of course, it's still early in the election cycle. None of this is to say that Trump is likely to win the Republican nomination. Pundits point out that at this time in 2011, Rick Perry's lead was giving way to a rising Herman Cain, neither of whom won even one state in the nomination process. And there are many reasons he would struggle in a general election. But outside groups like Jeb Bush's Super PAC and the economic conservative group Club for Growth are recognizing Trump's staying power and beginning to unload their dollars to topple him.
-Here are some recent poll numbers that suggest that the real estate mogul isn't just a passing phase: Trump's favorability ratings have turned 180 degrees. Right before Donald Trump announced his candidacy in mid-June, a Monmouth University poll showed only two in 10 Republicans had a positive view of the real estate mogul. By mid-July, it was 40 percent. In early August, it was 52 percent. Now, six in 10 Republicans have a favorable view of Donald Trump. Roughly three in 10 say they have a negative view. And these numbers hold up in early states. A Quinnipiac poll in Iowa last week found that 60 percent of Republicans there had a favorable view of Trump.
-Two-thirds of GOP voters would be happy with Trump as the nominee. In a CNN/ORC poll last week, 67 percent of Republicans said they would be either "enthusiastic" or "satisfied" if Trump were the nominee. Only two in 10 say they would be "upset" if he were the nominee. Only Ben Carson generates roughly the same level of enthusiasm as Trump (43 percent say they would be "enthusiastic" vs. 40 percent who say the same of Trump). The next closest in enthusiasm? Marco Rubio with only 21 percent.
-On the flip side, 47 percent of Republican voters say they would be "dissatisfied" or "upset" if establishment favorite Jeb Bush becomes the nominee. A majority of Republicans don't see Trump's temperament as a problem. While Donald Trump has been widely criticized for his bombast and insults, 52 percent of leaned Republican voters nationwide think that the real estate mogul has the right temperament to be president, according to Monday's ABC News/Washington Post poll. The same number holds in the first-in-the-nation caucus state of Iowa, where the same 52 percent of Republicans think he has the personality to be commander in chief, according to Quinnipiac last week.
-Still, 44 percent think he doesn't have the personality to serve effectively, and almost six in 10 independents say his temperament does not belong in the White House, according to ABC/Post. Republican voters are getting used to the idea. When they put on their pundit hats, Republican voters think Trump is for real. When asked who is most likely to win the GOP nomination, four in 10 said Trump was the best bet, according to a CNN/ORC poll out last week. That's a change from when four in 10 placed their money on Jeb Bush in late July. Full disclosure: GOP voters haven't had the clearest crystal ball in the past.
-At this time last cycle, four in 10 Republicans picked Rick Perry to win the nomination, vs. only 28 percent for eventual nominee Mitt Romney. Still, it shows that a plurality of GOP voters see Trump's campaign as plausible. Even if Republicans rallied around another candidate, Trump still beats almost everyone. Some pundits point out that the splintered field is likely contributing to Trump's lead, while anti-Trump support is be spread diffusely among more than a dozen other candidates. But a Monmouth University poll in early September shows that, in a hypothetical head-to-head matchup between Trump and most other Republican candidates, Trump almost always garners majority support.
-He leads Carly Fiorina by 13 points, Marco Rubio by 14 points, Walker by 15 points, Jeb Bush by 19 points, and, finally, Rand Paul, John Kasich and Chris Christie by 33 points each. He's in a dead heat with Ted Cruz. The only candidate who beats him? Ben Carson would lead the businessman by a wide 19 points in a hypothetical head-to-head. A bare majority of Donald Trump's supporters say they've made up their minds. A new CBS/NYT poll out on Tuesday shows that just more than half of voters who support Trump say they have locked in their votes. Obviously, a lot can happen to change that, and no one can really say they would never change their mind.
-46 percent said they are leaving the door open to switching candidates. Still, Trump's strongest competition at the moment is from fellow outsider neurosurgeon Ben Carson, but voters who say they have made up their minds are twice as likely to go for Trump. Six in 10 Republicans say they agree with Trump on immigration. Even since Donald Trump called immigrants from Mexico "rapists" in his campaign announcement speech two months ago, immigration has been front and center in the 2016 conversation. Some are worried that Trump's bombast will drive crucial Hispanic voters away from the Republican Party and damage rebranding efforts.
-But according to Monday's new ABC/Post poll, six in 10 Republicans say they agree with Trump on immigration issues. So as long as immigration remains in the spotlight, it seems Donald Trump will remain too. Frustration with government is climbing to new highs. Donald Trump and Ben Carson now account for roughly half of the support from Republican voters, largely due to their outsider status. Six in 10 Republicans in Monday's new ABC/Post poll say they want a political outsider over someone with government experience. And they are angry at Washington, too.
-A Des Moines Register/Bloomberg poll in Iowa from two weeks ago shows that three in four Iowa Republicans are frustrated with Republicans in Congress, with 54 percent "unsatisfied" and 21 percent "mad as hell." Jeremy Corbyn to make debut at Prime Minister's Questions Since his election, Mr Corbyn's debut at PMQs has been keenly awaited New Labour leader Jeremy Corbyn is to make his debut at Prime Minister's Questions later, taking on David Cameron for the first time.
-Mr Corbyn will rise to ask the first of his six allotted questions shortly after midday, with his performance likely to be closely scrutinised by the media and Labour MPs. He has called for "less theatre and more facts" at the weekly showpiece. He has also said he could skip some sessions, leaving them to colleagues. The encounter will be the first parliamentary test of Mr Corbyn's leadership, coming after his appointment of a shadow cabinet and his speech to the TUC annual congress on Tuesday.
-Meanwhile, the Labour leader's decision to stand in silence during the singing of the national anthem at a service on Tuesday to mark the 75th anniversary of the Battle of Britain has attracted criticism from a number of Tory MPs and is the focus of several front page stories in the newspapers. Mr Corbyn's decision not to sing the national anthem has attracted attention A spokesman for Mr Corbyn said he had "stood in respectful silence" and did recognise the "heroism of the Royal Air Force in the Battle of Britain."
-But a member of Mr Corbyn's shadow cabinet, Owen Smith, told BBC Two's Newsnight programme he would have advised the Labour leader to sing the national anthem "irrespective" of his belief that the monarchy should be abolished. Nearly a dozen shadow ministers have refused to serve in Mr Corbyn's top team, citing differences over the economy, defence and foreign affairs, while less than a sixth of the parliamentary party originally backed him as leader. BBC political correspondent Robin Brant says policy differences are also "stacking up" within Labour following Mr Corbyn's appointment over its position on the European Union and the government's cap on benefits.
-Mr Corbyn told the TUC conference Labour was putting forward amendments to remove the whole idea of a cap altogether. Hours later Mr Smith, the shadow work and pensions secretary, said the party was "very clear" that it was only opposing government plans to reduce the level of cap from £26,000 to £23,000. Mr Corbyn will be the fifth Labour leader that David Cameron has faced across the despatch box over the past decade since he became Tory leader. The Labour leader, who has promised a different approach to politics, says he has "crowd sourced" ideas for questions to ask Mr Cameron and has been given more than 30,000 suggestions.
-The Islington North MP has said PMQs is too confrontational and that he will refrain from both "repartee" and trading barbs, instead vowing to focus on serious issues such as poverty, inequality and the challenges facing young people. Mr Corbyn has said that Angela Eagle, the shadow business secretary, will deputise for him at PMQs when he does not attend - for instance when Mr Cameron is travelling abroad. He has also floated the idea of allowing other colleagues to take the floor on occasion, saying he had approached the Commons Speaker John Bercow to discuss the issue.
-When he became leader in 2005, Mr Cameron said he wanted to move away from the "Punch and Judy" style of politics often associated with PMQs but admitted some years later that he had failed. Since it was first televised in 1990, PMQs has been seen as a key barometer of a leader's judgement, their command of the Commons and their standing among their fellow MPs although critics have argued it has become a caricature and is in need of far-reaching reforms. 'Shot in Joburg': Homeless youth trained as photographers Downtown Johannesburg is a tough place to be homeless.
-But one group of former street children have found a way to learn a skill and make a living. "I was shot in Joburg" is a non-profit studio that teaches homeless youngsters how to take photographs of their neighbourhood and make a profit from it. BBC News went to meet one of the project's first graduates. JD Sports boss says higher wages could hurt expansion JD Sports Executive Chairman Peter Cowgill says a higher minimum wage for UK workers could mean "more spending power in the pockets of potential consumers." But that spending power is unlikely to outweigh the higher labour costs at his firm, he says.
-The costs could hit JD Sports' expansion plans, he added, which could mean fewer extra jobs. Thanasi Kokkinakis backed by Tennis Australia president Steve Healy Thanasi Kokkinakis deserves kudos rather than criticism for his behaviour. Thanasi Kokkinakis has been the collateral damage in the recent storm around his friend Nick Kyrgios and deserves kudos rather than criticism for his own behaviour, according to Tennis Australia president Steve Healy.
+UN Chief Says There Is No Military Solution in Syria
+Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people.
+The U.N. chief again urged all parties, including the divided U.N. Security Council, to unite and support inclusive negotiations to find a political solution.
+Ban told a news conference Wednesday that he plans to meet with foreign ministers of the five permanent council nations - the U.S., Russia, China, Britain and France - on the sidelines of the General Assembly's ministerial session later this month to discuss Syria.
+He expressed regret that divisions in the council and among the Syrian people and regional powers "made this situation unsolvable."
+Ban urged the five permanent members to show the solidarity and unity they did in achieving an Iran nuclear deal in addressing the Syria crisis.
+8 Poll Numbers That Show Donald Trump Is For Real
+Some have tried to label him a flip-flopper.
+Others have dismissed him as a joke.
+And some are holding out for an implosion.
+But no matter how some Republicans are trying to drag Donald Trump down from atop the polls, it hasn't worked (yet).
+Ten of the last 11 national polls have shown Donald Trump's lead at double digits, and some are starting to ask seriously what it means for the real estate mogul's nomination chances.
+Of course, it's still early in the election cycle.
+None of this is to say that Trump is likely to win the Republican nomination.
+Pundits point out that at this time in 2011, Rick Perry's lead was giving way to a rising Herman Cain, neither of whom won even one state in the nomination process.
+And there are many reasons he would struggle in a general election.
+But outside groups like Jeb Bush's Super PAC and the economic conservative group Club for Growth are recognizing Trump's staying power and beginning to unload their dollars to topple him.
+Here are some recent poll numbers that suggest that the real estate mogul isn't just a passing phase:
+Trump's favorability ratings have turned 180 degrees.
+Right before Donald Trump announced his candidacy in mid-June, a Monmouth University poll showed only two in 10 Republicans had a positive view of the real estate mogul.
+By mid-July, it was 40 percent.
+In early August, it was 52 percent.
+Now, six in 10 Republicans have a favorable view of Donald Trump.
+Roughly three in 10 say they have a negative view.
+And these numbers hold up in early states.
+A Quinnipiac poll in Iowa last week found that 60 percent of Republicans there had a favorable view of Trump.
+Two-thirds of GOP voters would be happy with Trump as the nominee.
+In a CNN/ORC poll last week, 67 percent of Republicans said they would be either "enthusiastic" or "satisfied" if Trump were the nominee.
+Only two in 10 say they would be "upset" if he were the nominee.
+Only Ben Carson generates roughly the same level of enthusiasm as Trump (43 percent say they would be "enthusiastic" vs. 40 percent who say the same of Trump).
+The next closest in enthusiasm?
+Marco Rubio with only 21 percent.
+On the flip side, 47 percent of Republican voters say they would be "dissatisfied" or "upset" if establishment favorite Jeb Bush becomes the nominee.
+A majority of Republicans don't see Trump's temperament as a problem.
+While Donald Trump has been widely criticized for his bombast and insults, 52 percent of leaned Republican voters nationwide think that the real estate mogul has the right temperament to be president, according to Monday's ABC News/Washington Post poll.
+The same number holds in the first-in-the-nation caucus state of Iowa, where the same 52 percent of Republicans think he has the personality to be commander in chief, according to Quinnipiac last week.
+Still, 44 percent think he doesn't have the personality to serve effectively, and almost six in 10 independents say his temperament does not belong in the White House, according to ABC/Post.
+Republican voters are getting used to the idea.
+When they put on their pundit hats, Republican voters think Trump is for real.
+When asked who is most likely to win the GOP nomination, four in 10 said Trump was the best bet, according to a CNN/ORC poll out last week.
+That's a change from when four in 10 placed their money on Jeb Bush in late July.
+Full disclosure: GOP voters haven't had the clearest crystal ball in the past.
+At this time last cycle, four in 10 Republicans picked Rick Perry to win the nomination, vs. only 28 percent for eventual nominee Mitt Romney.
+Still, it shows that a plurality of GOP voters see Trump's campaign as plausible.
+Even if Republicans rallied around another candidate, Trump still beats almost everyone.
+Some pundits point out that the splintered field is likely contributing to Trump's lead, while anti-Trump support is be spread diffusely among more than a dozen other candidates.
+But a Monmouth University poll in early September shows that, in a hypothetical head-to-head matchup between Trump and most other Republican candidates, Trump almost always garners majority support.
+He leads Carly Fiorina by 13 points, Marco Rubio by 14 points, Walker by 15 points, Jeb Bush by 19 points, and, finally, Rand Paul, John Kasich and Chris Christie by 33 points each.
+He's in a dead heat with Ted Cruz.
+The only candidate who beats him?
+Ben Carson would lead the businessman by a wide 19 points in a hypothetical head-to-head.
+A bare majority of Donald Trump's supporters say they've made up their minds.
+A new CBS/NYT poll out on Tuesday shows that just more than half of voters who support Trump say they have locked in their votes.
+Obviously, a lot can happen to change that, and no one can really say they would never change their mind.
+46 percent said they are leaving the door open to switching candidates.
+Still, Trump's strongest competition at the moment is from fellow outsider neurosurgeon Ben Carson, but voters who say they have made up their minds are twice as likely to go for Trump.
+Six in 10 Republicans say they agree with Trump on immigration.
+Even since Donald Trump called immigrants from Mexico "rapists" in his campaign announcement speech two months ago, immigration has been front and center in the 2016 conversation.
+Some are worried that Trump's bombast will drive crucial Hispanic voters away from the Republican Party and damage rebranding efforts.
+But according to Monday's new ABC/Post poll, six in 10 Republicans say they agree with Trump on immigration issues.
+So as long as immigration remains in the spotlight, it seems Donald Trump will remain too.
+Frustration with government is climbing to new highs.
+Donald Trump and Ben Carson now account for roughly half of the support from Republican voters, largely due to their outsider status.
+Six in 10 Republicans in Monday's new ABC/Post poll say they want a political outsider over someone with government experience.
+And they are angry at Washington, too.
+A Des Moines Register/Bloomberg poll in Iowa from two weeks ago shows that three in four Iowa Republicans are frustrated with Republicans in Congress, with 54 percent "unsatisfied" and 21 percent "mad as hell."
+Jeremy Corbyn to make debut at Prime Minister's Questions
+Since his election, Mr Corbyn's debut at PMQs has been keenly awaited
+New Labour leader Jeremy Corbyn is to make his debut at Prime Minister's Questions later, taking on David Cameron for the first time.
+Mr Corbyn will rise to ask the first of his six allotted questions shortly after midday, with his performance likely to be closely scrutinised by the media and Labour MPs.
+He has called for "less theatre and more facts" at the weekly showpiece.
+He has also said he could skip some sessions, leaving them to colleagues.
+The encounter will be the first parliamentary test of Mr Corbyn's leadership, coming after his appointment of a shadow cabinet and his speech to the TUC annual congress on Tuesday.
+Meanwhile, the Labour leader's decision to stand in silence during the singing of the national anthem at a service on Tuesday to mark the 75th anniversary of the Battle of Britain has attracted criticism from a number of Tory MPs and is the focus of several front page stories in the newspapers.
+Mr Corbyn's decision not to sing the national anthem has attracted attention
+A spokesman for Mr Corbyn said he had "stood in respectful silence" and did recognise the "heroism of the Royal Air Force in the Battle of Britain."
+But a member of Mr Corbyn's shadow cabinet, Owen Smith, told BBC Two's Newsnight programme he would have advised the Labour leader to sing the national anthem "irrespective" of his belief that the monarchy should be abolished.
+Nearly a dozen shadow ministers have refused to serve in Mr Corbyn's top team, citing differences over the economy, defence and foreign affairs, while less than a sixth of the parliamentary party originally backed him as leader.
+BBC political correspondent Robin Brant says policy differences are also "stacking up" within Labour following Mr Corbyn's appointment over its position on the European Union and the government's cap on benefits.
+Mr Corbyn told the TUC conference Labour was putting forward amendments to remove the whole idea of a cap altogether.
+Hours later Mr Smith, the shadow work and pensions secretary, said the party was "very clear" that it was only opposing government plans to reduce the level of cap from £26,000 to £23,000.
+Mr Corbyn will be the fifth Labour leader that David Cameron has faced across the despatch box over the past decade since he became Tory leader.
+The Labour leader, who has promised a different approach to politics, says he has "crowd sourced" ideas for questions to ask Mr Cameron and has been given more than 30,000 suggestions.
+The Islington North MP has said PMQs is too confrontational and that he will refrain from both "repartee" and trading barbs, instead vowing to focus on serious issues such as poverty, inequality and the challenges facing young people.
+Mr Corbyn has said that Angela Eagle, the shadow business secretary, will deputise for him at PMQs when he does not attend - for instance when Mr Cameron is travelling abroad.
+He has also floated the idea of allowing other colleagues to take the floor on occasion, saying he had approached the Commons Speaker John Bercow to discuss the issue.
+When he became leader in 2005, Mr Cameron said he wanted to move away from the "Punch and Judy" style of politics often associated with PMQs but admitted some years later that he had failed.
+Since it was first televised in 1990, PMQs has been seen as a key barometer of a leader's judgement, their command of the Commons and their standing among their fellow MPs although critics have argued it has become a caricature and is in need of far-reaching reforms.
+'Shot in Joburg': Homeless youth trained as photographers
+Downtown Johannesburg is a tough place to be homeless.
+But one group of former street children have found a way to learn a skill and make a living.
+"I was shot in Joburg" is a non-profit studio that teaches homeless youngsters how to take photographs of their neighbourhood and make a profit from it.
+BBC News went to meet one of the project's first graduates.
+JD Sports boss says higher wages could hurt expansion
+JD Sports Executive Chairman Peter Cowgill says a higher minimum wage for UK workers could mean "more spending power in the pockets of potential consumers."
+But that spending power is unlikely to outweigh the higher labour costs at his firm, he says.
+The costs could hit JD Sports' expansion plans, he added, which could mean fewer extra jobs.
+Thanasi Kokkinakis backed by Tennis Australia president Steve Healy
+Thanasi Kokkinakis deserves kudos rather than criticism for his behaviour.
+Thanasi Kokkinakis has been the collateral damage in the recent storm around his friend Nick Kyrgios and deserves kudos rather than criticism for his own behaviour, according to Tennis Australia president Steve Healy.
@@ -1,20 +1,100 @@
-Șeful ONU declară că nu există soluții militare în Siria Secretarul General Ban Ki-moon afirmă că răspunsul său la suportul militar al Rusiei pentru Siria este că „nu există o soluție militară” la conflictul care durează de aproape cinci ani iar mai multe arme nu ar face decât să agraveze violența și suferința a milioane de oameni. Șeful ONU a solicitat din nou tuturor părților, inclusiv Consiliului de securitate ONU divizat să se unifice și să susțină negocierile pentru a găsi o soluție politică. Ban a declarat miercuri în cadrul unei conferințe că intenționează să se întâlnească luna aceasta cu miniștrii de externe din cinci țări permanent prezente în consiliu - SUA, Rusia, China, Anglia și Franța - pe marginea sesiunii ministeriale a Adunării Generale pentru a discuta despre Siria.
-Ban și-a exprimat regretul că divizările în consiliu și între poporul sirian și puterile regionale „au făcut această situație de nerezolvat”. Ban le-a cerut celor cinci membri permanenți să dea dovadă de solidaritatea și unitatea arătate atunci când au reușit să încheie un acord referitor la armele nucleare ale Iranului, abordând astfel criza din Siria. 8 cifre din sondaje care arată că Donald Trump are șanse reale Unii au încercat să îl eticheteze ca politician „flip-flop”. Alții l-au numit o glumă. Iar alții așteaptă implozia. Însă indiferent de modul în care unii republicani încearcă să îl dărâme pe Donald Trump din vârful sondajelor, nu a funcționat (încă).
-Zece din ultimele 11 sondaje naționale au arătat că Donald Trump conduce cu un procent din două cifre iar unele voci încep să se întrebe serios ce înseamnă acest lucru pentru șansele de numire ale mogulului imobiliar. Desigur, este încă prematur. Nimic din toate acestea nu spune că Trump va câștiga cursa pentru nominalizarea republicanilor. Pundits arată că, în aceeași perioadă a anului 2011, avansul lui Rick Perry îi făcea loc lui Herman Cain în sondaje, dar niciunul dintre ei nu a câștigat în vreun stat în cursa de nominalizare. Iar motivele pentru care s-ar lupta din greu la alegerile generale sunt numeroase. Însă grupurile din exterior precum Super PAC al lui Jeb Bush și grupul conservator economic Club for Growth admit puterea lui Trump și încep să îl susțină cu bani.
-În continuare vă prezentăm câteva cifre din sondaje recente care sugerează că mogulul imobiliar nu este doar ceva trecător: Cifrele care indică susținerea față de Trump s-au întors la 180 grade. Chiar înainte ca Donald Trump să își anunțe candidatura, la mijlocul lui iunie, un sondaj realizat de Universitatea din Monmouth arăta că doar doi din 10 republicani aveau o părere pozitivă despre mogulul imobiliar. Până la mijlocul lui iulie, procentul a urcat la 40%. La începutul lui august, era 52%. În prezent, șase din 10 republicani au o părere favorabilă despre Donald Trump. Aproximativ trei din 10 declară că au o părere negativă. Aceste cifre se mențin. Un sondaj realizat săptămâna trecută de Quinnipiac în Iowa a concluzionat că 60% dintre republicanii din regiune au o părere favorabilă despre Trump.
-Două treimi dintre alegătorii GOP ar fi fericiți dacă Trump ar câștiga cursa pentru nominalizare. Într-un sondaj realizat săptămâna trecută de CNN/ORC, 67% dintre republicani au declarat că ar fi „entuziasmați” sau „mulțumiți” dacă Trump ar câștiga cursa pentru nominalizare. Doar doi din 10 declară că ar fi „supărați” dacă Trump ar câștiga cursa pentru nominalizare. Doar Ben Carson generează aproximativ același nivel de entuziasm ca Trump (43% declară că ar fi „entuziasmați” față de 40% care declară același lucru despre Trump). Cel mai aproape în ceea ce privește entuziasmul? Marco Rubio, cu doar 21%.
-De partea cealaltă, 47% dintre alegătorii republicani afirmă că ar fi „nemulțumiți” sau „supărați” dacă favoritul Jeb Bush câștigă cursa pentru nominalizare. Majoritatea republicanilor nu consideră temperamentul lui Trump o problemă. Deși Donald Trump a fost puternic criticat pentru insultele aduse și stilul său bombastic, 52% dintre alegătorii republicani la nivel național consideră că mogulul imobiliar are temperamentul potrivit pentru a fi președinte, conform sondajului realizat luni de ABC News/Washington Post. Regăsim aceleași cifre în statul Iowa, unde tot 52% dintre republicani cred că Trump are personalitatea potrivită pentru a fi conducător, conform sondajului realizat săptămâna trecută de Quinnipiac.
-Totuși, 44% sunt de părere că nu are personalitatea necesară pentru a acționa eficient și aproape șase din 10 independenți afirmă că temperamentul său nu are ce căuta la Casa Albă, conform ABC/Post. Alegătorii republicani se obișnuiesc cu ideea. Atunci când iau atitudinea de intelectuali, alegătorii republicani consideră că Trump este autentic. Conform unui sondaj realizat săptămâna trecută de CNN/ORC, la întrebarea cine are cele mai multe șanse să câștige cursa pentru nominalizare GOP, patru din 10 au declarat că Trump. Situația s-a schimbat față de finalul lui iulie, când patru din 10 ar fi pariat pe Jeb Bush. Informare completă: în trecut, alegătorii GOP nu au citit foarte bine viitorul.
-În aceeași perioadă a ultimelor alegeri, patru din 10 republicani l-au ales pe Rick Perry în cursa pentru nominalizare, față de doar 28% pentru Mitt Romney. Însă, aceste cifre arată că majoritatea alegătorilor GOP consideră plauzibilă campania lui Trump. Chiar dacă republicanii sau repliat spre un alt candidat. Trump încă se află în fruntea tuturor. Unele voci spun că situația divizată va contribui probabil la victoria lui Trump, în timp ce susținerea contra lui Trump se va împărți la mai mult de doisprezece candidați. Însă un sondaj derulat la începutul lui septembrie de Universitatea din Monmouth arată că, în situația ipotetică a unei colaborări între Trump și majoritatea celorlalți candidați republicani, aproape întotdeauna Trump va beneficia de susținerea majoritară.
-Trump se află la distanță de 13 puncte de Carly Fiorina, la 14 puncte de Marco Rubio, la 15 puncte de Walker, la 19 puncte de Jeb Bush și, în cele din urmă, la câte 33 de puncte față de Rand Paul, John Kasich și Chris Christie. Este aproape la egalitate cu Ted Cruz. Singurul candidat care îl învinge? Ben Carson l-ar învinge pe omul de afaceri cu 19 puncte într-o confruntare ipotetică de unu la unu. Majoritatea susținătorilor lui Donald Trump declară că s-au decis. Un nou sondaj realizat marți de CBS/NYT arată că peste jumătate dintre alegătorii care îl susțin pe Trump declară că nu își schimbă opțiunea de vot. Evident, se pot întâmpla multe în acest sens și nimeni nu poate spune că aceștia nu se vor răzgândi niciodată.
-46% afirmă că lasă portița deschisă posibilității de a-și schimba opțiunea. Cu toate acestea, cel mai important adversar al lui Trump este în prezent neurochirurgul Ben Carson, însă este de două ori mai probabil ca alegătorii care declară că s-au decis să voteze cu Trump. Șase din 10 republicani afirmă că sunt de acord cu Trump în problema imigrării. De când Donald Trump i-a numit pe imigranții din Mexic „violatori” în discursul de deschidere a campaniei sale, în urmă cu două luni, imigrarea a fost subiectul central în campania pentru 2016. Unii sunt îngrijorați că stilul bombastic al lui Trump va duce la o scindare între alegătorii hispanici importanți și Partidul Republican și va prejudicia eforturile de rebranding.
-Însă, conform sondajului realizat luni de ABC/Post, șase din 10 republicani afirmă că sunt de acord cu Trump în problema imigrării. Așa că, se pare că atâta timp cât problema imigrării rămâne în lumina reflectoarelor, la fel va rămâne și Doland Trump. Frustrarea față de autorități atinge noi culmi. Donald Trump și Ben Carson sunt acum susținuți de aproape jumătate dintre alegătorii republicani, în mare parte datorită statutului lor de outsideri. Conform sondajului realizat luni de ABC/Post, șase din 10 republicani afirmă că preferă un outsider politic în detrimentul cuiva cu experiență în guvernare. Oamenii sunt de asemenea supărați pe autoritățile de la Washington.
-Un sondaj derulat în urmă cu două săptămâni în Iowa de către Des Moines Register/Bloomberg arată că trei din patru republicani din Iowa sunt frustrați de prestația republicanilor din COngres, 54% declarându-se „nemulțumiți” iar 21% „nervoși la culme”. Jeremy Corbyn își face debutul la Prime Minister's Questions Încă de la alegerea sa, debutul domnului Corbyn la PMQs a fost îndelung așteptat Noul lider al Partidului Laburist, Jeremy Corbyn, își va face mai târziu debutul la Prime Minister's Questions, confruntându-se pentru prima dată cu David Cameron.
-Dl Corbyn va adresa primele dintre cele șase întrebări la care are dreptul la scurt timp după prânz; prestația sa va fi probabil analizată îndeaproape de mass-media și parlamentarii laburiști. În cadrul aparițiilor săptămânale, el a cerut „mai puțin teatru și mai multe fapte”. A declarat de asemenea că poate renunța la câteva participări și că le cedează colegilor săi. Confruntarea va fi primul test parlamentar al Dl Corbyn în poziție de lider, venind după ce a numit un „cabinet fantomă” și după discursul pe care l-a ținut marți la congresul anual TUC.
-Între timp, decizia liderului Partidului laburist de a păstra tăcerea la rostirea imnului național în cadrul unei slujbe ținute marți cu ocazia aniversării a 75 de ani de la Bătălia Angliei a atras critici din partea unor parlamentari conservatori și a ținut prima pagină a ziarelor. Decizia domnului Corbyn de a nu cânta imnul național a atras atenția Un purtător de cuvânt al Dl Corbyn a declarat că acesta „a păstrat tăcerea în mod respectuos” și a recunoscut „eroismul Forțelor aeriene britanice în Bătălia Angliei.”
-Însă un membru al cabinetului fantomă al Dl Corbyn, Owen Smith, a declarat pentru emisiunea Two's Newsnight transmisă de BBC că i-ar fi recomandat liderului laburist să cânte imnul național „indiferent” de credința sa că monarhia ar trebui abolită. În jur de doisprezece miniștri din cabinetul fantomă au refuzat să facă parte din echipa de frunte a Dl Corbyn, argumentând prin diferențe de opinie legate de economie, apărare și externe, în timp ce mai puțin de o șesime din partidul parlamentar l-a susținut ca lider. Corespondentul politic al BBC, Robin Brant, declară că diferențele de politică „se cumulează” în Partidul Laburist după numirea domnului Corbyn referitor la poziția sa față de Uniunea Europeană și limita de beneficii.
-Dl Corbyn a declarat la conferința TUC că Partidul Laburist va aduce modificări prin care se va elimina integral ideea limitării. Câteva ore mai târziu, Dl Smith, Ministrul Muncii și Pensiilor, a declarat că partidul „este foarte clar” în opoziția exclusivă față de planurile guvernului de a reduce nivelul „cap” de la 26.000 lire la 23.000 lire. Dl Corbyn va fi al cincilea lider laburist cu care se confruntă David Cameron la tribună în ultimul deceniu, de când a preluat conducerea Partidului Conservator. Liderul laburist, care a promis o abordare diferită a politicii, spune că are idei „din surse externe” pentru întrebări pe care să i le adreseze Domnului Cameron și că a primit peste 30.000 de sugestii.
-Parlamentarul Islington North a afirmat că PMQs implică un nivel de confruntare prea înalt și că se va abține de la replici și atacuri, angajându-se să se concentreze în schimb pe probleme serioase precum sărăcia, inegalitatea și provocările cu care se confruntă tinerii. Dl Corbyn a declarat că Angela Eagle, Ministrul de finanțe, îi va ține locul la PMQs atunci când el nu poate participa - de exemplu atunci când Dl Cameron se deplasează în străinătate. A exprimat de asemenea ideea că va permite altor colegi să ia cuvântul ocazional, spunând că l-a abordat pe Președintele Camerei Deputaților, John Bercow, pentru a discuta acest aspect.
-În 2005, când a preluat conducerea, Dl Cameron a declarat că dorește să renunțe la stilul politic „Punch and Judy” asociat adesea cu PMQs însă a recunoscut câțiva ani mai târziu că nu a reușit în demersul său. De la prima transmisie, în 1990, PMQs a fost considerată un barometru cheie al raționamentului unui lider, al modului în care acesta conduce Camera Deputaților și a poziției sale în rândul colegilor parlamentari, deși criticii afirmă a ca devenit o caricatură și că are nevoie de o reformare profundă. „Cadru în Joburg”: Tineri fără adăpost beneficiază de cursuri de fotografie Este dificil să fii un om fără adăpost în Johannesburg.
-Însă un grup de oameni care au trăit pe străzi în copilărie au găsit un mod de a învăța o meserie și de a-și câștiga traiul. „I was shot în Joburg” este un studio non-profit care îi învață pe tinerii fără adăpost să facă fotografii ale zonelor în care trăiesc și să câștige bani din asta. BBC News s-a întâlnit cu unul dintre primii absolvenți ai proiectului. Șeful JD Sports spune că salariile mai mari ar putea dăuna extinderii Președintele JD Sports, Peter Cowgill, declară că o creștere a salariului minim în Marea Britanie ar putea însemna „o putere de cumpărare mai mare în buzunarele potențialilor consumatori.” Este însă puțin probabil ca respectiva putere de cumpărare să depășească costurile mai mari pentru forța de muncă în cadrul firmei, afirmă el.
-Costurile ar putea avea impact asupra planurilor de extindere ale JD Sports, a adăugat el, ceea ce ar putea însemna mai puține locuri de muncă noi. Thanasi Kokkinakis susținut de președintele Tennis Australia, Steve Healy Thanasi Kokkinakis ar merita să fie lăudat și nu criticat pentru comportamentul său. Thanasi Kokkinakis a fost victimă colaterală în „furtuna” creată în jurul prietenului său, Nick Kyrgios, iar comportamentul său merită mai degrabă cuvinte de laudă și nu critică, în opinia președintelui Tennis Australia, Steve Healy.
+Șeful ONU declară că nu există soluții militare în Siria
+Secretarul General Ban Ki-moon afirmă că răspunsul său la suportul militar al Rusiei pentru Siria este că „nu există o soluție militară” la conflictul care durează de aproape cinci ani iar mai multe arme nu ar face decât să agraveze violența și suferința a milioane de oameni.
+Șeful ONU a solicitat din nou tuturor părților, inclusiv Consiliului de securitate ONU divizat să se unifice și să susțină negocierile pentru a găsi o soluție politică.
+Ban a declarat miercuri în cadrul unei conferințe că intenționează să se întâlnească luna aceasta cu miniștrii de externe din cinci țări permanent prezente în consiliu - SUA, Rusia, China, Anglia și Franța - pe marginea sesiunii ministeriale a Adunării Generale pentru a discuta despre Siria.
+Ban și-a exprimat regretul că divizările în consiliu și între poporul sirian și puterile regionale „au făcut această situație de nerezolvat”.
+Ban le-a cerut celor cinci membri permanenți să dea dovadă de solidaritatea și unitatea arătate atunci când au reușit să încheie un acord referitor la armele nucleare ale Iranului, abordând astfel criza din Siria.
+8 cifre din sondaje care arată că Donald Trump are șanse reale
+Unii au încercat să îl eticheteze ca politician „flip-flop”.
+Alții l-au numit o glumă.
+Iar alții așteaptă implozia.
+Însă indiferent de modul în care unii republicani încearcă să îl dărâme pe Donald Trump din vârful sondajelor, nu a funcționat (încă).
+Zece din ultimele 11 sondaje naționale au arătat că Donald Trump conduce cu un procent din două cifre iar unele voci încep să se întrebe serios ce înseamnă acest lucru pentru șansele de numire ale mogulului imobiliar.
+Desigur, este încă prematur.
+Nimic din toate acestea nu spune că Trump va câștiga cursa pentru nominalizarea republicanilor.
+Pundits arată că, în aceeași perioadă a anului 2011, avansul lui Rick Perry îi făcea loc lui Herman Cain în sondaje, dar niciunul dintre ei nu a câștigat în vreun stat în cursa de nominalizare.
+Iar motivele pentru care s-ar lupta din greu la alegerile generale sunt numeroase.
+Însă grupurile din exterior precum Super PAC al lui Jeb Bush și grupul conservator economic Club for Growth admit puterea lui Trump și încep să îl susțină cu bani.
+În continuare vă prezentăm câteva cifre din sondaje recente care sugerează că mogulul imobiliar nu este doar ceva trecător:
+Cifrele care indică susținerea față de Trump s-au întors la 180 grade.
+Chiar înainte ca Donald Trump să își anunțe candidatura, la mijlocul lui iunie, un sondaj realizat de Universitatea din Monmouth arăta că doar doi din 10 republicani aveau o părere pozitivă despre mogulul imobiliar.
+Până la mijlocul lui iulie, procentul a urcat la 40%.
+La începutul lui august, era 52%.
+În prezent, șase din 10 republicani au o părere favorabilă despre Donald Trump.
+Aproximativ trei din 10 declară că au o părere negativă.
+Aceste cifre se mențin.
+Un sondaj realizat săptămâna trecută de Quinnipiac în Iowa a concluzionat că 60% dintre republicanii din regiune au o părere favorabilă despre Trump.
+Două treimi dintre alegătorii GOP ar fi fericiți dacă Trump ar câștiga cursa pentru nominalizare.
+Într-un sondaj realizat săptămâna trecută de CNN/ORC, 67% dintre republicani au declarat că ar fi „entuziasmați” sau „mulțumiți” dacă Trump ar câștiga cursa pentru nominalizare.
+Doar doi din 10 declară că ar fi „supărați” dacă Trump ar câștiga cursa pentru nominalizare.
+Doar Ben Carson generează aproximativ același nivel de entuziasm ca Trump (43% declară că ar fi „entuziasmați” față de 40% care declară același lucru despre Trump).
+Cel mai aproape în ceea ce privește entuziasmul?
+Marco Rubio, cu doar 21%.
+De partea cealaltă, 47% dintre alegătorii republicani afirmă că ar fi „nemulțumiți” sau „supărați” dacă favoritul Jeb Bush câștigă cursa pentru nominalizare.
+Majoritatea republicanilor nu consideră temperamentul lui Trump o problemă.
+Deși Donald Trump a fost puternic criticat pentru insultele aduse și stilul său bombastic, 52% dintre alegătorii republicani la nivel național consideră că mogulul imobiliar are temperamentul potrivit pentru a fi președinte, conform sondajului realizat luni de ABC News/Washington Post.
+Regăsim aceleași cifre în statul Iowa, unde tot 52% dintre republicani cred că Trump are personalitatea potrivită pentru a fi conducător, conform sondajului realizat săptămâna trecută de Quinnipiac.
+Totuși, 44% sunt de părere că nu are personalitatea necesară pentru a acționa eficient și aproape șase din 10 independenți afirmă că temperamentul său nu are ce căuta la Casa Albă, conform ABC/Post.
+Alegătorii republicani se obișnuiesc cu ideea.
+Atunci când iau atitudinea de intelectuali, alegătorii republicani consideră că Trump este autentic.
+Conform unui sondaj realizat săptămâna trecută de CNN/ORC, la întrebarea cine are cele mai multe șanse să câștige cursa pentru nominalizare GOP, patru din 10 au declarat că Trump.
+Situația s-a schimbat față de finalul lui iulie, când patru din 10 ar fi pariat pe Jeb Bush.
+Informare completă: în trecut, alegătorii GOP nu au citit foarte bine viitorul.
+În aceeași perioadă a ultimelor alegeri, patru din 10 republicani l-au ales pe Rick Perry în cursa pentru nominalizare, față de doar 28% pentru Mitt Romney.
+Însă, aceste cifre arată că majoritatea alegătorilor GOP consideră plauzibilă campania lui Trump.
+Chiar dacă republicanii sau repliat spre un alt candidat. Trump încă se află în fruntea tuturor.
+Unele voci spun că situația divizată va contribui probabil la victoria lui Trump, în timp ce susținerea contra lui Trump se va împărți la mai mult de doisprezece candidați.
+Însă un sondaj derulat la începutul lui septembrie de Universitatea din Monmouth arată că, în situația ipotetică a unei colaborări între Trump și majoritatea celorlalți candidați republicani, aproape întotdeauna Trump va beneficia de susținerea majoritară.
+Trump se află la distanță de 13 puncte de Carly Fiorina, la 14 puncte de Marco Rubio, la 15 puncte de Walker, la 19 puncte de Jeb Bush și, în cele din urmă, la câte 33 de puncte față de Rand Paul, John Kasich și Chris Christie.
+Este aproape la egalitate cu Ted Cruz.
+Singurul candidat care îl învinge?
+Ben Carson l-ar învinge pe omul de afaceri cu 19 puncte într-o confruntare ipotetică de unu la unu.
+Majoritatea susținătorilor lui Donald Trump declară că s-au decis.
+Un nou sondaj realizat marți de CBS/NYT arată că peste jumătate dintre alegătorii care îl susțin pe Trump declară că nu își schimbă opțiunea de vot.
+Evident, se pot întâmpla multe în acest sens și nimeni nu poate spune că aceștia nu se vor răzgândi niciodată.
+46% afirmă că lasă portița deschisă posibilității de a-și schimba opțiunea.
+Cu toate acestea, cel mai important adversar al lui Trump este în prezent neurochirurgul Ben Carson, însă este de două ori mai probabil ca alegătorii care declară că s-au decis să voteze cu Trump.
+Șase din 10 republicani afirmă că sunt de acord cu Trump în problema imigrării.
+De când Donald Trump i-a numit pe imigranții din Mexic „violatori” în discursul de deschidere a campaniei sale, în urmă cu două luni, imigrarea a fost subiectul central în campania pentru 2016.
+Unii sunt îngrijorați că stilul bombastic al lui Trump va duce la o scindare între alegătorii hispanici importanți și Partidul Republican și va prejudicia eforturile de rebranding.
+Însă, conform sondajului realizat luni de ABC/Post, șase din 10 republicani afirmă că sunt de acord cu Trump în problema imigrării.
+Așa că, se pare că atâta timp cât problema imigrării rămâne în lumina reflectoarelor, la fel va rămâne și Doland Trump.
+Frustrarea față de autorități atinge noi culmi.
+Donald Trump și Ben Carson sunt acum susținuți de aproape jumătate dintre alegătorii republicani, în mare parte datorită statutului lor de outsideri.
+Conform sondajului realizat luni de ABC/Post, șase din 10 republicani afirmă că preferă un outsider politic în detrimentul cuiva cu experiență în guvernare.
+Oamenii sunt de asemenea supărați pe autoritățile de la Washington.
+Un sondaj derulat în urmă cu două săptămâni în Iowa de către Des Moines Register/Bloomberg arată că trei din patru republicani din Iowa sunt frustrați de prestația republicanilor din COngres, 54% declarându-se „nemulțumiți” iar 21% „nervoși la culme”.
+Jeremy Corbyn își face debutul la Prime Minister's Questions
+Încă de la alegerea sa, debutul domnului Corbyn la PMQs a fost îndelung așteptat
+Noul lider al Partidului Laburist, Jeremy Corbyn, își va face mai târziu debutul la Prime Minister's Questions, confruntându-se pentru prima dată cu David Cameron.
+Dl Corbyn va adresa primele dintre cele șase întrebări la care are dreptul la scurt timp după prânz; prestația sa va fi probabil analizată îndeaproape de mass-media și parlamentarii laburiști.
+În cadrul aparițiilor săptămânale, el a cerut „mai puțin teatru și mai multe fapte”.
+A declarat de asemenea că poate renunța la câteva participări și că le cedează colegilor săi.
+Confruntarea va fi primul test parlamentar al Dl Corbyn în poziție de lider, venind după ce a numit un „cabinet fantomă” și după discursul pe care l-a ținut marți la congresul anual TUC.
+Între timp, decizia liderului Partidului laburist de a păstra tăcerea la rostirea imnului național în cadrul unei slujbe ținute marți cu ocazia aniversării a 75 de ani de la Bătălia Angliei a atras critici din partea unor parlamentari conservatori și a ținut prima pagină a ziarelor.
+Decizia domnului Corbyn de a nu cânta imnul național a atras atenția
+Un purtător de cuvânt al Dl Corbyn a declarat că acesta „a păstrat tăcerea în mod respectuos” și a recunoscut „eroismul Forțelor aeriene britanice în Bătălia Angliei.”
+Însă un membru al cabinetului fantomă al Dl Corbyn, Owen Smith, a declarat pentru emisiunea Two's Newsnight transmisă de BBC că i-ar fi recomandat liderului laburist să cânte imnul național „indiferent” de credința sa că monarhia ar trebui abolită.
+În jur de doisprezece miniștri din cabinetul fantomă au refuzat să facă parte din echipa de frunte a Dl Corbyn, argumentând prin diferențe de opinie legate de economie, apărare și externe, în timp ce mai puțin de o șesime din partidul parlamentar l-a susținut ca lider.
+Corespondentul politic al BBC, Robin Brant, declară că diferențele de politică „se cumulează” în Partidul Laburist după numirea domnului Corbyn referitor la poziția sa față de Uniunea Europeană și limita de beneficii.
+Dl Corbyn a declarat la conferința TUC că Partidul Laburist va aduce modificări prin care se va elimina integral ideea limitării.
+Câteva ore mai târziu, Dl Smith, Ministrul Muncii și Pensiilor, a declarat că partidul „este foarte clar” în opoziția exclusivă față de planurile guvernului de a reduce nivelul „cap” de la 26.000 lire la 23.000 lire.
+Dl Corbyn va fi al cincilea lider laburist cu care se confruntă David Cameron la tribună în ultimul deceniu, de când a preluat conducerea Partidului Conservator.
+Liderul laburist, care a promis o abordare diferită a politicii, spune că are idei „din surse externe” pentru întrebări pe care să i le adreseze Domnului Cameron și că a primit peste 30.000 de sugestii.
+Parlamentarul Islington North a afirmat că PMQs implică un nivel de confruntare prea înalt și că se va abține de la replici și atacuri, angajându-se să se concentreze în schimb pe probleme serioase precum sărăcia, inegalitatea și provocările cu care se confruntă tinerii.
+Dl Corbyn a declarat că Angela Eagle, Ministrul de finanțe, îi va ține locul la PMQs atunci când el nu poate participa - de exemplu atunci când Dl Cameron se deplasează în străinătate.
+A exprimat de asemenea ideea că va permite altor colegi să ia cuvântul ocazional, spunând că l-a abordat pe Președintele Camerei Deputaților, John Bercow, pentru a discuta acest aspect.
+În 2005, când a preluat conducerea, Dl Cameron a declarat că dorește să renunțe la stilul politic „Punch and Judy” asociat adesea cu PMQs însă a recunoscut câțiva ani mai târziu că nu a reușit în demersul său.
+De la prima transmisie, în 1990, PMQs a fost considerată un barometru cheie al raționamentului unui lider, al modului în care acesta conduce Camera Deputaților și a poziției sale în rândul colegilor parlamentari, deși criticii afirmă a ca devenit o caricatură și că are nevoie de o reformare profundă.
+„Cadru în Joburg”: Tineri fără adăpost beneficiază de cursuri de fotografie
+Este dificil să fii un om fără adăpost în Johannesburg.
+Însă un grup de oameni care au trăit pe străzi în copilărie au găsit un mod de a învăța o meserie și de a-și câștiga traiul.
+„I was shot în Joburg” este un studio non-profit care îi învață pe tinerii fără adăpost să facă fotografii ale zonelor în care trăiesc și să câștige bani din asta.
+BBC News s-a întâlnit cu unul dintre primii absolvenți ai proiectului.
+Șeful JD Sports spune că salariile mai mari ar putea dăuna extinderii
+Președintele JD Sports, Peter Cowgill, declară că o creștere a salariului minim în Marea Britanie ar putea însemna „o putere de cumpărare mai mare în buzunarele potențialilor consumatori.”
+Este însă puțin probabil ca respectiva putere de cumpărare să depășească costurile mai mari pentru forța de muncă în cadrul firmei, afirmă el.
+Costurile ar putea avea impact asupra planurilor de extindere ale JD Sports, a adăugat el, ceea ce ar putea însemna mai puține locuri de muncă noi.
+Thanasi Kokkinakis susținut de președintele Tennis Australia, Steve Healy
+Thanasi Kokkinakis ar merita să fie lăudat și nu criticat pentru comportamentul său.
+Thanasi Kokkinakis a fost victimă colaterală în „furtuna” creată în jurul prietenului său, Nick Kyrgios, iar comportamentul său merită mai degrabă cuvinte de laudă și nu critică, în opinia președintelui Tennis Australia, Steve Healy.
@@ -1,8 +1,100 @@
-Membership of Parliament: see Minutes Approval of Minutes of previous sitting: see Minutes Membership of Parliament: see Minutes Verification of credentials: see Minutes Documents received: see Minutes Written statements and oral questions (tabling): see Minutes Petitions: see Minutes Texts of agreements forwarded by the Council: see Minutes Action taken on Parliament's resolutions: see Minutes Agenda for next sitting: see Minutes Closure of sitting (The sitting was closed at 7.45 p.m.)
-Election of Vice-Presidents of the European Parliament (deadline for submitting nominations): see Minutes (The sitting was suspended at 12.40 p.m. and resumed at 3.00 p.m.) Election of Quaestors of the European Parliament (deadline for submitting nominations): see Minutes (The sitting was suspended at 3.25 p.m. and resumed at 6.00 p.m.) Agenda for next sitting: see Minutes Closure of sitting (The sitting was closed at 6.15 p.m.) Opening of the sitting (The sitting was opened at 9.35 a.m.) Documents received: see Minutes Approval of Minutes of previous sitting: see Minutes Membership of Parliament: see Minutes
-Membership of committees (deadline for tabling amendments): see Minutes (The sitting was suspended at 7 p.m. and resumed at 9 p.m.) Agenda for next sitting: see Minutes Closure of sitting (The sitting was suspended at 23.25 p.m.) Documents received: see Minutes Communication of Council common positions: see Minutes (The sitting was suspended at 11.35 a.m. and resumed for voting time at noon) Approval of Minutes of previous sitting: see Minutes Committee of Inquiry into the crisis of the Equitable Life Assurance Society (extension of mandate): see Minutes
-Announcement by the President: see Minutes 1. Membership of committees (vote) 2. Amendment of the ACP-EC Partnership Agreement (vote) 4. Certification of train drivers operating locomotives and trains on the railway system in the Community (vote) 6. Law applicable to non-contractual obligations ("ROME II") (vote) 8. Seventh and eighth annual reports on arms exports (vote) Corrections to votes and voting intentions: see Minutes Membership of committees and delegations: see Minutes Request for waiver of parliamentary immunity: see Minutes Decisions concerning certain documents: see Minutes
-Written statements for entry in the register (Rule 116): see Minutes Forwarding of texts adopted during the sitting: see Minutes Dates for next sittings: see Minutes Adjournment of the session I declare the session of the European Parliament adjourned. (The sitting was closed at 1 p.m.) Approval of Minutes of previous sitting: see Minutes Membership of Parliament: see Minutes Request for the defence of parliamentary immunity: see Minutes Appointments to committees (proposal by the Conference of Presidents): see Minutes Documents received: see Minutes Texts of agreements forwarded by the Council: see Minutes
-Action taken on Parliament's resolutions: see Minutes Oral questions and written statements (tabling): see Minutes Written statements (Rule 116): see Minutes Agenda: see Minutes 1. Appointments to parliamentary committees (vote): see Minutes Voting time Agenda for next sitting: see Minutes Closure of sitting (The sitting was closed at 12 midnight) Opening of the sitting (The sitting was opened at 09.05) Documents received: see Minutes Approval of Minutes of previous sitting: see Minutes 1. Protection of passengers against displaced luggage (vote) 2.
-Approval of motor vehicles with regard to the forward field of vision of the driver (vote) 3. EC-Korea Agreement on scientific and technological cooperation (vote) 4. Mainstreaming sustainability in development cooperation policies (vote) 5. Draft Amending Budget No 1/2007 (vote) 7. EC-Gabon Fisheries Partnership (vote) 10. Limitation periods in cross-border disputes involving personal injuries and fatal accidents (vote) 12. Strategy for a strengthened partnership with the Pacific Islands (vote) 13. The European private company statute (vote) That concludes the vote.
-Corrections to votes and voting intentions: see Minutes Assignment conferred on a Member: see Minutes Membership of committees and delegations: see Minutes Decisions concerning certain documents: see Minutes Forwarding of texts adopted during the sitting: see Minutes Dates for next sittings: see Minutes
+Membership of Parliament: see Minutes
+Approval of Minutes of previous sitting: see Minutes
+Membership of Parliament: see Minutes
+Verification of credentials: see Minutes
+Documents received: see Minutes
+Written statements and oral questions (tabling): see Minutes
+Petitions: see Minutes
+Texts of agreements forwarded by the Council: see Minutes
+Action taken on Parliament's resolutions: see Minutes
+Agenda for next sitting: see Minutes
+Closure of sitting
+(The sitting was closed at 7.45 p.m.)
+Election of Vice-Presidents of the European Parliament (deadline for submitting nominations): see Minutes
+(The sitting was suspended at 12.40 p.m. and resumed at 3.00 p.m.)
+Election of Quaestors of the European Parliament (deadline for submitting nominations): see Minutes
+(The sitting was suspended at 3.25 p.m. and resumed at 6.00 p.m.)
+Agenda for next sitting: see Minutes
+Closure of sitting
+(The sitting was closed at 6.15 p.m.)
+Opening of the sitting
+(The sitting was opened at 9.35 a.m.)
+Documents received: see Minutes
+Approval of Minutes of previous sitting: see Minutes
+Membership of Parliament: see Minutes
+Membership of committees (deadline for tabling amendments): see Minutes
+(The sitting was suspended at 7 p.m. and resumed at 9 p.m.)
+Agenda for next sitting: see Minutes
+Closure of sitting
+(The sitting was suspended at 23.25 p.m.)
+Documents received: see Minutes
+Communication of Council common positions: see Minutes
+(The sitting was suspended at 11.35 a.m. and resumed for voting time at noon)
+Approval of Minutes of previous sitting: see Minutes
+Committee of Inquiry into the crisis of the Equitable Life Assurance Society (extension of mandate): see Minutes
+Announcement by the President: see Minutes
+1.
+Membership of committees (vote)
+2.
+Amendment of the ACP-EC Partnership Agreement (vote)
+4.
+Certification of train drivers operating locomotives and trains on the railway system in the Community (vote)
+6.
+Law applicable to non-contractual obligations ("ROME II") (vote)
+8.
+Seventh and eighth annual reports on arms exports (vote)
+Corrections to votes and voting intentions: see Minutes
+Membership of committees and delegations: see Minutes
+Request for waiver of parliamentary immunity: see Minutes
+Decisions concerning certain documents: see Minutes
+Written statements for entry in the register (Rule 116): see Minutes
+Forwarding of texts adopted during the sitting: see Minutes
+Dates for next sittings: see Minutes
+Adjournment of the session
+I declare the session of the European Parliament adjourned.
+(The sitting was closed at 1 p.m.)
+Approval of Minutes of previous sitting: see Minutes
+Membership of Parliament: see Minutes
+Request for the defence of parliamentary immunity: see Minutes
+Appointments to committees (proposal by the Conference of Presidents): see Minutes
+Documents received: see Minutes
+Texts of agreements forwarded by the Council: see Minutes
+Action taken on Parliament's resolutions: see Minutes
+Oral questions and written statements (tabling): see Minutes
+Written statements (Rule 116): see Minutes
+Agenda: see Minutes
+1.
+Appointments to parliamentary committees (vote): see Minutes
+Voting time
+Agenda for next sitting: see Minutes
+Closure of sitting
+(The sitting was closed at 12 midnight)
+Opening of the sitting
+(The sitting was opened at 09.05)
+Documents received: see Minutes
+Approval of Minutes of previous sitting: see Minutes
+1.
+Protection of passengers against displaced luggage (vote)
+2.
+Approval of motor vehicles with regard to the forward field of vision of the driver (vote)
+3.
+EC-Korea Agreement on scientific and technological cooperation (vote)
+4.
+Mainstreaming sustainability in development cooperation policies (vote)
+5.
+Draft Amending Budget No 1/2007 (vote)
+7.
+EC-Gabon Fisheries Partnership (vote)
+10.
+Limitation periods in cross-border disputes involving personal injuries and fatal accidents (vote)
+12.
+Strategy for a strengthened partnership with the Pacific Islands (vote)
+13.
+The European private company statute (vote)
+That concludes the vote.
+Corrections to votes and voting intentions: see Minutes
+Assignment conferred on a Member: see Minutes
+Membership of committees and delegations: see Minutes
+Decisions concerning certain documents: see Minutes
+Forwarding of texts adopted during the sitting: see Minutes
+Dates for next sittings: see Minutes
@@ -1,8 +1,100 @@
-Componenţa Parlamentului: a se vedea procesul-verbal Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal Componenţa Parlamentului: a se vedea procesul-verbal Verificarea prerogativelor: a se vedea procesul-verbal Depunere de documente: a se vedea procesul-verbal Declaraţii scrise şi întrebări orale (depunere): consultaţi procesul-verbal Petiţii: a se vedea procesul-verbal Transmiterea de către Consiliu a textelor acordurilor: a se vedea procesul-verbal Cursul dat rezoluţiilor Parlamentului: a se vedea procesul-verbal Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal Ridicarea şedinţei (Se levanta la sesión a las 19.45 horas)
-Alegerea vicepreşedinţilor Parlamentului European (termenul de depunere a candidaturilor): consultaţi procesul-verbal (Die Sitzung wird um 12.40 Uhr unterbrochen und um 15.00 Uhr wiederaufgenommen). Alegerea chestorilor Parlamentului European (termenul de depunere a candidaturilor): consultaţi procesul-verbal (Die Sitzung wird um 15.25 Uhr unterbrochen und um 18.00 Uhr wiederaufgenommen). Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal Ridicarea şedinţei (Die Sitzung wird um 18.15 Uhr geschlossen.) Deschiderea şedinţei (Die Sitzung wird um 9.35 Uhr eröffnet.) Depunerea documentelor: a se vedea procesul-verbal Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal Componenţa Parlamentului: a se vedea procesul-verbal
-Componenţa comisiilor (termenul de depunere a amendamentelor): consultaţi procesul-verbal (La seduta, sospesa alle 19.00, è ripresa alle 21.00) Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal Ridicarea şedinţei (Die Sitzung wird um 23.25 Uhr geschlossen.) Depunerea documentelor: a se vedea procesul-verbal Comunicarea poziţiilor comune ale Parlamentului: a se vedea procesul-verbal (La séance, suspendue à 11h35 dans l'attente de l'Heure des votes, est reprise à midi) Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal Comisia de anchetă privind criza societăţii de asigurări "Equitable Life” (prelungirea mandatului): consultaţi procesul-verbal
-Comunicarea Preşedintelui: consultaţi procesul-verbal 1. Componenţa comisiilor (vot) 2. Modificarea Acordului de parteneriat ACP-CE ("Acordul de la Cotonou”) (vot) 4. Certificarea mecanicilor de locomotivă care conduc locomotive şi trenuri în sistemul feroviar comunitar (vot) 6. Legea aplicabilă obligaţiilor necontractuale ("Roma II”) (vot) 8. Al şaptelea şi al optulea raport anual privind exportul de armament (vot) Corectările voturilor şi intenţiile de vot: a se vedea procesul-verbal Componenţa comisiilor şi a delegaţiilor: a se vedea procesul-verbal Cerere de ridicare a imunităţii parlamentare: consultaţi procesul-verbal Decizii privind anumite documente: a se vedea procesul-verbal
-Declaraţii scrise înscrise în registru (articolul 116 din Regulamentul de procedură): a se vedea procesul-verbal Transmiterea textelor adoptate în cursul prezentei şedinţe: a se vedea procesul-verbal Calendarul următoarelor şedinţe: a se vedea procesul-verbal Întreruperea sesiunii Dichiaro interrotta la sessione del Parlamento europeo. (La seduta è tolta alle 13.00) Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal Componenţa Parlamentului: a se vedea procesul-verbal Cerere de apărare a imunităţii parlamentare: consultaţi procesul-verbal Numiri în comisii (propunerea Conferinţei preşedinţilor): consultaţi procesul-verbal Depunerea documentelor: a se vedea procesul-verbal Transmiterea de către Consiliu a textelor acordurilor: a se vedea procesul-verbal
-Continuări ale rezoluţiilor Parlamentului: consultaţi procesul-verbal Declaraţii scrise şi întrebări orale (depunere): consultaţi procesul-verbal Declaraţii scrise (articolul 116 din Regulamentul de procedură) Ordinea de zi: a se vedea procesul-verbal 1. Numiri în comisiile parlamentare (vot): consultaţi procesul-verbal Timpul afectat votului Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal Ridicarea şedinţei (La seduta è tolta alle 24.00) Deschiderea şedinţei (The sitting was opened at 09.05) Depunerea documentelor: a se vedea procesul-verbal Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal 1. Protecţia pasagerilor împotriva deplasării bagajelor (vot) 2.
-Omologarea vehiculelor cu motor cu privire la câmpul de vizibilitate înainte al conducătorului auto (vot) 3. Acordul CE-Coreea de cooperare ştiinţifică şi tehnologică (vot) 4. Integrarea durabilităţii în politicile de cooperare pentru dezvoltare (vot) 5. Proiect de buget rectificativ nr.1/2007 (vot) 7. Acordul de parteneriat în domeniul pescuitului între Comunitatea Europeană şi Republica Gaboneză (vot) 10. Termenele de prescripţie aplicabile în cadrul litigiilor transfrontaliere cu privire la vătămările corporale şi accidentele mortale (vot) 12. Relaţiile UE cu insulele din Pacific: Strategie pentru un parteneriat consolidat (vot) 13. Statutul societăţii private europene (vot) Damit ist die Abstimmungsstunde beendet.
-Corectările voturilor şi intenţiile de vot: a se vedea procesul-verbal Misiune încredinţată unui deputat: consultaţi procesul-verbal Componenţa comisiilor şi a delegaţiilor: a se vedea procesul-verbal Decizii privind anumite documente: a se vedea procesul-verbal Transmiterea textelor adoptate în cursul prezentei şedinţe: a se vedea procesul-verbal Calendarul următoarelor şedinţe: a se vedea procesul-verbal
+Componenţa Parlamentului: a se vedea procesul-verbal
+Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal
+Componenţa Parlamentului: a se vedea procesul-verbal
+Verificarea prerogativelor: a se vedea procesul-verbal
+Depunere de documente: a se vedea procesul-verbal
+Declaraţii scrise şi întrebări orale (depunere): consultaţi procesul-verbal
+Petiţii: a se vedea procesul-verbal
+Transmiterea de către Consiliu a textelor acordurilor: a se vedea procesul-verbal
+Cursul dat rezoluţiilor Parlamentului: a se vedea procesul-verbal
+Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal
+Ridicarea şedinţei
+(Se levanta la sesión a las 19.45 horas)
+Alegerea vicepreşedinţilor Parlamentului European (termenul de depunere a candidaturilor): consultaţi procesul-verbal
+(Die Sitzung wird um 12.40 Uhr unterbrochen und um 15.00 Uhr wiederaufgenommen).
+Alegerea chestorilor Parlamentului European (termenul de depunere a candidaturilor): consultaţi procesul-verbal
+(Die Sitzung wird um 15.25 Uhr unterbrochen und um 18.00 Uhr wiederaufgenommen).
+Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal
+Ridicarea şedinţei
+(Die Sitzung wird um 18.15 Uhr geschlossen.)
+Deschiderea şedinţei
+(Die Sitzung wird um 9.35 Uhr eröffnet.)
+Depunerea documentelor: a se vedea procesul-verbal
+Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal
+Componenţa Parlamentului: a se vedea procesul-verbal
+Componenţa comisiilor (termenul de depunere a amendamentelor): consultaţi procesul-verbal
+(La seduta, sospesa alle 19.00, è ripresa alle 21.00)
+Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal
+Ridicarea şedinţei
+(Die Sitzung wird um 23.25 Uhr geschlossen.)
+Depunerea documentelor: a se vedea procesul-verbal
+Comunicarea poziţiilor comune ale Parlamentului: a se vedea procesul-verbal
+(La séance, suspendue à 11h35 dans l'attente de l'Heure des votes, est reprise à midi)
+Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal
+Comisia de anchetă privind criza societăţii de asigurări "Equitable Life” (prelungirea mandatului): consultaţi procesul-verbal
+Comunicarea Preşedintelui: consultaţi procesul-verbal
+1.
+Componenţa comisiilor (vot)
+2.
+Modificarea Acordului de parteneriat ACP-CE ("Acordul de la Cotonou”) (vot)
+4.
+Certificarea mecanicilor de locomotivă care conduc locomotive şi trenuri în sistemul feroviar comunitar (vot)
+6.
+Legea aplicabilă obligaţiilor necontractuale ("Roma II”) (vot)
+8.
+Al şaptelea şi al optulea raport anual privind exportul de armament (vot)
+Corectările voturilor şi intenţiile de vot: a se vedea procesul-verbal
+Componenţa comisiilor şi a delegaţiilor: a se vedea procesul-verbal
+Cerere de ridicare a imunităţii parlamentare: consultaţi procesul-verbal
+Decizii privind anumite documente: a se vedea procesul-verbal
+Declaraţii scrise înscrise în registru (articolul 116 din Regulamentul de procedură): a se vedea procesul-verbal
+Transmiterea textelor adoptate în cursul prezentei şedinţe: a se vedea procesul-verbal
+Calendarul următoarelor şedinţe: a se vedea procesul-verbal
+Întreruperea sesiunii
+Dichiaro interrotta la sessione del Parlamento europeo.
+(La seduta è tolta alle 13.00)
+Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal
+Componenţa Parlamentului: a se vedea procesul-verbal
+Cerere de apărare a imunităţii parlamentare: consultaţi procesul-verbal
+Numiri în comisii (propunerea Conferinţei preşedinţilor): consultaţi procesul-verbal
+Depunerea documentelor: a se vedea procesul-verbal
+Transmiterea de către Consiliu a textelor acordurilor: a se vedea procesul-verbal
+Continuări ale rezoluţiilor Parlamentului: consultaţi procesul-verbal
+Declaraţii scrise şi întrebări orale (depunere): consultaţi procesul-verbal
+Declaraţii scrise (articolul 116 din Regulamentul de procedură)
+Ordinea de zi: a se vedea procesul-verbal
+1.
+Numiri în comisiile parlamentare (vot): consultaţi procesul-verbal
+Timpul afectat votului
+Ordinea de zi a următoarei şedinţe: a se vedea procesul-verbal
+Ridicarea şedinţei
+(La seduta è tolta alle 24.00)
+Deschiderea şedinţei
+(The sitting was opened at 09.05)
+Depunerea documentelor: a se vedea procesul-verbal
+Aprobarea procesului-verbal al şedinţei precedente: a se vedea procesul-verbal
+1.
+Protecţia pasagerilor împotriva deplasării bagajelor (vot)
+2.
+Omologarea vehiculelor cu motor cu privire la câmpul de vizibilitate înainte al conducătorului auto (vot)
+3.
+Acordul CE-Coreea de cooperare ştiinţifică şi tehnologică (vot)
+4.
+Integrarea durabilităţii în politicile de cooperare pentru dezvoltare (vot)
+5.
+Proiect de buget rectificativ nr.1/2007 (vot)
+7.
+Acordul de parteneriat în domeniul pescuitului între Comunitatea Europeană şi Republica Gaboneză (vot)
+10.
+Termenele de prescripţie aplicabile în cadrul litigiilor transfrontaliere cu privire la vătămările corporale şi accidentele mortale (vot)
+12.
+Relaţiile UE cu insulele din Pacific: Strategie pentru un parteneriat consolidat (vot)
+13.
+Statutul societăţii private europene (vot)
+Damit ist die Abstimmungsstunde beendet.
+Corectările voturilor şi intenţiile de vot: a se vedea procesul-verbal
+Misiune încredinţată unui deputat: consultaţi procesul-verbal
+Componenţa comisiilor şi a delegaţiilor: a se vedea procesul-verbal
+Decizii privind anumite documente: a se vedea procesul-verbal
+Transmiterea textelor adoptate în cursul prezentei şedinţe: a se vedea procesul-verbal
+Calendarul următoarelor şedinţe: a se vedea procesul-verbal
@@ -1,16 +1,100 @@
-Brazil's Former Presidential Chief-of-Staff to Stand Trial A federal judge on Tuesday accepted the charges filed against Brazil's former presidential chief of staff for his alleged involvement in a massive corruption scheme at state-owned oil company Petrobras. The federal prosecutor's office said Jose Dirceu will face trial on the corruption, racketeering and money laundering charges filed earlier this month. Fourteen other people will also be tried, including Joao Vaccari Neto, the former treasurer of Brazil's governing Workers' Party and Renato de Souza Duque, Petrobras' former head of corporate services.
-Dirceu is the most senior member of the ruling Workers' Party to be taken into custody in connection with the scheme. Dirceu served as former President Luiz Inacio Lula da Silva's chief of staff between 2003 and 2005. He was arrested early August in his home, where he already was under house arrest serving an 11-year sentence for his involvement in a cash-for-votes scheme in Congress more than 10 years ago. Prosecutors have said that Dirceu masterminded the kickback scheme at Petrobras, accepted bribes while in office and continued to receive payments from contractors after he was jailed in late 2013 for the vote-buying scandal.
-According to prosecutors, the scheme at Petrobras involved roughly $2 billion in bribes and other illegal funds. Some of that money was allegedly funneled back to campaign coffers of the ruling party and its allies. It also allegedly included the payment of bribes to Petrobras executives in return for inflated contracts. 'Miraculous' recovery for Peshawar massacre schoolboy A teenager paralysed after being shot four times in Pakistan's deadliest terror attack has made a "miraculous" recovery following treatment in the UK. Muhammad Ibrahim Khan, 13, had been told by doctors in Pakistan that he would never walk again.
-At least 140 people, mostly children, were killed when gunmen stormed Peshawar's Army Public School last December. Muhammad, who arrived in London last month for surgery, is being discharged from hospital later. Exactly nine months ago, on an ordinary Tuesday morning, Muhammad sat in his first aid class listening to his teachers intently. At the same time seven gunmen disguised in security uniforms were entering the Army Public School. They were strapped with explosives and had one simple mission in mind: Kill every man, woman and child they came across. "I can't forget what happened that day," Muhammad says with a severe stare.
-We were sitting in the auditorium, we were asking questions... and then we heard heavy gunfire outside. The terrorists moved inside and they started killing - our teacher was burned alive. Muhammad described pulling four other pupils out of the auditorium as the carnage unfolded. He said he then heard his friend, Hamza calling to him. He said, 'oh brother save me'. I held his hand. That's when I was shot in the back, and he was shot in the head. Most of the people killed in the attack were pupils Hamza died in Muhammad's arms. Muhammad recalled blacking out after that, and the next thing he knew he was in a hospital bed, paralysed from the waist down.
-Doctors in Peshawar in northern Pakistan, and then Rawalpindi, close to the capital, told his family there was no treatment, and he would never walk again. "Seeing him I felt like my soul had left my body," says Muhammad's father, Sher Khan Those nine months were the hardest in my life. But Mr Khan and his wife, Sherbano, refused to believe that their cricket-mad son would never be able to use his legs again. They campaigned, and appealed for help on Pakistani TV, gaining the support of high profile people such as cricketer turned politician Imran Khan.
-Finally, they were able to raise the funds to bring Muhammad to the UK and provide him with treatment at London's private Harley Street Clinic. Consultant neurosurgeon Irfan Malik described Muhammad as "terrified" when he first arrived at the hospital. "He'd spent the last [few] months lying on a bed, unable to move side to side," says Mr Malik. He was weak, he had a pressure sore on his back. He wasn't in great shape. A vertebra at the base of Muhammad's spine was destroyed Muhammad was shot in his shoulder, his hip, and his back during the attack, damaging his lower spine - leading to paralysis.
-But during six hours of surgery, Mr Malik and his team were able to reattach nerve endings and reconstruct the damaged part of the spine. Even Mr Malik was surprised at what happened next. Exactly one week after the surgery Muhammad stood up and started taking steps and walking. We were not expecting to get that sort of excellent result. That was miraculous," he says. Less than two weeks after his operation, Muhammad is ready to leave hospital and start the long road to recovery. Muhammad has defied the odds and started to walk again He says he wants to build his strength and continue his education in the UK. But he says he is determined to return to Pakistan, join the army and help fight terrorism.
-"I feel like I have a second chance at life," he says as he shows off pictures he's drawn of guns scribbled out next to school books and pens Muhammad grows physically stronger every day but the psychological trauma he continues to endure is unimaginable. "My anger is not diminishing" he says. In my school little kids were killed. What was their crime? His mother, wiping a tear from her eye, caressed his head and said: "I can see my son walking again." He'll be able to get on with his normal life. 'Super Voice' 4G service from Three offers better signal Three is making use of a lower frequency 4G spectrum that can travel more widely
-Mobile phone provider Three has launched a UK service it says will improve reception inside buildings and in rural black spots. Its 4G Super Voice enables customers to make calls and send texts using a lower frequency spectrum. Other networks are looking into introducing the technology, known as Voice Over Long-Term Evolution (VoLTE). It currently works on only the Samsung Galaxy S5, but recent iPhone handsets will be added in the coming months. Three said up to 5.5 million customers would have access to the service by 2017.
-Chief technology officer Bryn Jones said: "By the end of the year, one million of our customers will have access to better indoor coverage and be able to use their phones in more places than ever before." Stars prepare for panto season Pantomime season is big business for theatres up and down the UK, with many getting ready for this year's season now. Some of the biggest names in showbusiness now take part in the yuletide theatre. Matthew Kelly and Hayley Mills will be appearing in Cinderella - one as an ugly sister, the other as fairy godmother. They reveal their panto secrets to BBC Breakfast. Steven Wilson: 'If I don't do anything, I feel this creeping guilt'
-Steven Wilson was recently the big winner at the Progressive Music Awards Steven Wilson is often dubbed the hardest working musician in the world of progressive rock. The multi-talented musician won three prizes at this month's Progressive Music Awards in London, including album of the year for Hand. The Guardian's five-star review called it "a smart, soulful and immersive work of art." Since the 1980s, Wilson has been the driving force in a number of musical projects, the best known of which is the rock band Porcupine Tree. Now, ahead of two sell-out shows at the Royal Albert Hall, Wilson is releasing a vinyl-only double LP, Transience, to showcase the "more accessible" side of his solo output.
-He tells the BBC about his love of vinyl, his busy schedule and explains how comic actor Matt Berry came to be his support act. What does vinyl mean to you? I grew up at the very tail end of the vinyl era, and at the time, I remember, we couldn't wait for CD to come along because vinyl was so frustrating. You would buy the record, take it home, and it would have a scratch, and you would have to take it back again. I love CDs, and for some kinds of music - classical for example - it is better than vinyl. But the problem with the CD and digital downloads is that there's nothing you can really cherish or treasure. Owning vinyl is like having a beautiful painting hanging in your living room.
-It's something you can hold, pore over the lyrics and immerse yourself in the art work. I thought it was just a nostalgic thing, but it can't be if kids too young to remember vinyl are enjoying that kind of experience. Do you have a piece of vinyl that you treasure? The truth is I got rid of 100% of my vinyl in the 90s. All the vinyl I have is re-bought. I started off from the perspective that I wanted to recreate the collection I had when I was 15, but it's gone beyond that. The first record which I persuaded my parents to buy for me was Electric Light Orchestra's Out of the Blue.
-If I still had my original copy, it would have sentimental value, but, alas, it's in a charity shop somewhere. Steven Wilson hopes the album will be a doorway for potential new fans Why release your new compilation Transience on vinyl? It was originally conceived as an idea for Record Store Day, but we missed the boat on that. My record company had suggested I put together some of my shorter, more accessible songs. I got a bit obsessed by the idea to make something like "an introduction to Steven Wilson," and I was committed to it being a vinyl-only release. Anyone who buys the vinyl does also get a high-resolution download.
-Do you have a concern that the album won't show your work in a true light?
+Brazil's Former Presidential Chief-of-Staff to Stand Trial
+A federal judge on Tuesday accepted the charges filed against Brazil's former presidential chief of staff for his alleged involvement in a massive corruption scheme at state-owned oil company Petrobras.
+The federal prosecutor's office said Jose Dirceu will face trial on the corruption, racketeering and money laundering charges filed earlier this month.
+Fourteen other people will also be tried, including Joao Vaccari Neto, the former treasurer of Brazil's governing Workers' Party and Renato de Souza Duque, Petrobras' former head of corporate services.
+Dirceu is the most senior member of the ruling Workers' Party to be taken into custody in connection with the scheme.
+Dirceu served as former President Luiz Inacio Lula da Silva's chief of staff between 2003 and 2005.
+He was arrested early August in his home, where he already was under house arrest serving an 11-year sentence for his involvement in a cash-for-votes scheme in Congress more than 10 years ago.
+Prosecutors have said that Dirceu masterminded the kickback scheme at Petrobras, accepted bribes while in office and continued to receive payments from contractors after he was jailed in late 2013 for the vote-buying scandal.
+According to prosecutors, the scheme at Petrobras involved roughly $2 billion in bribes and other illegal funds.
+Some of that money was allegedly funneled back to campaign coffers of the ruling party and its allies.
+It also allegedly included the payment of bribes to Petrobras executives in return for inflated contracts.
+'Miraculous' recovery for Peshawar massacre schoolboy
+A teenager paralysed after being shot four times in Pakistan's deadliest terror attack has made a "miraculous" recovery following treatment in the UK.
+Muhammad Ibrahim Khan, 13, had been told by doctors in Pakistan that he would never walk again.
+At least 140 people, mostly children, were killed when gunmen stormed Peshawar's Army Public School last December.
+Muhammad, who arrived in London last month for surgery, is being discharged from hospital later.
+Exactly nine months ago, on an ordinary Tuesday morning, Muhammad sat in his first aid class listening to his teachers intently.
+At the same time seven gunmen disguised in security uniforms were entering the Army Public School.
+They were strapped with explosives and had one simple mission in mind: Kill every man, woman and child they came across.
+"I can't forget what happened that day," Muhammad says with a severe stare.
+We were sitting in the auditorium, we were asking questions... and then we heard heavy gunfire outside.
+The terrorists moved inside and they started killing - our teacher was burned alive.
+Muhammad described pulling four other pupils out of the auditorium as the carnage unfolded.
+He said he then heard his friend, Hamza calling to him.
+He said, 'oh brother save me'.
+I held his hand.
+That's when I was shot in the back, and he was shot in the head.
+Most of the people killed in the attack were pupils
+Hamza died in Muhammad's arms.
+Muhammad recalled blacking out after that, and the next thing he knew he was in a hospital bed, paralysed from the waist down.
+Doctors in Peshawar in northern Pakistan, and then Rawalpindi, close to the capital, told his family there was no treatment, and he would never walk again.
+"Seeing him I felt like my soul had left my body," says Muhammad's father, Sher Khan
+Those nine months were the hardest in my life.
+But Mr Khan and his wife, Sherbano, refused to believe that their cricket-mad son would never be able to use his legs again.
+They campaigned, and appealed for help on Pakistani TV, gaining the support of high profile people such as cricketer turned politician Imran Khan.
+Finally, they were able to raise the funds to bring Muhammad to the UK and provide him with treatment at London's private Harley Street Clinic.
+Consultant neurosurgeon Irfan Malik described Muhammad as "terrified" when he first arrived at the hospital.
+"He'd spent the last [few] months lying on a bed, unable to move side to side," says Mr Malik.
+He was weak, he had a pressure sore on his back.
+He wasn't in great shape.
+A vertebra at the base of Muhammad's spine was destroyed
+Muhammad was shot in his shoulder, his hip, and his back during the attack, damaging his lower spine - leading to paralysis.
+But during six hours of surgery, Mr Malik and his team were able to reattach nerve endings and reconstruct the damaged part of the spine.
+Even Mr Malik was surprised at what happened next.
+Exactly one week after the surgery Muhammad stood up and started taking steps and walking.
+We were not expecting to get that sort of excellent result.
+That was miraculous," he says.
+Less than two weeks after his operation, Muhammad is ready to leave hospital and start the long road to recovery.
+Muhammad has defied the odds and started to walk again
+He says he wants to build his strength and continue his education in the UK.
+But he says he is determined to return to Pakistan, join the army and help fight terrorism.
+"I feel like I have a second chance at life," he says as he shows off pictures he's drawn of guns scribbled out next to school books and pens
+Muhammad grows physically stronger every day but the psychological trauma he continues to endure is unimaginable.
+"My anger is not diminishing" he says.
+In my school little kids were killed.
+What was their crime?
+His mother, wiping a tear from her eye, caressed his head and said: "I can see my son walking again."
+He'll be able to get on with his normal life.
+'Super Voice' 4G service from Three offers better signal
+Three is making use of a lower frequency 4G spectrum that can travel more widely
+Mobile phone provider Three has launched a UK service it says will improve reception inside buildings and in rural black spots.
+Its 4G Super Voice enables customers to make calls and send texts using a lower frequency spectrum.
+Other networks are looking into introducing the technology, known as Voice Over Long-Term Evolution (VoLTE).
+It currently works on only the Samsung Galaxy S5, but recent iPhone handsets will be added in the coming months.
+Three said up to 5.5 million customers would have access to the service by 2017.
+Chief technology officer Bryn Jones said: "By the end of the year, one million of our customers will have access to better indoor coverage and be able to use their phones in more places than ever before."
+Stars prepare for panto season
+Pantomime season is big business for theatres up and down the UK, with many getting ready for this year's season now.
+Some of the biggest names in showbusiness now take part in the yuletide theatre.
+Matthew Kelly and Hayley Mills will be appearing in Cinderella - one as an ugly sister, the other as fairy godmother.
+They reveal their panto secrets to BBC Breakfast.
+Steven Wilson: 'If I don't do anything, I feel this creeping guilt'
+Steven Wilson was recently the big winner at the Progressive Music Awards
+Steven Wilson is often dubbed the hardest working musician in the world of progressive rock.
+The multi-talented musician won three prizes at this month's Progressive Music Awards in London, including album of the year for Hand.
+The Guardian's five-star review called it "a smart, soulful and immersive work of art."
+Since the 1980s, Wilson has been the driving force in a number of musical projects, the best known of which is the rock band Porcupine Tree.
+Now, ahead of two sell-out shows at the Royal Albert Hall, Wilson is releasing a vinyl-only double LP, Transience, to showcase the "more accessible" side of his solo output.
+He tells the BBC about his love of vinyl, his busy schedule and explains how comic actor Matt Berry came to be his support act.
+What does vinyl mean to you?
+I grew up at the very tail end of the vinyl era, and at the time, I remember, we couldn't wait for CD to come along because vinyl was so frustrating.
+You would buy the record, take it home, and it would have a scratch, and you would have to take it back again.
+I love CDs, and for some kinds of music - classical for example - it is better than vinyl.
+But the problem with the CD and digital downloads is that there's nothing you can really cherish or treasure.
+Owning vinyl is like having a beautiful painting hanging in your living room.
+It's something you can hold, pore over the lyrics and immerse yourself in the art work.
+I thought it was just a nostalgic thing, but it can't be if kids too young to remember vinyl are enjoying that kind of experience.
+Do you have a piece of vinyl that you treasure?
+The truth is I got rid of 100% of my vinyl in the 90s.
+All the vinyl I have is re-bought.
+I started off from the perspective that I wanted to recreate the collection I had when I was 15, but it's gone beyond that.
+The first record which I persuaded my parents to buy for me was Electric Light Orchestra's Out of the Blue.
+If I still had my original copy, it would have sentimental value, but, alas, it's in a charity shop somewhere.
+Steven Wilson hopes the album will be a doorway for potential new fans
+Why release your new compilation Transience on vinyl?
+It was originally conceived as an idea for Record Store Day, but we missed the boat on that.
+My record company had suggested I put together some of my shorter, more accessible songs.
+I got a bit obsessed by the idea to make something like "an introduction to Steven Wilson," and I was committed to it being a vinyl-only release.
+Anyone who buys the vinyl does also get a high-resolution download.
+Do you have a concern that the album won't show your work in a true light?
@@ -1,16 +1,100 @@
-Fostul șef al cabinetului prezidențial brazilian este adus în fața instanței Marți, un judecător federal a acceptat acuzațiile aduse împotriva fostului șef al cabinetului prezidențial brazilian pentru presupusa implicare a acestuia într-o schemă masivă de corupție privind compania petrolieră de stat Petrobras. Biroul procurorului federal a declarat că Jose Dirceu va fi trimis în judecată pentru acuzațiile de corupție, înșelătorie și spălare de bani aduse în această lună. Alte paisprezece persoane vor fi judecate, printre acestea numărându-se Joao Vaccari Neto, fostul trezorier al Partidului Muncitorilor, aflat la putere în Brazilia, și Renato de Souza Duque, fostul președinte al serviciilor pentru întreprinderi ale Petrobras.
-Dirceu este cel mai vechi membru al Partidului Muncitorilor aflat la guvernare luat în custodie pentru legăturile cu această schemă. Dirceu a servit ca șef de cabinet al fostului președinte Luiz Inacio Lula da Silva între 2003 și 2005. A fost arestat la începutul lui august de acasă, unde deja se afla sub arest la domiciliu, cu o pedeapsă de 11 ani pentru implicarea într-o schemă de cumpărare a voturilor în Congres cu peste 10 ani în urmă. Procurorii au declarat că Dirceu a dezvoltat schema de luare de mită de la Petrobras, a acceptat mită în timp ce se afla în funcție și a continuat să primească plăți de la antreprenori după ce a fost închis la sfârșitul lui 2013 pentru scandalul voturilor cumpărate.
-Conform procurorilor, schema de la Petrobras a implicat aproximativ 2 miliarde de dolari sub formă de mită și alte fonduri ilegale. O parte din acei bani s-ar fi întors în fondul de campanie al partidului aflat la guvernare și al aliaților acestora. De asemenea, ar fi inclus mită către directorii Petrobras în schimbul unor contracte umflate. Recuperarea „miraculoasă” a unui elev supraviețuitor al masacrului de la Peshawar Un adolescent paralizat după ce fusese împușcat de patru ori în cel mai cumplit atac terorist din Pakistan a reușit o recuperare „miraculoasă” după ce a urmat un tratament în Regatul Unit. Lui Mohamed Ibrahim Khan, în vârstă de 13 ani, doctorii din Pakistan îi spuseseră că nu va mai putea să meargă niciodată.
-Cel puțin 140 de persoane, majoritatea copii, au fost ucise când bărbați înarmați au atacat școala publică a armatei din Peshawar în luna decembrie a anului trecut. Mohamed, care a sosit la Londra luna trecută pentru operație, va fi externat mai târziu din spital. Exact cu nouă luni în urmă, într-o dimineață obișnuită de marți, Mohamed stătea la ora de primul ajutor și își asculta atent profesorii. Chiar atunci, șapte bărbați înarmați deghizați în uniformele agenților de pază intrau în școala publică a armatei. Purtau centuri cu explozivi și aveau de îndeplinit o misiune simplă: să îi ucidă pe toți bărbații, femeile și copiii care le ieșeau în cale. „Nu pot uita ce s-a întâmplat în acea zi”, spune Mohamed cu o privire aspră.
-Stăteam în amfiteatru, puneam întrebări... apoi am auzit focuri de armă afară. Teroriștii au intrat înăuntru și au început să ucidă. Profesorul nostru a fost ars de viu. Mohamed descrie cum a scos patru elevi din amfiteatru în timp ce se desfășura carnagiul. Apoi spune că și-a auzit prietenul, pe Hamza, strigându-l. Spunea „oh, frate, salvează-mă”. L-am ținut de mână. Atunci eu am fost împușcat în spate, iar el în cap. Cei mai mulți dintre cei uciși în atac erau elevi Hamza a murit în brațele lui Mohamed. Mohamed își amintește că imediat după asta a leșinat și că următorul lucru pe care l-a știut a fost că se afla pe un pat de spital, paralizat de la brâu în jos.
-Doctorii din Peshawar din nordul Pakistanului, apoi cei din Rawalpindi, aproape de capitală, i-au spus familiei sale că nu exista tratament și că nu va mai putea merge niciodată. „Când l-am văzut, am simțit cum îmi iese sufletul”, spune Sher Khan, tatăl lui Mohamed. Acele nouă luni au fost cele mai grele din viața mea. Însă Khan și soția lui, Sherbano, au refuzat să creadă că fiul lor atât de pasionat de crichet nu-și va mai putea folosi vreodată picioarele. Au făcut o campanie și au cerut ajutor de la televiziunea pakistaneză, atrăgând sprijinul unor oameni faimoși precum Imran Khan, jucător de crichet devenit politician.
-Într-un final, au reușit să strângă fonduri pentru a-l duce pe Mohamed în Regatul Unit și a-i oferi tratament la clinica privată Harley Street din Londra. Neurochirurgul consultant Irfan Malik l-a descris pe Mohamed drept „înspăimântat” când acesta a ajuns la spital. „Își petrecuse ultimele [câteva] luni zăcând în pat, fără să se poată mișca de pe o parte pe alta, spune Malik. Era slăbit, se pusese multă presiune pe spatele lui. Nu era într-o formă prea bună. O vertebră de la baza coloanei vertebrale a lui Mohamed fusese distrusă Mohamed fusese împușcat în umăr, în șold și în spate în timpul atacului, iar coloana vertebrală inferioară îi fusese distrusă, ducând la paralizie.
-Însă, în timpul unei operații care a durat șase ore, Malik și echipa lui au reușit să lege din nou terminațiile nervoase și să reconstruiască partea distrusă a coloanei. Chiar și Malik a fost surprins de ceea ce s-a întâmplat în continuare. Exact la o săptămână după operație, Mohamed s-a ridicat și a început să facă pași și să meargă. Nu ne așteptam la un rezultat atât de bun. A fost un miracol”, spune acesta. În mai puțin de două săptămâni de la operație, Mohamed este gata să părăsească spitalul și să înceapă procesul lung de recuperare. Mohamed a sfidat soarta și a început să meargă din nou Vrea să devină puternic și să își continue studiile în Regatul Unit. Însă este hotărât să revină în Pakistan, să se înroleze în armată și să lupte împotriva terorismului.
-„Simt că am încă o șansă la viață” spune el, arătând imaginile cu arme desenate de el lângă manuale școlare și stilouri Fizic, Mohamed devine tot mai puternic în fiecare zi, însă trauma psihologică prin care trece și acum este de neimaginat. „Furia mea nu a scăzut”, mărturisește el. În școala mea au fost uciși copii mici. Ce crimă au comis ei? Mama lui își șterge o lacrimă, îl mângâie pe creștet și spune: „Îmi văd fiul mergând din nou”. Va putea să-și continue firesc viața. Serviciul 4G „Super Voice” de la Three oferă semnal mai bun Three folosește un spectru 4G cu o frecvență mai joasă, care poate acoperi o zonă mai extinsă
-Furnizorul de telefonie mobilă Three a lansat în Regatul Unit un serviciu despre care spune că va îmbunătăți recepția în interiorul clădirilor și în zonele rurale fără semnal. Serviciul 4G Super Voice le permite clienților să efectueze apeluri și să trimită mesaje text folosind un spectru cu o frecvență mai joasă. Și alte rețele intenționează să introducă aceeași tehnologie, cunoscută ca „Voice Over Long-Term Evolution (VoLTE)”. Aceasta funcționează momentan doar cu Samsung Galaxy S5, însă telefoanele iPhone recente vor beneficia de ea în lunile următoare. Three menționează că până la 5,5 milioane de clienți vor avea acces la serviciu până în 2017.
-Responsabilul șef pentru tehnologie, Bryn Jones a declarat: „Până la sfârșitul anului, un milion dintre clienții noștri vor avea acces la o acoperire mai bună în interior și își vor putea folosi telefoanele în mai multe locuri ca până acum”. Vedetele se pregătesc pentru stagiunea de pantomimă Stagiunea de pantomimă este foarte importantă pentru teatrele din tot Regatul Unit, multe dintre ele pregătindu-se acum pentru stagiunea din acest an. Acum, la teatrul de Crăciun participă unele dintre numele cele mai mari din showbusiness. Matthew Kelly și Hayley Mills vor apărea în Cenușăreasa - primul în rolul uneia dintre surorile rele, iar a doua în rolul zânei. Aceștia dezvăluie secretele pantomimei lor la BBC Breakfast. Steven Wilson: „Dacă nu fac nimic, mă simt vinovat”
-Steven Wilson a fost desemnat recent drept marele câștigător al Progressive Music Awards Steven Wilson a fost numit de multe ori drept cel mai muncitor muzician din lumea rockului progresiv. Talentatul muzician a câștigat trei premii la Progressive Music Awards, care a avut loc luna aceasta la Londra, printre care și premiul pentru cel mai bun album al anului pentru Hand. În recenzia sa de cinci stele, The Guardian a numit albumul „o operă de artă inteligentă, expresivă și captivantă”. Încă din anii 1980, Wilson este motorul mai multor proiecte muzicale, cel mai cunoscut dintre acestea fiind trupa de rock Porcupine Tree. Acum, înainte de două spectacole cu casa închisă la Royal Albert Hall, Wilson lansează un dublu LP doar în format vinil, Transience, pentru a arăta latura „mai accesibilă” a activității sale solo.
-A povestit pentru BBC despre dragostea lui pentru viniluri și despre programul său încărcat și a explicat cum a ajuns actorul de comedie Matt Berry să îi deschidă spectacolele. Ce înseamnă vinil pentru tine? Am crescut chiar în perioada de sfârșit a erei vinilurilor și îmi amintesc că atunci abia așteptam apariția CD-ului, căci vinilul era atât de enervant. Cumpărai un disc, mergeai cu el acasă, avea o zgârietură și trebuia să îl aduci înapoi. Iubesc CD-urile, iar pentru anumite tipuri de muzică, de exemplu cea clasică, sunt mai bune decât vinilurile. Însă problema cu CD-urile și cu descărcările digitale este aceea că nu mai există nimic pe care să îl prețuiești cu adevărat. Să ai un vinil e ca și cum ai avea un tablou frumos agățat în sufragerie.
-E ceva ce poți ține în mână, în timp ce te lași absorbit de versuri și copleșit de actul artistic. Am crezut că e doar o chestie nostalgică, însă nu are cum să fie așa dacă unor puști prea tineri să-și amintească de viniluri le place acest gen de experiență. Ai vreun vinil la care ții în mod special? Recunosc că am scăpat de toate vinilurile în anii '90. Toate vinilurile pe care le am sunt cumpărate din nou. Am pornit de la ideea de a reface colecția pe care o aveam la 15 ani, însă am trecut de limita aceea. Primul disc pe care mi-am convins părinții să mi-l cumpere a fost Out of the Blue de la Electric Light Orchestra.
-Dacă aș mai fi avut încă exemplarul inițial, acesta ar fi avut valoare sentimentală, însă, din păcate, se află pe undeva printr-un magazin de caritate. Steven Wilson speră că albumul va fi o poartă către posibili fani noi De ce ți-ai lansat noua compilație Transience pe vinil? Aceasta a fost concepută inițial ca idee pentru Ziua magazinelor de discuri, însă am ratat ocazia. Casa mea de discuri sugerase să adun câteva dintre melodiile mele mai scurte și mai accesibile. Am ajuns să fiu ușor obsedat de ideea de a face ceva gen „introducere în muzica lui Steven Wilson” și am ținut neapărat ca proiectul să fie lansat doar pe vinil. Cine cumpără vinilul primește, de asemenea, și o variantă descărcată la rezoluție înaltă.
-Ești îngrijorat că albumul nu va arăta muzica ta în adevărata ei lumină?
+Fostul șef al cabinetului prezidențial brazilian este adus în fața instanței
+Marți, un judecător federal a acceptat acuzațiile aduse împotriva fostului șef al cabinetului prezidențial brazilian pentru presupusa implicare a acestuia într-o schemă masivă de corupție privind compania petrolieră de stat Petrobras.
+Biroul procurorului federal a declarat că Jose Dirceu va fi trimis în judecată pentru acuzațiile de corupție, înșelătorie și spălare de bani aduse în această lună.
+Alte paisprezece persoane vor fi judecate, printre acestea numărându-se Joao Vaccari Neto, fostul trezorier al Partidului Muncitorilor, aflat la putere în Brazilia, și Renato de Souza Duque, fostul președinte al serviciilor pentru întreprinderi ale Petrobras.
+Dirceu este cel mai vechi membru al Partidului Muncitorilor aflat la guvernare luat în custodie pentru legăturile cu această schemă.
+Dirceu a servit ca șef de cabinet al fostului președinte Luiz Inacio Lula da Silva între 2003 și 2005.
+A fost arestat la începutul lui august de acasă, unde deja se afla sub arest la domiciliu, cu o pedeapsă de 11 ani pentru implicarea într-o schemă de cumpărare a voturilor în Congres cu peste 10 ani în urmă.
+Procurorii au declarat că Dirceu a dezvoltat schema de luare de mită de la Petrobras, a acceptat mită în timp ce se afla în funcție și a continuat să primească plăți de la antreprenori după ce a fost închis la sfârșitul lui 2013 pentru scandalul voturilor cumpărate.
+Conform procurorilor, schema de la Petrobras a implicat aproximativ 2 miliarde de dolari sub formă de mită și alte fonduri ilegale.
+O parte din acei bani s-ar fi întors în fondul de campanie al partidului aflat la guvernare și al aliaților acestora.
+De asemenea, ar fi inclus mită către directorii Petrobras în schimbul unor contracte umflate.
+Recuperarea „miraculoasă” a unui elev supraviețuitor al masacrului de la Peshawar
+Un adolescent paralizat după ce fusese împușcat de patru ori în cel mai cumplit atac terorist din Pakistan a reușit o recuperare „miraculoasă” după ce a urmat un tratament în Regatul Unit.
+Lui Mohamed Ibrahim Khan, în vârstă de 13 ani, doctorii din Pakistan îi spuseseră că nu va mai putea să meargă niciodată.
+Cel puțin 140 de persoane, majoritatea copii, au fost ucise când bărbați înarmați au atacat școala publică a armatei din Peshawar în luna decembrie a anului trecut.
+Mohamed, care a sosit la Londra luna trecută pentru operație, va fi externat mai târziu din spital.
+Exact cu nouă luni în urmă, într-o dimineață obișnuită de marți, Mohamed stătea la ora de primul ajutor și își asculta atent profesorii.
+Chiar atunci, șapte bărbați înarmați deghizați în uniformele agenților de pază intrau în școala publică a armatei.
+Purtau centuri cu explozivi și aveau de îndeplinit o misiune simplă: să îi ucidă pe toți bărbații, femeile și copiii care le ieșeau în cale.
+„Nu pot uita ce s-a întâmplat în acea zi”, spune Mohamed cu o privire aspră.
+Stăteam în amfiteatru, puneam întrebări... apoi am auzit focuri de armă afară.
+Teroriștii au intrat înăuntru și au început să ucidă. Profesorul nostru a fost ars de viu.
+Mohamed descrie cum a scos patru elevi din amfiteatru în timp ce se desfășura carnagiul.
+Apoi spune că și-a auzit prietenul, pe Hamza, strigându-l.
+Spunea „oh, frate, salvează-mă”.
+L-am ținut de mână.
+Atunci eu am fost împușcat în spate, iar el în cap.
+Cei mai mulți dintre cei uciși în atac erau elevi
+Hamza a murit în brațele lui Mohamed.
+Mohamed își amintește că imediat după asta a leșinat și că următorul lucru pe care l-a știut a fost că se afla pe un pat de spital, paralizat de la brâu în jos.
+Doctorii din Peshawar din nordul Pakistanului, apoi cei din Rawalpindi, aproape de capitală, i-au spus familiei sale că nu exista tratament și că nu va mai putea merge niciodată.
+„Când l-am văzut, am simțit cum îmi iese sufletul”, spune Sher Khan, tatăl lui Mohamed.
+Acele nouă luni au fost cele mai grele din viața mea.
+Însă Khan și soția lui, Sherbano, au refuzat să creadă că fiul lor atât de pasionat de crichet nu-și va mai putea folosi vreodată picioarele.
+Au făcut o campanie și au cerut ajutor de la televiziunea pakistaneză, atrăgând sprijinul unor oameni faimoși precum Imran Khan, jucător de crichet devenit politician.
+Într-un final, au reușit să strângă fonduri pentru a-l duce pe Mohamed în Regatul Unit și a-i oferi tratament la clinica privată Harley Street din Londra.
+Neurochirurgul consultant Irfan Malik l-a descris pe Mohamed drept „înspăimântat” când acesta a ajuns la spital.
+„Își petrecuse ultimele [câteva] luni zăcând în pat, fără să se poată mișca de pe o parte pe alta, spune Malik.
+Era slăbit, se pusese multă presiune pe spatele lui.
+Nu era într-o formă prea bună.
+O vertebră de la baza coloanei vertebrale a lui Mohamed fusese distrusă
+Mohamed fusese împușcat în umăr, în șold și în spate în timpul atacului, iar coloana vertebrală inferioară îi fusese distrusă, ducând la paralizie.
+Însă, în timpul unei operații care a durat șase ore, Malik și echipa lui au reușit să lege din nou terminațiile nervoase și să reconstruiască partea distrusă a coloanei.
+Chiar și Malik a fost surprins de ceea ce s-a întâmplat în continuare.
+Exact la o săptămână după operație, Mohamed s-a ridicat și a început să facă pași și să meargă.
+Nu ne așteptam la un rezultat atât de bun.
+A fost un miracol”, spune acesta.
+În mai puțin de două săptămâni de la operație, Mohamed este gata să părăsească spitalul și să înceapă procesul lung de recuperare.
+Mohamed a sfidat soarta și a început să meargă din nou
+Vrea să devină puternic și să își continue studiile în Regatul Unit.
+Însă este hotărât să revină în Pakistan, să se înroleze în armată și să lupte împotriva terorismului.
+„Simt că am încă o șansă la viață” spune el, arătând imaginile cu arme desenate de el lângă manuale școlare și stilouri
+Fizic, Mohamed devine tot mai puternic în fiecare zi, însă trauma psihologică prin care trece și acum este de neimaginat.
+„Furia mea nu a scăzut”, mărturisește el.
+În școala mea au fost uciși copii mici.
+Ce crimă au comis ei?
+Mama lui își șterge o lacrimă, îl mângâie pe creștet și spune: „Îmi văd fiul mergând din nou”.
+Va putea să-și continue firesc viața.
+Serviciul 4G „Super Voice” de la Three oferă semnal mai bun
+Three folosește un spectru 4G cu o frecvență mai joasă, care poate acoperi o zonă mai extinsă
+Furnizorul de telefonie mobilă Three a lansat în Regatul Unit un serviciu despre care spune că va îmbunătăți recepția în interiorul clădirilor și în zonele rurale fără semnal.
+Serviciul 4G Super Voice le permite clienților să efectueze apeluri și să trimită mesaje text folosind un spectru cu o frecvență mai joasă.
+Și alte rețele intenționează să introducă aceeași tehnologie, cunoscută ca „Voice Over Long-Term Evolution (VoLTE)”.
+Aceasta funcționează momentan doar cu Samsung Galaxy S5, însă telefoanele iPhone recente vor beneficia de ea în lunile următoare.
+Three menționează că până la 5,5 milioane de clienți vor avea acces la serviciu până în 2017.
+Responsabilul șef pentru tehnologie, Bryn Jones a declarat: „Până la sfârșitul anului, un milion dintre clienții noștri vor avea acces la o acoperire mai bună în interior și își vor putea folosi telefoanele în mai multe locuri ca până acum”.
+Vedetele se pregătesc pentru stagiunea de pantomimă
+Stagiunea de pantomimă este foarte importantă pentru teatrele din tot Regatul Unit, multe dintre ele pregătindu-se acum pentru stagiunea din acest an.
+Acum, la teatrul de Crăciun participă unele dintre numele cele mai mari din showbusiness.
+Matthew Kelly și Hayley Mills vor apărea în Cenușăreasa - primul în rolul uneia dintre surorile rele, iar a doua în rolul zânei.
+Aceștia dezvăluie secretele pantomimei lor la BBC Breakfast.
+Steven Wilson: „Dacă nu fac nimic, mă simt vinovat”
+Steven Wilson a fost desemnat recent drept marele câștigător al Progressive Music Awards
+Steven Wilson a fost numit de multe ori drept cel mai muncitor muzician din lumea rockului progresiv.
+Talentatul muzician a câștigat trei premii la Progressive Music Awards, care a avut loc luna aceasta la Londra, printre care și premiul pentru cel mai bun album al anului pentru Hand.
+În recenzia sa de cinci stele, The Guardian a numit albumul „o operă de artă inteligentă, expresivă și captivantă”.
+Încă din anii 1980, Wilson este motorul mai multor proiecte muzicale, cel mai cunoscut dintre acestea fiind trupa de rock Porcupine Tree.
+Acum, înainte de două spectacole cu casa închisă la Royal Albert Hall, Wilson lansează un dublu LP doar în format vinil, Transience, pentru a arăta latura „mai accesibilă” a activității sale solo.
+A povestit pentru BBC despre dragostea lui pentru viniluri și despre programul său încărcat și a explicat cum a ajuns actorul de comedie Matt Berry să îi deschidă spectacolele.
+Ce înseamnă vinil pentru tine?
+Am crescut chiar în perioada de sfârșit a erei vinilurilor și îmi amintesc că atunci abia așteptam apariția CD-ului, căci vinilul era atât de enervant.
+Cumpărai un disc, mergeai cu el acasă, avea o zgârietură și trebuia să îl aduci înapoi.
+Iubesc CD-urile, iar pentru anumite tipuri de muzică, de exemplu cea clasică, sunt mai bune decât vinilurile.
+Însă problema cu CD-urile și cu descărcările digitale este aceea că nu mai există nimic pe care să îl prețuiești cu adevărat.
+Să ai un vinil e ca și cum ai avea un tablou frumos agățat în sufragerie.
+E ceva ce poți ține în mână, în timp ce te lași absorbit de versuri și copleșit de actul artistic.
+Am crezut că e doar o chestie nostalgică, însă nu are cum să fie așa dacă unor puști prea tineri să-și amintească de viniluri le place acest gen de experiență.
+Ai vreun vinil la care ții în mod special?
+Recunosc că am scăpat de toate vinilurile în anii '90.
+Toate vinilurile pe care le am sunt cumpărate din nou.
+Am pornit de la ideea de a reface colecția pe care o aveam la 15 ani, însă am trecut de limita aceea.
+Primul disc pe care mi-am convins părinții să mi-l cumpere a fost Out of the Blue de la Electric Light Orchestra.
+Dacă aș mai fi avut încă exemplarul inițial, acesta ar fi avut valoare sentimentală, însă, din păcate, se află pe undeva printr-un magazin de caritate.
+Steven Wilson speră că albumul va fi o poartă către posibili fani noi
+De ce ți-ai lansat noua compilație Transience pe vinil?
+Aceasta a fost concepută inițial ca idee pentru Ziua magazinelor de discuri, însă am ratat ocazia.
+Casa mea de discuri sugerase să adun câteva dintre melodiile mele mai scurte și mai accesibile.
+Am ajuns să fiu ușor obsedat de ideea de a face ceva gen „introducere în muzica lui Steven Wilson” și am ținut neapărat ca proiectul să fie lansat doar pe vinil.
+Cine cumpără vinilul primește, de asemenea, și o variantă descărcată la rezoluție înaltă.
+Ești îngrijorat că albumul nu va arăta muzica ta în adevărata ei lumină?
@@ -8,17 +8,15 @@ from pathlib import Path
 from unittest.mock import patch

 import pytest
-import pytorch_lightning as pl
 import torch
 from pytest import param
 from torch.utils.data import DataLoader

-import lightning_base
 from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, MBartTokenizer
 from transformers.testing_utils import require_multigpu

 from .distillation import distill_main, evaluate_checkpoint
-from .finetune import SummarizationModule, main
+from .finetune import main
 from .pack_dataset import pack_data_dir
 from .run_eval import generate_summaries_or_translations, run_generate
 from .utils import MBartDataset, Seq2SeqDataset, label_smoothed_nll_loss, lmap, load_json
@@ -45,7 +43,7 @@ CHEAP_ARGS = {
    "student_decoder_layers": 1,
    "val_check_interval": 1.0,
    "output_dir": "",
-    "fp16": False,  # TODO(SS): set this to CUDA_AVAILABLE if ci installs apex or start using native amp
+    "fp16": CUDA_AVAILABLE,
    "no_teacher": False,
    "fp16_opt_level": "O1",
    "gpus": 1 if CUDA_AVAILABLE else 0,
@@ -62,7 +60,6 @@ CHEAP_ARGS = {
    "tokenizer_name": "facebook/bart-large",
    "do_lower_case": False,
    "learning_rate": 0.3,
-    "lr_scheduler": "linear",
    "weight_decay": 0.0,
    "adam_epsilon": 1e-08,
    "warmup_steps": 0,
@@ -147,7 +144,7 @@ class TestSummarizationDistiller(unittest.TestCase):
        evaluate_checkpoint(ckpts[0], dest_dir=Path(tempfile.mkdtemp()))

    def test_loss_fn(self):
-        model = AutoModelForSeq2SeqLM.from_pretrained(BART_TINY, return_dict=True)
+        model = AutoModelForSeq2SeqLM.from_pretrained(BART_TINY)
        input_ids, mask = model.dummy_inputs["input_ids"], model.dummy_inputs["attention_mask"]
        target_ids = torch.tensor([[0, 4, 8, 2], [0, 8, 2, 1]], dtype=torch.long, device=model.device)
        decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
@@ -280,114 +277,6 @@ def test_finetune(model):
        assert bart.decoder.embed_tokens == bart.shared


-def test_finetune_extra_model_args():
-    args_d: dict = CHEAP_ARGS.copy()
-
-    task = "summarization"
-    tmp_dir = make_test_data_dir()
-
-    args_d.update(
-        data_dir=tmp_dir,
-        tokenizer_name=None,
-        train_batch_size=2,
-        eval_batch_size=2,
-        do_predict=False,
-        task=task,
-        src_lang="en_XX",
-        tgt_lang="ro_RO",
-        freeze_encoder=True,
-        freeze_embeds=True,
-    )
-
-    # test models whose config includes the extra_model_args
-    model = BART_TINY
-    output_dir = tempfile.mkdtemp(prefix="output_1_")
-    args_d1 = args_d.copy()
-    args_d1.update(
-        model_name_or_path=model, output_dir=output_dir,
-    )
-    extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout")
-    for p in extra_model_params:
-        args_d1[p] = 0.5
-    args = argparse.Namespace(**args_d1)
-    model = main(args)
-    for p in extra_model_params:
-        assert getattr(model.config, p) == 0.5, f"failed to override the model config for param {p}"
-
-    # test models whose config doesn't include the extra_model_args
-    model = T5_TINY
-    output_dir = tempfile.mkdtemp(prefix="output_2_")
-    args_d2 = args_d.copy()
-    args_d2.update(
-        model_name_or_path=model, output_dir=output_dir,
-    )
-    unsupported_param = "encoder_layerdrop"
-    args_d2[unsupported_param] = 0.5
-    args = argparse.Namespace(**args_d2)
-    with pytest.raises(Exception) as excinfo:
-        model = main(args)
-    assert str(excinfo.value) == f"model config doesn't have a `{unsupported_param}` attribute"
-
-
-def test_finetune_lr_shedulers(capsys):
-    args_d: dict = CHEAP_ARGS.copy()
-
-    task = "summarization"
-    tmp_dir = make_test_data_dir()
-
-    model = BART_TINY
-    output_dir = tempfile.mkdtemp(prefix="output_1_")
-
-    args_d.update(
-        data_dir=tmp_dir,
-        model_name_or_path=model,
-        output_dir=output_dir,
-        tokenizer_name=None,
-        train_batch_size=2,
-        eval_batch_size=2,
-        do_predict=False,
-        task=task,
-        src_lang="en_XX",
-        tgt_lang="ro_RO",
-        freeze_encoder=True,
-        freeze_embeds=True,
-    )
-
-    # emulate finetune.py
-    parser = argparse.ArgumentParser()
-    parser = pl.Trainer.add_argparse_args(parser)
-    parser = SummarizationModule.add_model_specific_args(parser, os.getcwd())
-    args = {"--help": True}
-
-    # --help test
-    with pytest.raises(SystemExit) as excinfo:
-        args = parser.parse_args(args)
-        assert False, "--help is expected to sys.exit"
-    assert excinfo.type == SystemExit
-    captured = capsys.readouterr()
-    expected = lightning_base.arg_to_scheduler_metavar
-    assert expected in captured.out, "--help is expected to list the supported schedulers"
-
-    # --lr_scheduler=non_existing_scheduler test
-    unsupported_param = "non_existing_scheduler"
-    args = {f"--lr_scheduler={unsupported_param}"}
-    with pytest.raises(SystemExit) as excinfo:
-        args = parser.parse_args(args)
-        assert False, "invalid argument is expected to sys.exit"
-    assert excinfo.type == SystemExit
-    captured = capsys.readouterr()
-    expected = f"invalid choice: '{unsupported_param}'"
-    assert expected in captured.err, f"should have bailed on invalid choice of scheduler {unsupported_param}"
-
-    # --lr_scheduler=existing_scheduler test
-    supported_param = "cosine"
-    args_d1 = args_d.copy()
-    args_d1["lr_scheduler"] = supported_param
-    args = argparse.Namespace(**args_d1)
-    model = main(args)
-    assert getattr(model.hparams, "lr_scheduler") == supported_param, f"lr_scheduler={supported_param} shouldn't fail"
-
-
 def test_pack_dataset():
    tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-cc25")

@@ -411,17 +300,14 @@ def test_mbart_dataset_truncation():
    tmp_dir = make_test_data_dir()
    max_len_source = max(len(tokenizer.encode(a)) for a in ARTICLES)
    max_len_target = max(len(tokenizer.encode(a)) for a in SUMMARIES)
-    max_src_len = 4
-    max_tgt_len = 8
-    assert max_len_target > max_src_len  # Truncated
-    assert max_len_source > max_src_len
+    trunc = 4
    src_lang, tgt_lang = "ro_RO", "de_DE"  # NOT WHAT IT WAS TRAINED ON
    train_dataset = MBartDataset(
        tokenizer,
        data_dir=tmp_dir,
        type_path="train",
-        max_source_length=max_src_len,
-        max_target_length=max_tgt_len,  # ignored
+        max_source_length=trunc,
+        max_target_length=1000,  # ignored
        src_lang=src_lang,
        tgt_lang=tgt_lang,
    )
@@ -430,15 +316,17 @@ def test_mbart_dataset_truncation():
        assert isinstance(batch, dict)
        assert batch["attention_mask"].shape == batch["input_ids"].shape
        # show that articles were trimmed.
-        assert batch["input_ids"].shape[1] == max_src_len
+        assert batch["input_ids"].shape[1] == trunc
        # show that targets are the same len
-        assert batch["decoder_input_ids"].shape[1] == max_tgt_len
+        assert batch["decoder_input_ids"].shape[1] == trunc
        # check language codes in correct place
        assert batch["decoder_input_ids"][0, 0].item() == tokenizer.lang_code_to_id[tgt_lang]
        assert batch["decoder_input_ids"][0, -1].item() == tokenizer.eos_token_id
        assert batch["input_ids"][0, -2].item() == tokenizer.eos_token_id
        assert batch["input_ids"][0, -1].item() == tokenizer.lang_code_to_id[src_lang]

+        assert max_len_target > trunc  # Truncated
+        assert max_len_source > trunc
        break  # No need to test every batch


@@ -5,14 +5,16 @@ python finetune.py \
    --learning_rate=3e-5 \
    --fp16 \
    --do_train \
-    --val_check_interval=0.25 \
+    --do_predict \
+    --val_check_interval 0.25 \
    --adam_eps 1e-06 \
    --num_train_epochs 6 --src_lang en_XX --tgt_lang ro_RO \
    --data_dir $ENRO_DIR \
    --max_source_length $MAX_LEN --max_target_length $MAX_LEN --val_max_target_length $MAX_LEN --test_max_target_length $MAX_LEN \
-    --train_batch_size=$BS --eval_batch_size=$BS \
+    --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS \
    --task translation \
    --warmup_steps 500 \
    --freeze_embeds \
-    --model_name_or_path=facebook/mbart-large-cc25 \
+    --early_stopping_patience 4 \
+    --model_name_or_path facebook/mbart-large-cc25 \
    $@
@@ -128,6 +128,12 @@ class Seq2SeqDataset(Dataset):
    def get_char_lens(data_file):
        return [len(x) for x in Path(data_file).open().readlines()]

+    @staticmethod
+    def trim_seq2seq_batch(batch, pad_token_id) -> tuple:
+        y = trim_batch(batch["decoder_input_ids"], pad_token_id)
+        source_ids, source_mask = trim_batch(batch["input_ids"], pad_token_id, attention_mask=batch["attention_mask"])
+        return source_ids, source_mask, y
+
    def collate_fn(self, batch) -> Dict[str, torch.Tensor]:
        input_ids = torch.stack([x["input_ids"] for x in batch])
        masks = torch.stack([x["attention_mask"] for x in batch])
@@ -151,8 +157,7 @@ class MBartDataset(Seq2SeqDataset):
        super().__init__(*args, **kwargs)
        if self.max_source_length != self.max_target_length:
            warnings.warn(
-                f"Mbart is using sequence lengths {self.max_source_length}, {self.max_target_length}. "
-                f"Imbalanced sequence lengths may be undesired for translation tasks"
+                f"Mbart will ignore max_target_length = {self.max_target_length} and use {self.max_source_length} for both sides."
            )

    def __getitem__(self, index) -> Dict[str, str]:
@@ -173,7 +178,6 @@ class MBartDataset(Seq2SeqDataset):
            tgt_texts=[x["tgt_texts"] for x in batch],
            tgt_lang=self.tgt_lang,
            max_length=self.max_source_length,
-            max_target_length=self.max_target_length,
        )
        return batch_encoding.data

@@ -9,7 +9,6 @@ from enum import Enum
 from typing import Dict, Optional

 import numpy as np
-import tensorflow as tf
 import tensorflow_datasets as tfds

 from transformers import (
@@ -36,11 +35,7 @@ class Split(Enum):


 def get_tfds(
-    task_name: str,
-    tokenizer: PreTrainedTokenizer,
-    max_seq_length: Optional[int] = None,
-    mode: Split = Split.train,
-    data_dir: str = None,
+    task_name: str, tokenizer: PreTrainedTokenizer, max_seq_length: Optional[int] = None, mode: Split = Split.train
 ):
    if task_name == "mnli-mm" and mode == Split.dev:
        tfds_name = "mnli_mismatched"
@@ -55,11 +50,9 @@ def get_tfds(
    else:
        tfds_name = task_name

-    ds, info = tfds.load("glue/" + tfds_name, split=mode.value, with_info=True, data_dir=data_dir)
-    ds = glue_convert_examples_to_features(ds, tokenizer, max_seq_length, task_name)
-    ds = ds.apply(tf.data.experimental.assert_cardinality(info.splits[mode.value].num_examples))
+    ds = tfds.load("glue/" + tfds_name, split=mode.value)

-    return ds
+    return glue_convert_examples_to_features(ds, tokenizer, max_seq_length, task_name)


 logger = logging.getLogger(__name__)
@@ -76,7 +69,6 @@ class GlueDataTrainingArguments:
    """

    task_name: str = field(metadata={"help": "The name of the task to train on: " + ", ".join(glue_processors.keys())})
-    data_dir: Optional[str] = field(default=None, metadata={"help": "The input/output data dir for TFDS."})
    max_seq_length: int = field(
        default=128,
        metadata={
@@ -179,22 +171,13 @@ def main():

    # Get datasets
    train_dataset = (
-        get_tfds(
-            task_name=data_args.task_name,
-            tokenizer=tokenizer,
-            max_seq_length=data_args.max_seq_length,
-            data_dir=data_args.data_dir,
-        )
+        get_tfds(task_name=data_args.task_name, tokenizer=tokenizer, max_seq_length=data_args.max_seq_length)
        if training_args.do_train
        else None
    )
    eval_dataset = (
        get_tfds(
-            task_name=data_args.task_name,
-            tokenizer=tokenizer,
-            max_seq_length=data_args.max_seq_length,
-            mode=Split.dev,
-            data_dir=data_args.data_dir,
+            task_name=data_args.task_name, tokenizer=tokenizer, max_seq_length=data_args.max_seq_length, mode=Split.dev
        )
        if training_args.do_eval
        else None
@@ -17,6 +17,7 @@

 import logging
 import os
+import warnings
 from dataclasses import dataclass, field
 from typing import Dict, List, Optional, Tuple

@@ -184,6 +185,11 @@ def main():

        for i in range(batch_size):
            for j in range(seq_len):
+                if label_ids[i, j] == -1:
+                    label_ids[i, j] = -100
+                    warnings.warn(
+                        "Using `-1` to mask the loss for the token is depreciated. Please use `-100` instead."
+                    )
                if label_ids[i, j] != -100:
                    out_label_list[i].append(label_map[label_ids[i][j]])
                    preds_list[i].append(label_map[preds[i][j]])
@@ -146,7 +146,7 @@ if is_tf_available():
        """

        features: List[InputFeatures]
-        pad_token_label_id: int = -100
+        pad_token_label_id: int = -1
        # Use cross entropy ignore_index as padding label id so that only
        # real label ids contribute to the loss later.

@@ -221,8 +221,6 @@ if is_tf_available():
                )

        def get_dataset(self):
-            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
-
            return self.dataset

        def __len__(self):
@@ -1,18 +1,2 @@
 This model is used detecting **hatespeech** in **Arabic language**. The mono in the name refers to the monolingual setting, where the model is trained using only Arabic language data. It is finetuned on multilingual bert model.
 The model is trained with different learning rates and the best validation score achieved is 0.877609 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)
-
-### For more details about our paper
-
-Sai Saketh Aluru, Binny Mathew, Punyajoy Saha and Animesh Mukherjee. "[Deep Learning Models for Multilingual Hate Speech Detection](https://arxiv.org/abs/2004.06465)". Accepted at ECML-PKDD 2020.
-
-***Please cite our paper in any published work that uses any of these resources.***
-
-~~~
-@article{aluru2020deep,
-  title={Deep Learning Models for Multilingual Hate Speech Detection},
-  author={Aluru, Sai Saket and Mathew, Binny and Saha, Punyajoy and Mukherjee, Animesh},
-  journal={arXiv preprint arXiv:2004.06465},
-  year={2020}
-}
-
-~~~
@@ -1,47 +0,0 @@
---
-language: de
-license: mit
---
-
-# bert-german-dbmdz-uncased-sentence-stsb
-
-## How to use
-**The usage description above - provided by Hugging Face - is wrong! Please use this:**
-
-Install the `sentence-transformers` package. See here: <https://github.com/UKPLab/sentence-transformers>
-```python
-from sentence_transformers import models
-from sentence_transformers import SentenceTransformer
-
-# load BERT model from Hugging Face
-word_embedding_model = models.Transformer(
-    'T-Systems-onsite/bert-german-dbmdz-uncased-sentence-stsb')
-
-# Apply mean pooling to get one fixed sized sentence vector
-pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension(),
-                               pooling_mode_mean_tokens=True,
-                               pooling_mode_cls_token=False,
-                               pooling_mode_max_tokens=False)
-
-# join BERT model and pooling to get the sentence transformer
-model = SentenceTransformer(modules=[word_embedding_model, pooling_model])
-```
-
-## Model description
-This is a German [sentence embedding](https://github.com/UKPLab/sentence-transformers) trained on the [German STSbenchmark Dataset](https://github.com/t-systems-on-site-services-gmbh/german-STSbenchmark). It was trained from [Philip May](https://eniak.de/) and open-sourced by [T-Systems-onsite](https://www.t-systems-onsite.de/).The base language model is the [dbmdz/bert-base-german-uncased](https://huggingface.co/dbmdz/bert-base-german-uncased) from [Bayerische Staatsbibliothek ](https://huggingface.co/dbmdz).
-
-## Intended uses
-> Sentence-BERT (SBERT) is a  modification  of  the  pretrained BERT network that use siamese and triplet network structures to derive semantically mean-ingful sentence embeddings that can be compared using cosine-similarity. This reduces the effort for finding the most similar pair from 65hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT.
-
-Source: [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://arxiv.org/abs/1908.10084)
-
-## Training procedure
-We did an automatic hyperprameter optimization with [Optuna](https://github.com/optuna/optuna) and found the following hyperprameters:
- batch_size = 5
- num_epochs = 11
- lr = 2.637549780860126e-05
- eps = 5.0696075038683e-06
- weight_decay = 0.02817210102940054
- warmup_steps = 27.342745941760147 % of total steps
-
-The final model was trained on the combination of all three datasets: `sts_de_dev.csv`, `sts_de_test.csv` and `sts_de_train.csv`
@@ -1,51 +0,0 @@
---
-language: "en"
-tags:
- paraphrase-generation
- text-generation
- Conditional Generation
-inference: false
---
-
-# Paraphrase-Generation
-
-## Model description
-
-T5 Model for generating paraphrases of english sentences. Trained on the [Google PAWS](https://github.com/google-research-datasets/paws) dataset.
-
-## How to use
-
-PyTorch and TF models available
-
-```python
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
-
-tokenizer = AutoTokenizer.from_pretrained("Vamsi/T5_Paraphrase_Paws")  
-model = AutoModelForSeq2SeqLM.from_pretrained("Vamsi/T5_Paraphrase_Paws")
-
-sentence = "This is something which i cannot understand at all"
-
-text =  "paraphrase: " + sentence + " </s>"
-
-encoding = tokenizer.encode_plus(text,pad_to_max_length=True, return_tensors="pt")
-input_ids, attention_masks = encoding["input_ids"].to("cuda"), encoding["attention_mask"].to("cuda")
-
-
-outputs = model.generate(
-    input_ids=input_ids, attention_mask=attention_masks,
-    max_length=256,
-    do_sample=True,
-    top_k=120,
-    top_p=0.95,
-    early_stopping=True,
-    num_return_sequences=5
-)
-
-for output in outputs:
-    line = tokenizer.decode(output, skip_special_tokens=True,clean_up_tokenization_spaces=True)
-    print(line)
-
-
-```
-
-For more reference on training your own T5 model or using this model, do check out [Paraphrase Generation](https://github.com/Vamsi995/Paraphrase-Generator).
@@ -1,25 +0,0 @@
---
-language: en
-datasets:
- yelp_polarity
---
-
-# RoBERTa-base-finetuned-yelp-polarity
-
-This is a [RoBERTa-base](https://huggingface.co/roberta-base) checkpoint fine-tuned on binary sentiment classifcation from [Yelp polarity](https://huggingface.co/nlp/viewer/?dataset=yelp_polarity).
-It gets **98.08%** accuracy on the test set.
-
-## Hyper-parameters
-
-We used the following hyper-parameters to train the model on one GPU:
-```python
-num_train_epochs            = 2.0
-learning_rate               = 1e-05
-weight_decay                = 0.0
-adam_epsilon                = 1e-08
-max_grad_norm               = 1.0
-per_device_train_batch_size = 32
-gradient_accumulation_steps = 1
-warmup_steps                = 3500
-seed                        = 42
-```
@@ -1,35 +1,20 @@
 ---
 language: ar
-datasets:
- oscar
- wikipedia
 ---

 # Arabic BERT Model

 Pretrained BERT base language model for Arabic

+_If you use this model in your work, please cite this paper (to appear in 2020):_

-_If you use this model in your work, please cite this paper:_
-
-<!--```
+```
@inproceedings{
  title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
  author={Safaya, Ali and Abdullatif, Moutasem and Yuret, Deniz},
  booktitle={Proceedings of the International Workshop on Semantic Evaluation (SemEval)},
  year={2020}
 }
-```-->
-
-```
-@misc{safaya2020kuisail,
-    title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
-    author={Ali Safaya and Moutasem Abdullatif and Deniz Yuret},
-    year={2020},
-    eprint={2007.13184},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
 ```

 ## Pretraining Corpus
@@ -1,35 +1,20 @@
 ---
 language: ar
-datasets:
- oscar
- wikipedia
 ---

-
 # Arabic BERT Large Model

 Pretrained BERT Large language model for Arabic

-_If you use this model in your work, please cite this paper:_
+_If you use this model in your work, please cite this paper (to appear in 2020):_

-<!--```
+```
@inproceedings{
  title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
  author={Safaya, Ali and Abdullatif, Moutasem and Yuret, Deniz},
  booktitle={Proceedings of the International Workshop on Semantic Evaluation (SemEval)},
  year={2020}
 }
-```-->
-
-```
-@misc{safaya2020kuisail,
-    title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
-    author={Ali Safaya and Moutasem Abdullatif and Deniz Yuret},
-    year={2020},
-    eprint={2007.13184},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
 ```

 ## Pretraining Corpus
@@ -1,34 +1,20 @@
 ---
 language: ar
-datasets:
- oscar
- wikipedia
 ---

 # Arabic BERT Medium Model

 Pretrained BERT Medium language model for Arabic

-_If you use this model in your work, please cite this paper:_
+_If you use this model in your work, please cite this paper (to appear in 2020):_

-<!--```
+```
@inproceedings{
  title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
  author={Safaya, Ali and Abdullatif, Moutasem and Yuret, Deniz},
  booktitle={Proceedings of the International Workshop on Semantic Evaluation (SemEval)},
  year={2020}
 }
-```-->
-
-```
-@misc{safaya2020kuisail,
-    title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
-    author={Ali Safaya and Moutasem Abdullatif and Deniz Yuret},
-    year={2020},
-    eprint={2007.13184},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
 ```

 ## Pretraining Corpus
@@ -9,26 +9,15 @@ datasets:

 Pretrained BERT Mini language model for Arabic

-_If you use this model in your work, please cite this paper:_
+_If you use this model in your work, please cite this paper (to appear in 2020):_

-<!--```
+```
@inproceedings{
  title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
  author={Safaya, Ali and Abdullatif, Moutasem and Yuret, Deniz},
  booktitle={Proceedings of the International Workshop on Semantic Evaluation (SemEval)},
  year={2020}
 }
-```-->
-
-```
-@misc{safaya2020kuisail,
-    title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
-    author={Ali Safaya and Moutasem Abdullatif and Deniz Yuret},
-    year={2020},
-    eprint={2007.13184},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
 ```

 ## Pretraining Corpus
@@ -1,6 +1,5 @@
 ---
 language: de
-license: mit
 thumbnail: https://static.tildacdn.com/tild6438-3730-4164-b266-613634323466/german_bert.png
 tags:
 - exbert
@@ -1,113 +0,0 @@
---
-language: en
-thumbnail: 
-license: mit
-tags:
- question-answering
- bert
- bert-base
-datasets:
- squad
-metrics:
- squad
-widget:
- text: "Which name is also used to describe the Amazon rainforest in English?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
- text: "How many square kilometers of rainforest is covered in the basin?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
---
-
-## BERT-base uncased model fine-tuned on SQuAD v1
-
-This model was fine-tuned from the HuggingFace [BERT](https://www.aclweb.org/anthology/N19-1423/) base uncased checkpoint on [SQuAD1.1](https://rajpurkar.github.io/SQuAD-explorer).
-This model is case-insensitive: it does not make a difference between english and English.
-
-## Details
-
-| Dataset  | Split | # samples |
-| -------- | ----- | --------- |
-| SQuAD1.1 | train | 90.6K      |
-| SQuAD1.1 | eval  | 11.1k     |
-
-
-### Fine-tuning
- Python: `3.7.5`
-
- Machine specs: 
-
-  `CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz`
-  
-  `Memory: 32 GiB`
-
-  `GPUs: 2 GeForce GTX 1070, each with 8GiB memory`
-  
-  `GPU driver: 418.87.01, CUDA: 10.1`
-
- script:
-
-  ```shell
-  # after install https://github.com/huggingface/transformers
-
-  cd examples/question-answering
-  mkdir -p data
-
-  wget -O data/train-v1.1.json https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json
-
-  wget -O data/dev-v1.1.json  https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json
-
-  python run_squad.py \
-    --model_type bert \
-    --model_name_or_path bert-base-uncased \
-    --do_train \
-    --do_eval \
-    --do_lower_case \
-    --train_file train-v1.1.json \
-    --predict_file dev-v1.1.json \
-    --per_gpu_train_batch_size 12 \
-    --per_gpu_eval_batch_size=16 \
-    --learning_rate 3e-5 \
-    --num_train_epochs 2.0 \
-    --max_seq_length 320 \
-    --doc_stride 128 \
-    --data_dir data \
-    --output_dir data/bert-base-uncased-squad-v1 2>&1 | tee train-energy-bert-base-squad-v1.log
-  ```
-
-It took about 2 hours to finish.
-
-### Results
-
-**Model size**: `418M`
-
-| Metric | # Value   | # Original ([Table 2](https://www.aclweb.org/anthology/N19-1423.pdf))|
-| ------ | --------- | --------- |
-| **EM** | **80.9** | **80.8** |
-| **F1** | **88.2** | **88.5** |
-
-Note that the above results didn't involve any hyperparameter search.
-
-## Example Usage
-
-
-```python
-from transformers import pipeline
-
-qa_pipeline = pipeline(
-    "question-answering",
-    model="csarron/bert-base-uncased-squad-v1",
-    tokenizer="csarron/bert-base-uncased-squad-v1"
-)
-
-predictions = qa_pipeline({
-    'context': "The game was played on February 7, 2016 at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California.",
-    'question': "What day was the game played on?"
-})
-
-print(predictions)
-# output:
-# {'score': 0.8730505704879761, 'start': 23, 'end': 39, 'answer': 'February 7, 2016'}
-```
-
-> Created by [Qingqing Cao](https://awk.ai/) | [GitHub](https://github.com/csarron) | [Twitter](https://twitter.com/sysnlp) 
-
-> Made with ❤️ in New York.
@@ -1,116 +0,0 @@
---
-language: en
-thumbnail: 
-license: mit
-tags:
- question-answering
- mobilebert
-datasets:
- squad
-metrics:
- squad
-widget:
- text: "Which name is also used to describe the Amazon rainforest in English?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
- text: "How many square kilometers of rainforest is covered in the basin?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
---
-
-## MobileBERT fine-tuned on SQuAD v1
-
-[MobileBERT](https://arxiv.org/abs/2004.02984) is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance
-between self-attentions and feed-forward networks.
-
-This model was fine-tuned from the HuggingFace checkpoint `google/mobilebert-uncased` on [SQuAD1.1](https://rajpurkar.github.io/SQuAD-explorer).
-
-## Details
-
-| Dataset  | Split | # samples |
-| -------- | ----- | --------- |
-| SQuAD1.1 | train | 90.6K      |
-| SQuAD1.1 | eval  | 11.1k     |
-
-
-### Fine-tuning
- Python: `3.7.5`
-
- Machine specs: 
-
-  `CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz`
-  
-  `Memory: 32 GiB`
-
-  `GPUs: 2 GeForce GTX 1070, each with 8GiB memory`
-  
-  `GPU driver: 418.87.01, CUDA: 10.1`
-
- script:
-
-  ```shell
-  # after install https://github.com/huggingface/transformers
-
-  cd examples/question-answering
-  mkdir -p data
-
-  wget -O data/train-v1.1.json https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json
-
-  wget -O data/dev-v1.1.json  https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json
-
-  export SQUAD_DIR=`pwd`/data
-
-  python run_squad.py \
-    --model_type mobilebert \
-    --model_name_or_path google/mobilebert-uncased \
-    --do_train \
-    --do_eval \
-    --do_lower_case \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
-    --per_gpu_train_batch_size 16 \
-    --per_gpu_eval_batch_size 16 \
-    --learning_rate 4e-5 \
-    --num_train_epochs 5.0 \
-    --max_seq_length 320 \
-    --doc_stride 128 \
-    --warmup_steps 1400 \
-    --output_dir $SQUAD_DIR/mobilebert-uncased-warmup-squad_v1 2>&1 | tee train-mobilebert-warmup-squad_v1.log
-  ```
-
-It took about 3 hours to finish.
-
-### Results
-
-**Model size**: `95M`
-
-| Metric | # Value   | # Original ([Table 5](https://arxiv.org/pdf/2004.02984.pdf))|
-| ------ | --------- | --------- |
-| **EM** | **82.6** | **82.9** |
-| **F1** | **90.0** | **90.0** |
-
-Note that the above results didn't involve any hyperparameter search.
-
-## Example Usage
-
-
-```python
-from transformers import pipeline
-
-qa_pipeline = pipeline(
-    "question-answering",
-    model="csarron/mobilebert-uncased-squad-v1",
-    tokenizer="csarron/mobilebert-uncased-squad-v1"
-)
-
-predictions = qa_pipeline({
-    'context': "The game was played on February 7, 2016 at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California.",
-    'question': "What day was the game played on?"
-})
-
-print(predictions)
-# output:
-# {'score': 0.7754058241844177, 'start': 23, 'end': 39, 'answer': 'February 7, 2016'}
-```
-
-> Created by [Qingqing Cao](https://awk.ai/) | [GitHub](https://github.com/csarron) | [Twitter](https://twitter.com/sysnlp) 
-
-> Made with ❤️ in New York.
@@ -1,118 +0,0 @@
---
-language: en
-thumbnail: 
-license: mit
-tags:
- question-answering
- mobilebert
-datasets:
- squad_v2
-metrics:
- squad_v2
-widget:
- text: "Which name is also used to describe the Amazon rainforest in English?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
- text: "How many square kilometers of rainforest is covered in the basin?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
---
-
-## MobileBERT fine-tuned on SQuAD v2
-
-[MobileBERT](https://arxiv.org/abs/2004.02984) is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance
-between self-attentions and feed-forward networks.
-
-This model was fine-tuned from the HuggingFace checkpoint `google/mobilebert-uncased` on [SQuAD2.0](https://rajpurkar.github.io/SQuAD-explorer).
-
-## Details
-
-| Dataset  | Split | # samples |
-| -------- | ----- | --------- |
-| SQuAD2.0 | train | 130k      |
-| SQuAD2.0 | eval  | 12.3k     |
-
-
-### Fine-tuning
- Python: `3.7.5`
-
- Machine specs: 
-
-  `CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz`
-  
-  `Memory: 32 GiB`
-
-  `GPUs: 2 GeForce GTX 1070, each with 8GiB memory`
-  
-  `GPU driver: 418.87.01, CUDA: 10.1`
-
- script:
-
-  ```shell
-  # after install https://github.com/huggingface/transformers
-
-  cd examples/question-answering
-  mkdir -p data
-
-  wget -O data/train-v2.0.json https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json
-
-  wget -O data/dev-v2.0.json  https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json
-
-  export SQUAD_DIR=`pwd`/data
-
-  python run_squad.py \
-    --model_type mobilebert \
-    --model_name_or_path google/mobilebert-uncased \
-    --do_train \
-    --do_eval \
-    --do_lower_case \
-    --version_2_with_negative \
-    --train_file $SQUAD_DIR/train-v2.0.json \
-    --predict_file $SQUAD_DIR/dev-v2.0.json \
-    --per_gpu_train_batch_size 16 \
-    --per_gpu_eval_batch_size 16 \
-    --learning_rate 4e-5 \
-    --num_train_epochs 5.0 \
-    --max_seq_length 320 \
-    --doc_stride 128 \
-    --warmup_steps 1400 \
-    --save_steps 2000 \
-    --output_dir $SQUAD_DIR/mobilebert-uncased-warmup-squad_v2 2>&1 | tee train-mobilebert-warmup-squad_v2.log
-  ```
-
-It took about 3.5 hours to finish.
-
-### Results
-
-**Model size**: `95M`
-
-| Metric | # Value   | # Original ([Table 5](https://arxiv.org/pdf/2004.02984.pdf))|
-| ------ | --------- | --------- |
-| **EM** | **75.2** | **76.2** |
-| **F1** | **78.8** | **79.2** |
-
-Note that the above results didn't involve any hyperparameter search.
-
-## Example Usage
-
-
-```python
-from transformers import pipeline
-
-qa_pipeline = pipeline(
-    "question-answering",
-    model="csarron/mobilebert-uncased-squad-v2",
-    tokenizer="csarron/mobilebert-uncased-squad-v2"
-)
-
-predictions = qa_pipeline({
-    'context': "The game was played on February 7, 2016 at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California.",
-    'question': "What day was the game played on?"
-})
-
-print(predictions)
-# output:
-# {'score': 0.71434086561203, 'start': 23, 'end': 39, 'answer': 'February 7, 2016'}
-```
-
-> Created by [Qingqing Cao](https://awk.ai/) | [GitHub](https://github.com/csarron) | [Twitter](https://twitter.com/sysnlp) 
-
-> Made with ❤️ in New York.
@@ -1,112 +0,0 @@
---
-language: en
-thumbnail: 
-license: mit
-tags:
- question-answering
- roberta
- roberta-base
-datasets:
- squad
-metrics:
- squad
-widget:
- text: "Which name is also used to describe the Amazon rainforest in English?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
- text: "How many square kilometers of rainforest is covered in the basin?"
-  context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
---
-
-## RoBERTa-base fine-tuned on SQuAD v1
-
-This model was fine-tuned from the HuggingFace [RoBERTa](https://arxiv.org/abs/1907.11692) base checkpoint on [SQuAD1.1](https://rajpurkar.github.io/SQuAD-explorer).
-This model is case-sensitive: it makes a difference between english and English.
-
-## Details
-
-| Dataset  | Split | # samples |
-| -------- | ----- | --------- |
-| SQuAD1.1 | train | 96.8K      |
-| SQuAD1.1 | eval  | 11.8k     |
-
-
-### Fine-tuning
- Python: `3.7.5`
-
- Machine specs: 
-
-  `CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz`
-  
-  `Memory: 32 GiB`
-
-  `GPUs: 2 GeForce GTX 1070, each with 8GiB memory`
-  
-  `GPU driver: 418.87.01, CUDA: 10.1`
-
- script:
-
-  ```shell
-  # after install https://github.com/huggingface/transformers
-
-  cd examples/question-answering
-  mkdir -p data
-
-  wget -O data/train-v1.1.json https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json
-
-  wget -O data/dev-v1.1.json  https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json
-
-  python run_energy_squad.py \
-    --model_type roberta \
-    --model_name_or_path roberta-base \
-    --do_train \
-    --do_eval \
-    --train_file train-v1.1.json \
-    --predict_file dev-v1.1.json \
-    --per_gpu_train_batch_size 12 \
-    --per_gpu_eval_batch_size 16 \
-    --learning_rate 3e-5 \
-    --num_train_epochs 2.0 \
-    --max_seq_length 320 \
-    --doc_stride 128 \
-    --data_dir data \
-    --output_dir data/roberta-base-squad-v1 2>&1 | tee train-roberta-base-squad-v1.log
-  ```
-
-It took about 2 hours to finish.
-
-### Results
-
-**Model size**: `477M`
-
-| Metric | # Value   |
-| ------ | --------- |
-| **EM** | **83.0** |
-| **F1** | **90.4** |
-
-Note that the above results didn't involve any hyperparameter search.
-
-## Example Usage
-
-
-```python
-from transformers import pipeline
-
-qa_pipeline = pipeline(
-    "question-answering",
-    model="csarron/roberta-base-squad-v1",
-    tokenizer="csarron/roberta-base-squad-v1"
-)
-
-predictions = qa_pipeline({
-    'context': "The game was played on February 7, 2016 at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California.",
-    'question': "What day was the game played on?"
-})
-
-print(predictions)
-# output:
-# {'score': 0.8625259399414062, 'start': 23, 'end': 39, 'answer': 'February 7, 2016'}
-```
-
-> Created by [Qingqing Cao](https://awk.ai/) | [GitHub](https://github.com/csarron) | [Twitter](https://twitter.com/sysnlp) 
-
-> Made with ❤️ in New York.
@@ -1,6 +1,5 @@
 ---
 language: de
-license: mit
 thumbnail: https://static.tildacdn.com/tild6438-3730-4164-b266-613634323466/german_bert.png
 tags:
 - exbert
@@ -1,118 +0,0 @@
---
-datasets:
- squad_v2
---
-
-# MiniLM-L12-H384-uncased for QA
-
-## Overview
-**Language model:** microsoft/MiniLM-L12-H384-uncased
-**Language:** English
-**Downstream-task:** Extractive QA
-**Training data:** SQuAD 2.0
-**Eval data:** SQuAD 2.0
-**Code:**  See [example](https://github.com/deepset-ai/FARM/blob/master/examples/question_answering.py) in [FARM](https://github.com/deepset-ai/FARM/blob/master/examples/question_answering.py)
-**Infrastructure**: 1x Tesla v100
-
-## Hyperparameters
-
-```
-seed=42
-batch_size = 12
-n_epochs = 4
-base_LM_model = "microsoft/MiniLM-L12-H384-uncased"
-max_seq_len = 384
-learning_rate = 4e-5
-lr_schedule = LinearWarmup
-warmup_proportion = 0.2
-doc_stride=128
-max_query_length=64
-grad_acc_steps=4
-```
-
-## Performance
-Evaluated on the SQuAD 2.0 dev set with the [official eval script](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/).
-```
-"exact": 76.13071675229513,
-"f1": 79.49786500219953,
-"total": 11873,
-"HasAns_exact": 78.35695006747639,
-"HasAns_f1": 85.10090269418276,
-"HasAns_total": 5928,
-"NoAns_exact": 73.91084945332211,
-"NoAns_f1": 73.91084945332211,
-"NoAns_total": 5945
-```
-
-## Usage
-
-### In Transformers
-```python
-from transformers.pipelines import pipeline
-from transformers.modeling_auto import AutoModelForQuestionAnswering
-from transformers.tokenization_auto import AutoTokenizer
-
-model_name = "deepset/minilm-uncased-squad2"
-
-# a) Get predictions
-nlp = pipeline('question-answering', model=model_name, tokenizer=model_name)
-QA_input = {
-    'question': 'Why is model conversion important?',
-    'context': 'The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks.'
-}
-res = nlp(QA_input)
-
-# b) Load model & tokenizer
-model = AutoModelForQuestionAnswering.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-```
-
-### In FARM
-
-```python
-from farm.modeling.adaptive_model import AdaptiveModel
-from farm.modeling.tokenization import Tokenizer
-from farm.infer import Inferencer
-
-model_name = "deepset/minilm-uncased-squad2"
-
-# a) Get predictions
-nlp = Inferencer.load(model_name, task_type="question_answering")
-QA_input = [{"questions": ["Why is model conversion important?"],
-             "text": "The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks."}]
-res = nlp.inference_from_dicts(dicts=QA_input)
-
-# b) Load model & tokenizer
-model = AdaptiveModel.convert_from_transformers(model_name, device="cpu", task_type="question_answering")
-tokenizer = Tokenizer.load(model_name)
-```
-
-### In haystack
-For doing QA at scale (i.e. many docs instead of single paragraph), you can load the model also in [haystack](https://github.com/deepset-ai/haystack/):
-```python
-reader = FARMReader(model_name_or_path="deepset/minilm-uncased-squad2")
-# or
-reader = TransformersReader(model="deepset/minilm-uncased-squad2",tokenizer="deepset/minilm-uncased-squad2")
-```
-
-
-## Authors
-Vaishali Pal `vaishali.pal [at] deepset.ai`
-Branden Chan: `branden.chan [at] deepset.ai`
-Timo Möller: `timo.moeller [at] deepset.ai`
-Malte Pietsch: `malte.pietsch [at] deepset.ai`
-Tanay Soni: `tanay.soni [at] deepset.ai`
-
-## About us
-![deepset logo](https://raw.githubusercontent.com/deepset-ai/FARM/master/docs/img/deepset_logo.png)
-
-We bring NLP to the industry via open source!
-Our focus: Industry specific language models & large scale QA systems.
-
-Some of our work:
- [German BERT (aka "bert-base-german-cased")](https://deepset.ai/german-bert)
- [FARM](https://github.com/deepset-ai/FARM)
- [Haystack](https://github.com/deepset-ai/haystack/)
-
-Get in touch:
-[Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Website](https://deepset.ai)
@@ -1,133 +0,0 @@
---
-language: multilingual
-tags:
- question-answering
-datasets:
- squad_v2
---
-
-# Multilingual XLM-RoBERTa large for QA on various languages 
-
-## Overview
-**Language model:** xlm-roberta-large  
-**Language:** Multilingual  
-**Downstream-task:** Extractive QA  
-**Training data:** SQuAD 2.0  
-**Eval data:** SQuAD dev set - German MLQA - German XQuAD   
-**Training run:** [MLFlow link](https://public-mlflow.deepset.ai/#/experiments/124/runs/3a540e3f3ecf4dd98eae8fc6d457ff20)  
-**Infrastructure**: 4x Tesla v100
-
-## Hyperparameters
-
-```
-batch_size = 32
-n_epochs = 3
-base_LM_model = "xlm-roberta-large"
-max_seq_len = 256
-learning_rate = 1e-5
-lr_schedule = LinearWarmup
-warmup_proportion = 0.2
-doc_stride=128
-max_query_length=64
-``` 
-
-## Performance
-Evaluated on the SQuAD 2.0 English dev set with the [official eval script](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/).
-```
-  "exact": 79.45759285774446,
-  "f1": 83.79259828925511,
-  "total": 11873,
-  "HasAns_exact": 71.96356275303644,
-  "HasAns_f1": 80.6460053117963,
-  "HasAns_total": 5928,
-  "NoAns_exact": 86.93019343986543,
-  "NoAns_f1": 86.93019343986543,
-  "NoAns_total": 5945
-```
-
-Evaluated on German [MLQA: test-context-de-question-de.json](https://github.com/facebookresearch/MLQA)
-```
-"exact": 49.34691166703564,
-"f1": 66.15582561674236,
-"total": 4517,
-```
-
-Evaluated on German [XQuAD: xquad.de.json](https://github.com/deepmind/xquad)
-```
-"exact": 61.51260504201681,
-"f1": 78.80206098332569,
-"total": 1190,
-```
-
-## Usage
-
-### In Transformers
-```python
-from transformers.pipelines import pipeline
-from transformers.modeling_auto import AutoModelForQuestionAnswering
-from transformers.tokenization_auto import AutoTokenizer
-
-model_name = "deepset/xlm-roberta-large-squad2"
-
-# a) Get predictions
-nlp = pipeline('question-answering', model=model_name, tokenizer=model_name)
-QA_input = {
-    'question': 'Why is model conversion important?',
-    'context': 'The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks.'
-}
-res = nlp(QA_input)
-
-# b) Load model & tokenizer
-model = AutoModelForQuestionAnswering.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-```
-
-### In FARM
-
-```python
-from farm.modeling.adaptive_model import AdaptiveModel
-from farm.modeling.tokenization import Tokenizer
-from farm.infer import QAInferencer
-
-model_name = "deepset/xlm-roberta-large-squad2"
-
-# a) Get predictions
-nlp = QAInferencer.load(model_name)
-QA_input = [{"questions": ["Why is model conversion important?"],
-             "text": "The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks."}]
-res = nlp.inference_from_dicts(dicts=QA_input, rest_api_schema=True)
-
-# b) Load model & tokenizer
-model = AdaptiveModel.convert_from_transformers(model_name, device="cpu", task_type="question_answering")
-tokenizer = Tokenizer.load(model_name)
-```
-
-### In haystack
-For doing QA at scale (i.e. many docs instead of single paragraph), you can load the model also in [haystack](https://github.com/deepset-ai/haystack/):
-```python
-reader = FARMReader(model_name_or_path="deepset/xlm-roberta-large-squad2")
-# or 
-reader = TransformersReader(model="deepset/xlm-roberta-large-squad2",tokenizer="deepset/xlm-roberta-large-squad2")
-```
-
-
-## Authors
-Branden Chan: `branden.chan [at] deepset.ai`  
-Timo Möller: `timo.moeller [at] deepset.ai`  
-Malte Pietsch: `malte.pietsch [at] deepset.ai`  
-Tanay Soni: `tanay.soni [at] deepset.ai`
-
-## About us
-![deepset logo](https://raw.githubusercontent.com/deepset-ai/FARM/master/docs/img/deepset_logo.png)
-
-We bring NLP to the industry via open source!  
-Our focus: Industry specific language models & large scale QA systems.  
-  
-Some of our work: 
- [German BERT (aka "bert-base-german-cased")](https://deepset.ai/german-bert)
- [FARM](https://github.com/deepset-ai/FARM)
- [Haystack](https://github.com/deepset-ai/haystack/)
-
-Get in touch:
-[Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Website](https://deepset.ai)
-
@@ -1,108 +0,0 @@
---
-language: fr
-datasets:
- PIAF
- FQuAD
- SQuAD-FR
-widget:
- text: "Comment s'appelle le portail open data du gouvernement ?"
-  context: "Etalab est une administration publique française qui fait notamment office de Chief Data Officer de l'État et coordonne la conception et la mise en œuvre de sa stratégie dans le domaine de la donnée (ouverture et partage des données publiques ou open data, exploitation des données et intelligence artificielle...). Ainsi, Etalab développe et maintient le portail des données ouvertes du gouvernement français data.gouv.fr.
-Etalab promeut également une plus grande ouverture l'administration sur la société (gouvernement ouvert) : transparence de l'action publique, innovation ouverte, participation citoyenne... elle promeut l’innovation, l’expérimentation, les méthodes de travail ouvertes, agiles et itératives, ainsi que les synergies avec la société civile pour décloisonner l’administration et favoriser l’adoption des meilleures pratiques professionnelles dans le domaine du numérique. À ce titre elle étudie notamment l’opportunité de recourir à des technologies en voie de maturation issues du monde de la recherche.
-Cette entité chargée de l'innovation au sein de l'administration doit contribuer à l'amélioration du service public grâce au numérique. Elle est rattachée à la Direction interministérielle du numérique, dont les missions et l’organisation ont été fixées par le décret du 30 octobre 2019.  Dirigé par Laure Lucchesi depuis 2016, elle rassemble une équipe pluridisciplinaire d'une trentaine de personnes."
---
-
-# camembert-base-squadFR-fquad-piaf
-
-## Description
-
-Question-answering French model, using base [CamemBERT](https://camembert-model.fr/) fine-tuned on a combo of three French Q&A datasets:
-
-1. [PIAFv1.1](https://www.data.gouv.fr/en/datasets/piaf-le-dataset-francophone-de-questions-reponses/)
-2. [FQuADv1.0](https://fquad.illuin.tech/)
-3. [SQuAD-FR (SQuAD automatically translated to French)](https://github.com/Alikabbadj/French-SQuAD)
-
-## Training hyperparameters
-
-```shell
-python run_squad.py \
--model_type camembert \
--model_name_or_path camembert-base \
--do_train --do_eval \
--train_file data/SQuAD+fquad+piaf.json \
--predict_file data/fquad_valid.json \
--per_gpu_train_batch_size 12 \ 
--learning_rate 3e-5 \ 
--num_train_epochs 4 \  
--max_seq_length 384 \ 
--doc_stride 128 \
--save_steps 10000 
-``` 
-
-## Evaluation results
-### FQuAD v1.0 Evaluation
-```shell
-{"f1": 79.81, "exact_match": 55.14}
-```
-### SQuAD-FR Evaluation
-```shell
-{"f1": 59.54, "exact_match": 80.61}
-```
-
-## Usage
-
-```python
-from transformers import pipeline
-
-nlp = pipeline('question-answering', model='etalab-ia/camembert-base-squadFR-fquad-piaf', tokenizer='etalab-ia/camembert-base-squadFR-fquad-piaf')
-
-nlp({
-    'question': "Qui est Claude Monet?",
-    'context': "Claude Monet, né le 14 novembre 1840 à Paris et mort le 5 décembre 1926 à Giverny, est un peintre français et l’un des fondateurs de l'impressionnisme."
-})
-```
-
-## Citation
-
-### PIAF
-```
-@inproceedings{KeraronLBAMSSS20,
-  author    = {Rachel Keraron and
-               Guillaume Lancrenon and
-               Mathilde Bras and
-               Fr{\'{e}}d{\'{e}}ric Allary and
-               Gilles Moyse and
-               Thomas Scialom and
-               Edmundo{-}Pavel Soriano{-}Morales and
-               Jacopo Staiano},
-  title     = {Project {PIAF:} Building a Native French Question-Answering Dataset},
-  booktitle = {{LREC}},
-  pages     = {5481--5490},
-  publisher = {European Language Resources Association},
-  year      = {2020}
-}
-
-```
-
-### FQuAD
-```
-@article{dHoffschmidt2020FQuADFQ,
-  title={FQuAD: French Question Answering Dataset},
-  author={Martin d'Hoffschmidt and Maxime Vidal and Wacim Belblidia and Tom Brendl'e and Quentin Heinrich},
-  journal={ArXiv},
-  year={2020},
-  volume={abs/2002.06071}
-}
-```
-
-### SQuAD-FR
-```
- @MISC{kabbadj2018,
-   author =       "Kabbadj, Ali",
-   title =        "Something new in French Text Mining and Information Extraction (Universal Chatbot): Largest Q&A French training dataset (110 000+) ",
-   editor =       "linkedin.com",
-   month =        "November",
-   year =         "2018",
-   url =          "\url{https://www.linkedin.com/pulse/something-new-french-text-mining-information-chatbot-largest-kabbadj/}",
-   note =         "[Online; posted 11-November-2018]",
- }
- ```
@@ -1,33 +0,0 @@
-# Model name
-
-## Model description
-
-This model is a sequence-to-sequence question generator which takes an answer and context as an input, and generates a question as an output. It is based on a pretrained `t5-base` model.
-
-## Intended uses & limitations
-
-The model is trained to generate reading comprehension-style questions with answers extracted from a text. The model performs best with full sentence answers, but can also be used with single word or short phrase answers.
-
-#### How to use
-
-The model takes concatenated answers and context as an input sequence, and will generate a full question sentence as an output sequence. The max sequence length is 512 tokens. Inputs should be organised into the following format:
-```
-answer_token <answer-phrase> context_token <context-from-text>
-```
-The input sequence can then be encoded and passed as the `input_ids` argument in the model's `generate()` method.
-
-For best results, a large number of questions can be generated, and then filtered using [iarfmoose/bert-base-cased-qa-evaluator](https://huggingface.co/iarfmoose/bert-base-cased-qa-evaluator).
-
-For examples, please see https://github.com/iarfmoose/question_generator.
-
-#### Limitations and bias
-
-The model is limited to generating questions in the same style as those found in [SQuAD](https://rajpurkar.github.io/SQuAD-explorer/), [CoQA](https://stanfordnlp.github.io/coqa/), and [MSMARCO](https://microsoft.github.io/msmarco/). The generated questions can potentially be leading or reflect biases that are present in the context. If the context is too short or completely absent, or if the context and answer do not match, the generated question is likely to be incoherent.
-
-## Training data
-
-The model was fine-tuned on a dataset made up of several well-known QA datasets ([SQuAD](https://rajpurkar.github.io/SQuAD-explorer/), [CoQA](https://stanfordnlp.github.io/coqa/), and [MSMARCO](https://microsoft.github.io/msmarco/)). The datasets were restructured by concatenating the answer and context fields into the previously-mentioned format. The question field from the datasets was used as the target during training. The full training set was roughly 200,000 examples.
-
-## Training procedure
-
-The model was trained for 20 epochs over the training set with a learning rate of 1e-3. The batch size was only 4 due to GPU memory limitations when training on Google Colab.
@@ -1,72 +0,0 @@
---
-language: en
-tags:
- text-classification
- pytorch
-datasets:
- yahoo-answers
-widget:
- text: "Who are you voting for in 2020? <sep> This text is about politics."
---
-
-# bart-lage-mnli-yahoo-answers
-
-## Model Description
-
-This model takes [facebook/bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) and fine-tunes it on Yahoo Answers topic classification. It can be used to predict whether a topic label can be assigned to a given sequence, whether or not the label has been seen before.
-
-You can play with an interactive demo of this zero-shot technique with this model, as well as the non-finetuned [facebook/bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli), [here](https://huggingface.co/zero-shot/).
-
-## Inteded Usage
-
-This model was fine-tuned on topic classification and will perform best at zero-shot topic classification. Use `hypothesis_template="This text is about {}."` as this is the template used during fine-tuning.
-
-For settings other than topic classification, you can use any model pre-trained on MNLI such as [facebook/bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) or [roberta-large-mnli](https://huggingface.co/roberta-large-mnli) with the same code as written below.
-
-#### With the zero-shot classification pipeline
-
-The model can be used with the `zero-shot-classification` pipeline like so:
-
-```python
-from transformers import pipeline
-nlp = pipeline("zero-shot-classification", model="joeddav/bart-large-mnli-yahoo-answers")
-
-sequence_to_classify = "Who are you voting for in 2020?"
-candidate_labels = ["Europe", "public health", "politics", "elections"]
-hypothesis_template = "This text is about {}."
-nlp(sequence_to_classify, candidate_labels, multi_class=True, hypothesis_template=hypothesis_template)
-```
-
-#### With manual PyTorch
-
-```python
-# pose sequence as a NLI premise and label as a hypothesis
-from transformers import BartForSequenceClassification, BartTokenizer
-nli_model = BartForSequenceClassification.from_pretrained('joeddav/bart-large-mnli-yahoo-answers')
-tokenizer = BartTokenizer.from_pretrained('joeddav/bart-large-mnli-yahoo-answers')
-
-premise = sequence
-hypothesis = f'This text is about {label}.'
-
-# run through model pre-trained on MNLI
-x = tokenizer.encode(premise, hypothesis, return_tensors='pt',
-                        max_length=tokenizer.max_len,
-                        truncation_strategy='only_first')
-logits = nli_model(x.to(device))[0]
-
-# we throw away "neutral" (dim 1) and take the probability of
-# "entailment" (2) as the probability of the label being true 
-entail_contradiction_logits = logits[:,[0,2]]
-probs = entail_contradiction_logits.softmax(dim=1)
-prob_label_is_true = probs[:,1]
-```
-
-## Training
-
-The model is a pre-trained MNLI classifier further fine-tuned on Yahoo Answers topic classification in the manner originally described in [Yin et al. 2019](https://arxiv.org/abs/1909.00161) and [this blog post](https://joeddav.github.io/blog/2020/05/29/ZSL.html). That is, each sequence is fed to the pre-trained NLI model in place of the premise and each candidate label as the hypothesis, formatted like so: `This text is about {class name}.` For each example in the training set, a true and a randomly-selected false label hypothesis are fed to the model which must predict which labels are valid and which are false.
-
-Since this method studies the ability to classify unseen labels after being trained on a different set of labels, the model is only trained on 5 out of the 10 labels in Yahoo Answers. These are "Society & Culture", "Health", "Computers & Internet", "Business & Finance", and "Family & Relationships".
-
-## Evaluation Results
-
-This model was evaluated with the label-weighted F1 of the _seen_ and _unseen_ labels. That is, for each example the model must predict from one of the 10 corpus labels. The F1 is reported for the labels seen during training as well as the labels unseen during training. We found an F1 score of `.68` and `.72` for the unseen and seen labels, respectively. In order to adjust for the in-vs-out of distribution labels, we subtract a fixed amount of 30% from the normalized probabilities of the _seen_ labels, as described in [Yin et al. 2019](https://arxiv.org/abs/1909.00161) and [our blog post](https://joeddav.github.io/blog/2020/05/29/ZSL.html).
@@ -1,44 +0,0 @@
---
-language: en
-datasets:
- emo
---
-
-## Emo-MobileBERT: a thin version of BERT LARGE, trained on the EmoContext Dataset from scratch
-
-### Details of MobileBERT
-
-The **MobileBERT** model was presented in [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by *Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, Denny Zhou* and here is the abstract:
-
-Natural Language Processing (NLP) has recently achieved great success by using huge pre-trained models with hundreds of millions of parameters. However, these models suffer from heavy model sizes and high latency such that they cannot be deployed to resource-limited mobile devices. In this paper, we propose MobileBERT for compressing and accelerating the popular BERT model. Like the original BERT, MobileBERT is task-agnostic, that is, it can be generically applied to various downstream NLP tasks via simple fine-tuning. Basically, MobileBERT is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks. To train MobileBERT, we first train a specially designed teacher model, an inverted-bottleneck incorporated BERT_LARGE model. Then, we conduct knowledge transfer from this teacher to MobileBERT. Empirical studies show that MobileBERT is 4.3x smaller and 5.5x faster than BERT_BASE while achieving competitive results on well-known benchmarks. On the natural language inference tasks of GLUE, MobileBERT achieves a GLUEscore o 77.7 (0.6 lower than BERT_BASE), and 62 ms latency on a Pixel 4 phone. On the SQuAD v1.1/v2.0 question answering task, MobileBERT achieves a dev F1 score of 90.0/79.2 (1.5/2.1 higher than BERT_BASE).
-
-### Details of the downstream task (Emotion Recognition) - Dataset 📚
-
-SemEval-2019 Task 3: EmoContext Contextual Emotion Detection in Text
-
-In this dataset, given a textual dialogue i.e. an utterance along with two previous turns of context, the goal was to infer the underlying emotion of the utterance by choosing from four emotion classes:
-
- - sad 😢
- - happy 😃
- - angry 😡
- - others
-
-### Model training
-
-The training script is present [here](https://github.com/lordtt13/transformers-experiments/blob/master/Custom%20Tasks/emo-mobilebert.ipynb).
-
-### Pipelining the Model
-
-```python
-from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
-
-tokenizer = AutoTokenizer.from_pretrained("lordtt13/emo-mobilebert")
-
-model = AutoModelForSequenceClassification.from_pretrained("lordtt13/emo-mobilebert")
-
-nlp_sentence_classif = transformers.pipeline('sentiment-analysis', model = model, tokenizer = tokenizer)
-nlp_sentence_classif("I've never had such a bad day in my life")
-# Output: [{'label': 'sad', 'score': 0.93153977394104}]
-```
-
-> Created by [Tanmay Thakur](https://github.com/lordtt13) | [LinkedIn](https://www.linkedin.com/in/tanmay-thakur-6bb5a9154/)
@@ -1,161 +0,0 @@
---
-language: fa
-tags:
- albert-persian
- persian-lm
-license: apache-2.0
-datasets:
- Persian Wikidumps
- MirasText
- BigBang Page
- Chetor
- Eligasht
- DigiMag
- Ted Talks
- Books (Novels, ...)
---
-
-# ALBERT-Persian
-
-## ALBERT-Persian: A Lite BERT for Self-supervised Learning of Language Representations for the Persian Language
-
-## Introduction
-
-ALBERT-Persian trained on a massive amount of public corpora ([Persian Wikidumps](https://dumps.wikimedia.org/fawiki/), [MirasText](https://github.com/miras-tech/MirasText)) and six other manually crawled text data from a various type of websites ([BigBang Page](https://bigbangpage.com/) `scientific`, [Chetor](https://www.chetor.com/) `lifestyle`, [Eligasht](https://www.eligasht.com/Blog/) `itinerary`,  [Digikala](https://www.digikala.com/mag/) `digital magazine`, [Ted Talks](https://www.ted.com/talks) `general conversational`, Books `novels, storybooks, short stories from old to the contemporary era`).
-
-
-
-## Intended uses & limitations
-
-You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
-be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?search=albert-fa) to look for
-fine-tuned versions on a task that interests you.
-
-
-### How to use
-
-#### TensorFlow 2.0
-
-```python
-from transformers import AutoConfig, AutoTokenizer, TFAutoModel
-
-config = AutoConfig.from_pretrained("m3hrdadfi/albert-fa-base-v2")
-tokenizer = AutoTokenizer.from_pretrained("m3hrdadfi/albert-fa-base-v2")
-model = TFAutoModel.from_pretrained("m3hrdadfi/albert-fa-base-v2")
-
-text = "ما در هوشواره معتقدیم با انتقال صحیح دانش و آگاهی، همه افراد می‌توانند از ابزارهای هوشمند استفاده کنند. شعار ما هوش مصنوعی برای همه است."
-tokenizer.tokenize(text)
-
->>> ['▁ما', '▁در', '▁هوش', 'واره', '▁معتقد', 'یم', '▁با', '▁انتقال', '▁صحیح', '▁دانش', '▁و', '▁اگاه', 'ی', '،', '▁همه', '▁افراد', '▁می', '▁توانند', '▁از', '▁ابزارهای', '▁هوشمند', '▁استفاده', '▁کنند', '.', '▁شعار', '▁ما', '▁هوش', '▁مصنوعی', '▁برای', '▁همه', '▁است', '.']
-
-```
-
-#### Pytorch
-
-```python
-from transformers import AutoConfig, AutoTokenizer, AutoModel
-
-config = AutoConfig.from_pretrained("m3hrdadfi/albert-fa-base-v2")
-tokenizer = AutoTokenizer.from_pretrained("m3hrdadfi/albert-fa-base-v2")
-model = AutoModel.from_pretrained("m3hrdadfi/albert-fa-base-v2")
-```
-
-## Training
-
-ALBERT-Persian is the first attempt on ALBERT for the Persian Language. The model was trained based on Google's ALBERT BASE Version 2.0 over various writing styles from numerous subjects (e.g., scientific, novels, news) with more than `3.9M` documents, `73M` sentences, and `1.3B` words, like the way we did for [ParsBERT](https://github.com/hooshvare/parsbert).
-
-## Goals
-Objective goals during training are as below (after 140K steps).
-
-``` bash
-***** Eval results *****
-global_step = 140000
-loss = 2.0080082
-masked_lm_accuracy = 0.6141017
-masked_lm_loss = 1.9963315
-sentence_order_accuracy = 0.985
-sentence_order_loss = 0.06908702
-```
-
-
-## Derivative models
-
-### Base Config
-
-#### Albert Model
- [m3hrdadfi/albert-face-base-v2](https://huggingface.co/m3hrdadfi/albert-fa-base-v2) 
-
-#### Albert Sentiment Analysis
- [m3hrdadfi/albert-fa-base-v2-sentiment-digikala](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-digikala) 
- [m3hrdadfi/albert-fa-base-v2-sentiment-snappfood](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-snappfood) 
- [m3hrdadfi/albert-fa-base-v2-sentiment-deepsentipers-binary](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-deepsentipers-binary) 
- [m3hrdadfi/albert-fa-base-v2-sentiment-deepsentipers-multi](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-deepsentipers-multi) 
- [m3hrdadfi/albert-fa-base-v2-sentiment-binary](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-binary) 
- [m3hrdadfi/albert-fa-base-v2-sentiment-multi](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-multi) 
- [m3hrdadfi/albert-fa-base-v2-sentiment-multi](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-sentiment-multi) 
-
-#### Albert Text Classification
- [m3hrdadfi/albert-fa-base-v2-clf-digimag](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-clf-digimag) 
- [m3hrdadfi/albert-fa-base-v2-clf-persiannews](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-clf-persiannews) 
-
-#### Albert NER
- [m3hrdadfi/albert-fa-base-v2-ner](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-ner) 
- [m3hrdadfi/albert-fa-base-v2-ner-arman](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-ner-arman) 
- [m3hrdadfi/albert-fa-base-v2-ner-arman](https://huggingface.co/m3hrdadfi/albert-fa-base-v2-ner-arman) 
-
-## Eval results
-
-The following tables summarize the F1 scores obtained by ALBERT-Persian as compared to other models and architectures.
-
-
-### Sentiment Analysis (SA) Task
-
-|          Dataset         | ALBERT-fa-base-v2 | ParsBERT-v1 | mBERT | DeepSentiPers |
-|:------------------------:|:-----------------:|:-----------:|:-----:|:-------------:|
-|  Digikala User Comments  |       81.12       |    81.74    | 80.74 |       -       |
-|  SnappFood User Comments |       85.79       |    88.12    | 87.87 |       -       |
-|  SentiPers (Multi Class) |       66.12       |    71.11    |   -   |     69.33     |
-| SentiPers (Binary Class) |       91.09       |    92.13    |   -   |     91.98     |
-
-
-### Text Classification (TC) Task
-
-|      Dataset      | ALBERT-fa-base-v2 | ParsBERT-v1 | mBERT |
-|:-----------------:|:-----------------:|:-----------:|:-----:|
-| Digikala Magazine |       92.33       |    93.59    | 90.72 |
-|    Persian News   |       97.01       |    97.19    | 95.79 |
-
-
-### Named Entity Recognition (NER) Task
-
-| Dataset | ALBERT-fa-base-v2 | ParsBERT-v1 | mBERT | MorphoBERT | Beheshti-NER | LSTM-CRF | Rule-Based CRF | BiLSTM-CRF |
-|:-------:|:-----------------:|:-----------:|:-----:|:----------:|:------------:|:--------:|:--------------:|:----------:|
-|  PEYMA  |       88.99       |    93.10    | 86.64 |      -     |     90.59    |     -    |      84.00     |      -     |
-|  ARMAN  |       97.43       |    98.79    | 95.89 |    89.9    |     84.03    |   86.55  |        -       |    77.45   |
-
-
-### BibTeX entry and citation info
-
-Please cite in publications as the following:
-
-```bibtex
-@misc{ALBERT-Persian,
-  author = {Mehrdad Farahani},
-  title = {ALBERT-Persian: A Lite BERT for Self-supervised Learning of Language Representations for the Persian Language},
-  year = {2020},
-  publisher = {GitHub},
-  journal = {GitHub repository},
-  howpublished = {\url{https://github.com/m3hrdadfi/albert-persian}},
-}
-
-@article{ParsBERT,
-    title={ParsBERT: Transformer-based Model for Persian Language Understanding},
-    author={Mehrdad Farahani, Mohammad Gharachorloo, Marzieh Farahani, Mohammad Manthouri},
-    journal={ArXiv},
-    year={2020},
-    volume={abs/2005.12515}
-}
-```
-
-## Questions?
-Post a Github issue on the [ALBERT-Persian](https://github.com/m3hrdadfi/albert-persian) repo.
@@ -1,125 +1,5 @@
 ---
 language: es
-thumbnail: https://i.imgur.com/DUlT077.jpg
 widget:
 - text: "España es un país muy <mask> en la UE"
 ---
-
-# RuPERTa: the Spanish RoBERTa 🎃<img src="https://abs-0.twimg.com/emoji/v2/svg/1f1ea-1f1f8.svg" alt="spain flag" width="25"/>
-
-RuPERTa-base (uncased) is a [RoBERTa model](https://github.com/pytorch/fairseq/tree/master/examples/roberta) trained on a *uncased* verison of [big Spanish corpus](https://github.com/josecannete/spanish-corpora).
-RoBERTa iterates on BERT's pretraining procedure, including training the model longer, with bigger batches over more data; removing the next sentence prediction objective; training on longer sequences; and dynamically changing the masking pattern applied to the training data.
-The architecture is the same as `roberta-base`:
-
-`roberta.base:` **RoBERTa** using the **BERT-base architecture 125M** params
-
-## Benchmarks 🧾 
-WIP (I continue working on it) 🚧
-
-| Task     |    F1 | Precision | Recall |                                                                        Fine-tuned model |                                                                                                                                                                                                                                                                                               Reproduce it |
-| -------- | ----: | --------: | -----: | --------------------------------------------------------------------------------------: | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
-| POS      | 97.39 |     97.47 |  97.32 | [RuPERTa-base-finetuned-pos](https://huggingface.co/mrm8488/RuPERTa-base-finetuned-pos) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mrm8488/shared_colab_notebooks/blob/master/RuPERTa_base_finetuned_POS.ipynb)
-| NER      | 77.55 |     75.53 |  79.68 | [RuPERTa-base-finetuned-ner](https://huggingface.co/mrm8488/RuPERTa-base-finetuned-ner) |
-| SQUAD-es v1 |  to-do |       |    |
-| SQUAD-es v2 |  to-do |       |    |
-
-## Model in action 🔨
-
-### Usage for POS and NER 🏷
-
-```python
-import torch
-from transformers import AutoModelForTokenClassification, AutoTokenizer
-
-id2label = {
-    "0": "B-LOC",
-    "1": "B-MISC",
-    "2": "B-ORG",
-    "3": "B-PER",
-    "4": "I-LOC",
-    "5": "I-MISC",
-    "6": "I-ORG",
-    "7": "I-PER",
-    "8": "O"
-}
-
-tokenizer = AutoTokenizer.from_pretrained('mrm8488/RuPERTa-base-finetuned-ner')
-model = AutoModelForTokenClassification.from_pretrained('mrm8488/RuPERTa-base-finetuned-ner')
-
-text ="Julien, CEO de HF, nació en Francia."
-
-input_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)
-
-outputs = model(input_ids)
-last_hidden_states = outputs[0]
-
-for m in last_hidden_states:
-  for index, n in enumerate(m):
-    if(index > 0 and index <= len(text.split(" "))):
-      print(text.split(" ")[index-1] + ": " + id2label[str(torch.argmax(n).item())])
-
-# Output:
-'''
-Julien,: I-PER
-CEO: O
-de: O
-HF,: B-ORG
-nació: I-PER
-en: I-PER
-Francia.: I-LOC
-'''
-```
-
-For **POS** just change the `id2label` dictionary and the model path to [mrm8488/RuPERTa-base-finetuned-pos](https://huggingface.co/mrm8488/RuPERTa-base-finetuned-pos)
-
-### Fast usage for LM with `pipelines` 🧪
-
-```python
-from transformers import AutoModelWithLMHead, AutoTokenizer
-model = AutoModelWithLMHead.from_pretrained('mrm8488/RuPERTa-base')
-tokenizer = AutoTokenizer.from_pretrained("mrm8488/RuPERTa-base", do_lower_case=True)
-
-from transformers import pipeline
-
-pipeline_fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer)
-
-pipeline_fill_mask("España es un país muy <mask> en la UE")
-```
-
-```json
-[
-  {
-    "score": 0.1814306527376175,
-    "sequence": "<s> españa es un país muy importante en la ue</s>",
-    "token": 1560
-  },
-  {
-    "score": 0.024842597544193268,
-    "sequence": "<s> españa es un país muy fuerte en la ue</s>",
-    "token": 2854
-  },
-  {
-    "score": 0.02473250962793827,
-    "sequence": "<s> españa es un país muy pequeño en la ue</s>",
-    "token": 2948
-  },
-  {
-    "score": 0.023991240188479424,
-    "sequence": "<s> españa es un país muy antiguo en la ue</s>",
-    "token": 5240
-  },
-  {
-    "score": 0.0215945765376091,
-    "sequence": "<s> españa es un país muy popular en la ue</s>",
-    "token": 5782
-  }
-]
-```
-
-## Acknowledgments
-
-I thank [🤗/transformers team](https://github.com/huggingface/transformers) for answering my doubts and Google for helping me with the [TensorFlow Research Cloud](https://www.tensorflow.org/tfrc) program.
-
-> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488)
-
-> Made with <span style="color: #e25555;">&hearts;</span> in Spain
@@ -1,6 +0,0 @@
---
-language: en
-thumbnail:
-widget:
- text: "HuggingFace Cake:"
---
@@ -1,6 +0,0 @@
---
-language: en
-thumbnail:
-widget:
- text: "HuggingFace Cake:"
---
@@ -1,74 +0,0 @@
---
-language: en
-datasets:
- squad
---
-
-# MobileBERT + SQuAD (v1.1) 📱❓
-
-[mobilebert-uncased](https://huggingface.co/google/mobilebert-uncased) fine-tuned on [SQUAD v2.0 dataset](https://rajpurkar.github.io/SQuAD-explorer/explore/v2.0/dev/) for **Q&A** downstream task.
-
-## Details of the downstream task (Q&A) - Model 🧠
-
-**MobileBERT** is a thin version of *BERT_LARGE*, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks.
-
-The checkpoint used here is the original MobileBert Optimized Uncased English: (uncased_L-24_H-128_B-512_A-4_F-4_OPT) checkpoint.
-
-More about the model [here](https://arxiv.org/abs/2004.02984)
-
-## Details of the downstream task (Q&A) - Dataset 📚
-
-**S**tanford **Q**uestion **A**nswering **D**ataset (SQuAD) is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage, or the question might be unanswerable.
-SQuAD v1.1 contains **100,000+** question-answer pairs on **500+** articles.
-
-## Model training 🏋️‍
-
-The model was trained on a Tesla P100 GPU and 25GB of RAM with the following command:
-
-```bash
-python transformers/examples/question-answering/run_squad.py \
-  --model_type bert \
-  --model_name_or_path 'google/mobilebert-uncased' \
-  --do_eval \
-  --do_train \
-  --do_lower_case \
-  --train_file '/content/dataset/train-v1.1.json' \
-  --predict_file '/content/dataset/dev-v1.1.json' \
-  --per_gpu_train_batch_size 16 \
-  --learning_rate 3e-5 \
-  --num_train_epochs 5 \
-  --max_seq_length 384 \
-  --doc_stride 128 \
-  --output_dir '/content/output' \
-  --overwrite_output_dir \
-  --save_steps 1000
-```
-
-It is importatnt to say that this models converges much faster than other ones. So, it is also cheap to fine-tune.
-
-## Test set Results 🧾
-
-| Metric | # Value   |
-| ------ | --------- |
-| **EM** | **82.33** |
-| **F1** | **89.64** |
-| **Size**| **94 MB** |
-
-### Model in action 🚀
-
-Fast usage with **pipelines**:
-
-```python
-from transformers import pipeline
-QnA_pipeline = pipeline('question-answering', model='mrm8488/mobilebert-uncased-finetuned-squadv1')
-QnA_pipeline({
-    'context': 'A new strain of flu that has the potential to become a pandemic has been identified in China by scientists.',
-    'question': 'Who did identified it ?'
-    })
-    
-# Output: {'answer': 'scientists.', 'end': 106, 'score': 0.7885545492172241, 'start': 96}
-```
-
-> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
-
-> Made with <span style="color: #e25555;">&hearts;</span> in Spain
@@ -1,74 +0,0 @@
---
-language: en
-datasets:
- squad_v2
---
-
-# MobileBERT + SQuAD v2 📱❓
-
-[mobilebert-uncased](https://huggingface.co/google/mobilebert-uncased) fine-tuned on [SQUAD v2.0 dataset](https://rajpurkar.github.io/SQuAD-explorer/explore/v2.0/dev/) for **Q&A** downstream task.
-
-## Details of the downstream task (Q&A) - Model 🧠
-
-**MobileBERT** is a thin version of *BERT_LARGE*, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks.
-
-The checkpoint used here is the original MobileBert Optimized Uncased English: (uncased_L-24_H-128_B-512_A-4_F-4_OPT) checkpoint.
-
-More about the model [here](https://arxiv.org/abs/2004.02984)
-
-## Details of the downstream task (Q&A) - Dataset 📚
-
-**SQuAD2.0** combines the 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions written adversarially by crowdworkers to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but also determine when no answer is supported by the paragraph and abstain from answering.
-
-## Model training 🏋️‍
-
-The model was trained on a Tesla P100 GPU and 25GB of RAM with the following command:
-
-```bash
-python transformers/examples/question-answering/run_squad.py \
-  --model_type bert \
-  --model_name_or_path 'google/mobilebert-uncased' \
-  --do_eval \
-  --do_train \
-  --do_lower_case \
-  --train_file '/content/dataset/train-v2.0.json' \
-  --predict_file '/content/dataset/dev-v2.0.json' \
-  --per_gpu_train_batch_size 16 \
-  --learning_rate 3e-5 \
-  --num_train_epochs 5 \
-  --max_seq_length 384 \
-  --doc_stride 128 \
-  --output_dir '/content/output' \
-  --overwrite_output_dir \
-  --save_steps 1000 \
-  --version_2_with_negative
-```
-
-It is importatnt to say that this models converges much faster than other ones. So, it is also cheap to fine-tune.
-
-## Test set Results 🧾
-
-| Metric | # Value   |
-| ------ | --------- |
-| **EM** | **75.37** |
-| **F1** | **78.48** |
-| **Size**| **94 MB** |
-
-### Model in action 🚀
-
-Fast usage with **pipelines**:
-
-```python
-from transformers import pipeline
-QnA_pipeline = pipeline('question-answering', model='mrm8488/mobilebert-uncased-finetuned-squadv2')
-QnA_pipeline({
-    'context': 'A new strain of flu that has the potential to become a pandemic has been identified in China by scientists.',
-    'question': 'Who did identified it ?'
-    })
-    
-# Output: {'answer': 'scientists.', 'end': 106, 'score': 0.41531604528427124, 'start': 96}
-```
-
-> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
-
-> Made with <span style="color: #e25555;">&hearts;</span> in Spain
@@ -1,84 +0,0 @@
-## Model in Action 🚀
-
-```python
-import torch
-from transformers import T5ForConditionalGeneration,T5Tokenizer
-
-
-def set_seed(seed):
-  torch.manual_seed(seed)
-  if torch.cuda.is_available():
-    torch.cuda.manual_seed_all(seed)
-
-set_seed(42)
-
-model = T5ForConditionalGeneration.from_pretrained('ramsrigouthamg/t5_paraphraser')
-tokenizer = T5Tokenizer.from_pretrained('ramsrigouthamg/t5_paraphraser')
-
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print ("device ",device)
-model = model.to(device)
-
-sentence = "Which course should I take to get started in data science?"
-# sentence = "What are the ingredients required to bake a perfect cake?"
-# sentence = "What is the best possible approach to learn aeronautical engineering?"
-# sentence = "Do apples taste better than oranges in general?"
-
-
-text =  "paraphrase: " + sentence + " </s>"
-
-
-max_len = 256
-
-encoding = tokenizer.encode_plus(text,pad_to_max_length=True, return_tensors="pt")
-input_ids, attention_masks = encoding["input_ids"].to(device), encoding["attention_mask"].to(device)
-
-
-# set top_k = 50 and set top_p = 0.95 and num_return_sequences = 3
-beam_outputs = model.generate(
-    input_ids=input_ids, attention_mask=attention_masks,
-    do_sample=True,
-    max_length=256,
-    top_k=120,
-    top_p=0.98,
-    early_stopping=True,
-    num_return_sequences=10
-)
-
-
-print ("\nOriginal Question ::")
-print (sentence)
-print ("\n")
-print ("Paraphrased Questions :: ")
-final_outputs =[]
-for beam_output in beam_outputs:
-    sent = tokenizer.decode(beam_output, skip_special_tokens=True,clean_up_tokenization_spaces=True)
-    if sent.lower() != sentence.lower() and sent not in final_outputs:
-        final_outputs.append(sent)
-
-for i, final_output in enumerate(final_outputs):
-    print("{}: {}".format(i, final_output))
-
-```
-## Output
-```
-Original Question ::
-Which course should I take to get started in data science?
-
-
-Paraphrased Questions :: 
-0: What should I learn to become a data scientist?
-1: How do I get started with data science?
-2: How would you start a data science career?
-3: How can I start learning data science?
-4: How do you get started in data science?
-5: What's the best course for data science?
-6: Which course should I start with for data science?
-7: What courses should I follow to get started in data science?
-8: What degree should be taken by a data scientist?
-9: Which course should I follow to become a Data Scientist?
-```
-
-## Detailed blog post available here :
-https://towardsdatascience.com/paraphrase-any-question-with-t5-text-to-text-transfer-transformer-pretrained-model-and-cbb9e35f1555
-
@@ -1,26 +0,0 @@
---
-language: pt
-tags:
- portuguese
- brazil
- pt_BR
-widget:
- text: gostei muito dessa <mask>
---
-
-# BR_BERTo
-
-Portuguese (Brazil) model for text inference.
-
-## Params
-
-Trained on a corpus of 5_258_624 sentences, with 132_807_374 non unique tokens (992_418 unique tokens).
-
- Vocab size: 220_000
- RobertaForMaskedLM  size : 32
- Num train epochs: 2
- Time to train: ~23hs (on GCP with a Nvidia T4)
-
-I follow the great tutorial from HuggingFace team:
-
-[How to train a new language model from scratch using Transformers and Tokenizers](https://huggingface.co/blog/how-to-train)
@@ -1,7 +1,4 @@
 ---
-language: en
-datasets:
- c4
 tags:
 - summarization
 - translation
@@ -9,20 +6,3 @@ tags:
 license: apache-2.0
 ---

-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) 
-
-Pretraining Dataset: [C4](https://huggingface.co/datasets/c4)
-
-Other Community Checkpoints: [here](https://huggingface.co/models?search=t5)
-
-Paper: [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)
-
-Authors: *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* 
-
-
-## Abstract
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
@@ -1,7 +1,4 @@
 ---
-language: en
-datasets:
- c4
 tags:
 - summarization
 - translation
@@ -9,20 +6,3 @@ tags:
 license: apache-2.0
 ---

-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) 
-
-Pretraining Dataset: [C4](https://huggingface.co/datasets/c4)
-
-Other Community Checkpoints: [here](https://huggingface.co/models?search=t5)
-
-Paper: [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)
-
-Authors: *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* 
-
-
-## Abstract
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
@@ -1,7 +1,4 @@
 ---
-language: en
-datasets:
- c4
 tags:
 - summarization
 - translation
@@ -9,20 +6,3 @@ tags:
 license: apache-2.0
 ---

-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) 
-
-Pretraining Dataset: [C4](https://huggingface.co/datasets/c4)
-
-Other Community Checkpoints: [here](https://huggingface.co/models?search=t5)
-
-Paper: [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)
-
-Authors: *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* 
-
-
-## Abstract
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
@@ -1,7 +1,4 @@
 ---
-language: en
-datasets:
- c4
 tags:
 - summarization
 - translation
@@ -9,20 +6,3 @@ tags:
 license: apache-2.0
 ---

-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) 
-
-Pretraining Dataset: [C4](https://huggingface.co/datasets/c4)
-
-Other Community Checkpoints: [here](https://huggingface.co/models?search=t5)
-
-Paper: [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)
-
-Authors: *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* 
-
-
-## Abstract
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
@@ -1,7 +1,4 @@
 ---
-language: en
-datasets:
- c4
 tags:
 - summarization
 - translation
@@ -9,20 +6,3 @@ tags:
 license: apache-2.0
 ---

-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) 
-
-Pretraining Dataset: [C4](https://huggingface.co/datasets/c4)
-
-Other Community Checkpoints: [here](https://huggingface.co/models?search=t5)
-
-Paper: [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)
-
-Authors: *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* 
-
-
-## Abstract
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
@@ -1,44 +0,0 @@
---
-language: it
-tags:
- sentiment
- Italian
-license: MIT
-widget:
- text: 'Giuseppe Rossi è un ottimo politico'
---
-
-# 🤗 + polibert_SA - POLItic BERT based Sentiment Analysis
-  
-## Model description  
-  
-This model performs sentiment analysis on Italian political twitter sentences. It was trained starting from an instance of "bert-base-italian-uncased-xxl" and fine-tuned on an Italian dataset of tweets.
-  
-#### Hands-on  
-  
-```python
-import torch
-from torch import nn 
-
-text = "Giueseppe Rossi è un pessimo politico"
-input_ids = tokenizer.encode(text, add_special_tokens=True, return_tensors= 'pt')
-
-logits, = model(input_ids)
-logits = logits.squeeze(0)
-
-prob = nn.functional.softmax(logits, dim=0)
-
-# 0 Negative, 1 Neutral, 2 Positive 
-print(prob.argmax().tolist())
-```  
-  
-#### Hyperparameters
-
- Optimizer: **AdamW** with learning rate of **2e-5**, epsilon of **1e-8**
- Max epochs: **2**
- Batch size: **16**
-
-## Acknowledgments
-
-Thanks to the support from: 
-the [Hugging Face](https://huggingface.co/), Unione Professionisti (https://www.unioneprofessionisti.com/)
@@ -41,4 +41,3 @@ Pull Request so it can be included under the Community notebooks.
 |[Fine-tune T5 for Summarization](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|How to fine-tune T5 for summarization in PyTorch and track experiments with WandB|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|
 |[Speed up Fine-Tuning in Transformers with Dynamic Padding / Bucketing](https://github.com/ELS-RD/transformers-notebook/blob/master/Divide_Hugging_Face_Transformers_training_time_by_2_or_more.ipynb)|How to speed up fine-tuning by a factor of 2 using dynamic padding / bucketing|[Michael Benesty](https://github.com/pommedeterresautee) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CBfRU1zbfu7-ijiOqAAQUA-RJaxfcJoO?usp=sharing)|
 |[Pretrain Reformer for Masked Language Modeling](https://github.com/patrickvonplaten/notebooks/blob/master/Reformer_For_Masked_LM.ipynb)| How to train a Reformer model with bi-directional self-attention layers | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tzzh0i8PgDQGV3SMFUGxM7_gGae3K-uW?usp=sharing)|
-|[Expand and Fine Tune Sci-BERT](https://github.com/lordtt13/word-embeddings/blob/master/COVID-19%20Research%20Data/COVID-SciBERT.ipynb)| How to increase vocabulary of a pretrained SciBERT model from AllenAI on the CORD dataset and pipeline it. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1rqAR40goxbAfez1xvF3hBJphSCsvXmh8)|
@@ -9,11 +9,10 @@ known_third_party =
    fairseq
    faiss
    fastprogress
-    fire
-    fugashi
    git
    h5py
    matplotlib
+    MeCab
    nlp
    nltk
    numpy
@@ -44,5 +43,5 @@ multi_line_output = 3
 use_parentheses = True

 [flake8]
-ignore = E203, E501, E741, W503, W605
+ignore = E203, E501, E741, W503
 max-line-length = 119
@@ -65,7 +65,7 @@ if stale_egg_info.exists():

 extras = {}

-extras["ja"] = ["fugashi>=1.0", "ipadic>=1.0,<2.0"]
+extras["mecab"] = ["mecab-python3<1"]
 extras["sklearn"] = ["scikit-learn"]

 # keras2onnx and onnxconverter-common version is specific through a commit until 1.7.0 lands on pypi
@@ -97,7 +97,7 @@ extras["quality"] = [
    "isort @ git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort",
    "flake8",
 ]
-extras["dev"] = extras["testing"] + extras["quality"] + extras["ja"] + ["scikit-learn", "tensorflow", "torch"]
+extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3<1", "scikit-learn", "tensorflow", "torch"]

 setup(
    name="transformers",
@@ -104,8 +104,6 @@ from .modeling_tf_pytorch_utils import (

 # Pipelines
 from .pipelines import (
-    Conversation,
-    ConversationalPipeline,
    CsvPipelineDataFormat,
    FeatureExtractionPipeline,
    FillMaskPipeline,
@@ -120,7 +118,6 @@ from .pipelines import (
    TextGenerationPipeline,
    TokenClassificationPipeline,
    TranslationPipeline,
-    ZeroShotClassificationPipeline,
    pipeline,
 )

@@ -281,7 +278,6 @@ if is_torch_available():
        XLMForTokenClassification,
        XLMForQuestionAnswering,
        XLMForQuestionAnsweringSimple,
-        XLMForMultipleChoice,
        XLM_PRETRAINED_MODEL_ARCHIVE_LIST,
    )
    from .modeling_bart import (
@@ -360,8 +356,6 @@ if is_torch_available():
        FlaubertForTokenClassification,
        FlaubertForQuestionAnswering,
        FlaubertForQuestionAnsweringSimple,
-        FlaubertForTokenClassification,
-        FlaubertForMultipleChoice,
        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
    )

@@ -396,7 +390,6 @@ if is_torch_available():
        LongformerForMultipleChoice,
        LongformerForTokenClassification,
        LongformerForQuestionAnswering,
-        LongformerSelfAttention,
        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
    )

@@ -538,10 +531,8 @@ if is_tf_available():
    from .modeling_tf_electra import (
        TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
        TFElectraForMaskedLM,
-        TFElectraForMultipleChoice,
        TFElectraForPreTraining,
        TFElectraForQuestionAnswering,
-        TFElectraForSequenceClassification,
        TFElectraForTokenClassification,
        TFElectraModel,
        TFElectraPreTrainedModel,
@@ -566,14 +557,6 @@ if is_tf_available():
        TFGPT2PreTrainedModel,
    )

-    from .modeling_tf_longformer import (
-        TF_LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        TFLongformerModel,
-        TFLongformerForMaskedLM,
-        TFLongformerForQuestionAnswering,
-        TFLongformerSelfAttention,
-    )
-
    from .modeling_tf_mobilebert import (
        TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
        TFMobileBertModel,
@@ -62,8 +62,8 @@ class BenchmarkArguments:
    fp16: bool = field(default=False, metadata={"help": "Use FP16 to accelerate inference."})
    training: bool = field(default=False, metadata={"help": "Benchmark training of model"})
    verbose: bool = field(default=False, metadata={"help": "Verbose memory tracing"})
-    no_speed: bool = field(default=False, metadata={"help": "Don't perform speed measurements"})
-    no_memory: bool = field(default=False, metadata={"help": "Don't perform memory measurements"})
+    no_speed: bool = field(default=False, metadata={"help": "Don't perform speed measurments"})
+    no_memory: bool = field(default=False, metadata={"help": "Don't perform memory measurments"})
    trace_memory_line_by_line: bool = field(default=False, metadata={"help": "Trace memory line by line"})
    save_to_csv: bool = field(default=False, metadata={"help": "Save result to a CSV file"})
    log_print: bool = field(default=False, metadata={"help": "Save all print statements in a log file"})
@@ -165,7 +165,7 @@ class MemorySummary(NamedTuple):
        - `sequential`: a list of `MemoryState` namedtuple (see below) computed from the provided `memory_trace`
            by substracting the memory after executing each line from the memory before executing said line.
        - `cumulative`: a list of `MemoryState` namedtuple (see below) with cumulative increase in memory for each line
-            obtained by summing repeated memory increase for a line if it's executed several times.
+            obtained by summing repeted memory increase for a line if it's executed several times.
            The list is sorted from the frame with the largest memory consumption to the frame with the smallest (can be negative if memory is released)
        - `total`: total memory increase during the full tracing as a `Memory` named tuple (see below).
            Line with memory release (negative consumption) are ignored if `ignore_released_memory` is `True` (default).
@@ -310,7 +310,7 @@ def start_memory_tracing(
    gpus_to_trace: Optional[List[int]] = None,
 ) -> MemoryTrace:
    """ Setup line-by-line tracing to record rss mem (RAM) at each line of a module or sub-module.
-        See `./benchmark.py` for usage examples.
+        See `../../examples/benchmarks.py for a usage example.
        Current memory consumption is returned using psutil and in particular is the RSS memory
            "Resident Set Size” (the non-swapped physical memory the process is using).
            See https://psutil.readthedocs.io/en/latest/#psutil.Process.memory_info
@@ -468,7 +468,7 @@ def stop_memory_tracing(
                - `sequential`: a list of `MemoryState` namedtuple (see below) computed from the provided `memory_trace`
                    by substracting the memory after executing each line from the memory before executing said line.
                - `cumulative`: a list of `MemoryState` namedtuple (see below) with cumulative increase in memory for each line
-                    obtained by summing repeated memory increase for a line if it's executed several times.
+                    obtained by summing repeted memory increase for a line if it's executed several times.
                    The list is sorted from the frame with the largest memory consumption to the frame with the smallest (can be negative if memory is released)
                - `total`: total memory increase during the full tracing as a `Memory` named tuple (see below).
                    Line with memory release (negative consumption) are ignored if `ignore_released_memory` is `True` (default).
@@ -81,6 +81,7 @@ class TrainCommand(BaseTransformersCLICommand):
        self.framework = "tf" if is_tf_available() else "torch"

        os.makedirs(args.output, exist_ok=True)
+        assert os.path.isdir(args.output)
        self.output = args.output

        self.column_label = args.column_label
@@ -56,15 +56,6 @@ class EncoderDecoderConfig(PretrainedConfig):
            >>> # Accessing the model configuration
            >>> config_encoder = model.config.encoder
            >>> config_decoder  = model.config.decoder
-            >>> # set decoder config to causal lm
-            >>> config_decoder.is_decoder = True
-
-            >>> # Saving the model, including its configuration
-            >>> model.save_pretrained('my-model')
-
-            >>> # loading model and config from pretrained folder
-            >>> encoder_decoder_config = EncoderDecoderConfig.from_pretrained('my-model')
-            >>> model = EncoderDecoderModel.from_pretrained('my-model', config=encoder_decoder_config)
    """
    model_type = "encoder_decoder"

@@ -78,7 +69,7 @@ class EncoderDecoderConfig(PretrainedConfig):
        decoder_config = kwargs.pop("decoder")
        decoder_model_type = decoder_config.pop("model_type")

-        from .configuration_auto import AutoConfig
+        from transformers import AutoConfig

        self.encoder = AutoConfig.for_model(encoder_model_type, **encoder_config)
        self.decoder = AutoConfig.for_model(decoder_model_type, **decoder_config)
Author	SHA1	Message	Date
Sam Shleifer	203aa19ea8	Skip on CPU	2020-07-22 12:19:28 -04:00
Sam Shleifer	dd3336f18e	slow slow slow	2020-07-22 12:00:12 -04:00
Sam Shleifer	5930c8ec7c	add tiny test_data	2020-07-22 11:35:08 -04:00