Release: v2.9.1

* add first text for generation

* add generation pipeline to usage

* Created using Colaboratory

* correct docstring

* finish

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.9.0'
+release = u'2.9.1'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/model_doc/marian.rst

@@ -1,28 +1,90 @@
-MarianMTModel
+MarianMT
 ----------------------------------------------------
 **DISCLAIMER:** If you see something strange,
 file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer
-These models are for machine translation. The list of supported language pairs can be found `here <https://huggingface.co/Helsinki-NLP>`__.
-
-Opus Project
-~~~~~~~~~~~~
-The 1,000+ models were originally trained by `Jörg Tiedemann <https://researchportal.helsinki.fi/en/persons/j%C3%B6rg-tiedemann>`__ using the `Marian <https://marian-nmt.github.io/>`_ C++ library, which supports fast training and translation.
-All models are transformer encoder-decoders with 6 layers in each component. Each model's performance is documented in a model card.
+@sshleifer. Translations should be similar, but not identical to, output in the test set linked to in each model card.
 
 Implementation Notes
 ~~~~~~~~~~~~~~~~~~~~
 - each model is about 298 MB on disk, there are 1,000+ models.
-- Models are named with the following patter 'Helsinki-NLP/opus-mt-{src_langs}-{targ_langs}'. If there are multiple source or target languages they are joined by a '+' symbol.
+- The list of supported language pairs can be found `here <https://huggingface.co/Helsinki-NLP>`__.
+- The 1,000+ models were originally trained by `Jörg Tiedemann <https://researchportal.helsinki.fi/en/persons/j%C3%B6rg-tiedemann>`__ using the `Marian <https://marian-nmt.github.io/>`_ C++ library, which supports fast training and translation.
+- All models are transformer encoder-decoders with 6 layers in each component. Each model's performance is documented in a model card.
 - the 80 opus models that require BPE preprocessing are not supported.
-- There is an outstanding issue w.r.t multilingual models and language codes.
-- The modeling code is the same as ``BartModel`` with a few minor modifications:
+- The modeling code is the same as ``BartForConditionalGeneration`` with a few minor modifications:
     - static (sinusoid) positional embeddings (``MarianConfig.static_position_embeddings=True``)
     - a new final_logits_bias (``MarianConfig.add_bias_logits=True``)
     - no layernorm_embedding (``MarianConfig.normalize_embedding=False``)
     - the model starts generating with pad_token_id (which has 0 token_embedding) as the prefix. (Bart uses <s/>)
 - Code to bulk convert models can be found in ``convert_marian_to_pytorch.py``
 
+Naming
+~~~~~~
+- All  model names use the following format: ``Helsinki-NLP/opus-mt-{src}-{tgt}``
+- The language codes used to name models are inconsistent. Two digit codes can usually be found `here <https://developers.google.com/admin-sdk/directory/v1/languages>`_, three digit codes require googling "language code {code}".
+- Codes formatted like ``es_AR`` are usually ``code_{region}``. That one is spanish documents from Argentina.
+
+
+Multilingual Models
+~~~~~~~~~~~~~~~~~~~~
+
+All  model names use the following format: ``Helsinki-NLP/opus-mt-{src}-{tgt}``:
+    - if ``src`` is in all caps, the model supports multiple input languages, you can figure out which ones by looking at the model card, or the Group Members `mapping <https://gist.github.com/sshleifer/6d20e7761931b08e73c3219027b97b8a>`_ .
+    - if ``tgt`` is in all caps, the model can output multiple languages, and you should specify a language code by prepending the desired output language to the src_text
+    - You can see a tokenizer's supported language codes in ``tokenizer.supported_language_codes``
+
+Example of translating english to many romance languages, using language codes:
+
+.. code-block:: python
+
+    from transformers import MarianMTModel, MarianTokenizer
+    src_text = [
+        '>>fr<< this is a sentence in english that we want to translate to french',
+        '>>pt<< This should go to portuguese',
+        '>>es<< And this to Spanish'
+    ]
+
+    model_name = 'Helsinki-NLP/opus-mt-en-ROMANCE'
+    tokenizer = MarianTokenizer.from_pretrained(model_name)
+    print(tokenizer.supported_language_codes)
+    model = MarianMTModel.from_pretrained(model_name)
+    translated = model.generate(**tokenizer.prepare_translation_batch(src_text))
+    tgt_text = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
+    # ["c'est une phrase en anglais que nous voulons traduire en français",
+    # 'Isto deve ir para o português.',
+    # 'Y esto al español']
+
+Sometimes, models were trained on collections of languages that do not resolve to a group. In this case, _ is used as a separator for src or tgt, as in ``'Helsinki-NLP/opus-mt-en_el_es_fi-en_el_es_fi'``. These still require language codes.
+There are many supported regional language codes, like ``>>es_ES<<`` (Spain) and ``>>es_AR<<`` (Argentina), that do not seem to change translations. I have not found these to provide different results than just using ``>>es<<``.
+
+For Example:
+    - ``Helsinki-NLP/opus-mt-NORTH_EU-NORTH_EU``: translates from all NORTH_EU languages (see `mapping <https://gist.github.com/sshleifer/6d20e7761931b08e73c3219027b97b8a>`_) to all NORTH_EU languages. Use a special language code like ``>>de<<`` to specify output language.
+    - ``Helsinki-NLP/opus-mt-ROMANCE-en``: translates from many romance languages to english, no codes needed since there is only 1 tgt language.
+
+
+
+.. code-block:: python
+
+    GROUP_MEMBERS = {
+     'ZH': ['cmn', 'cn', 'yue', 'ze_zh', 'zh_cn', 'zh_CN', 'zh_HK', 'zh_tw', 'zh_TW', 'zh_yue', 'zhs', 'zht', 'zh'],
+     'ROMANCE': ['fr', 'fr_BE', 'fr_CA', 'fr_FR', 'wa', 'frp', 'oc', 'ca', 'rm', 'lld', 'fur', 'lij', 'lmo', 'es', 'es_AR', 'es_CL', 'es_CO', 'es_CR', 'es_DO', 'es_EC', 'es_ES', 'es_GT', 'es_HN', 'es_MX', 'es_NI', 'es_PA', 'es_PE', 'es_PR', 'es_SV', 'es_UY', 'es_VE', 'pt', 'pt_br', 'pt_BR', 'pt_PT', 'gl', 'lad', 'an', 'mwl', 'it', 'it_IT', 'co', 'nap', 'scn', 'vec', 'sc', 'ro', 'la'],
+     'NORTH_EU': ['de', 'nl', 'fy', 'af', 'da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],
+     'SCANDINAVIA': ['da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],
+     'SAMI': ['se', 'sma', 'smj', 'smn', 'sms'],
+     'NORWAY': ['nb_NO', 'nb', 'nn_NO', 'nn', 'nog', 'no_nb', 'no'],
+     'CELTIC': ['ga', 'cy', 'br', 'gd', 'kw', 'gv']
+    }
+
+Code to see available pretrained models:
+
+.. code-block:: python
+
+    from transformers.hf_api import HfApi
+    model_list = HfApi().model_list()
+    org = "Helsinki-NLP"
+    model_ids = [x.modelId for x in model_list if x.modelId.startswith(org)]
+    suffix = [x.split('/')[1] for x in model_ids]
+    multi_models = [f'{org}/{s}' for s in suffix if s != s.lower()]
 
 MarianMTModel
 ~~~~~~~~~~~~~

docs/source/usage.rst

@@ -404,48 +404,150 @@ Causal language modeling is the task of predicting the token following a sequenc
 model only attends to the left context (tokens on the left of the mask). Such a training is particularly interesting
 for generation tasks.
 
-There is currently no pipeline to do causal language modeling/generation.
+Usually, the next token is predicted by sampling from the logits of the last hidden state the model produces from the input sequence.
 
-Here is an example using the tokenizer and model. leveraging the :func:`~transformers.PreTrainedModel.generate` method
-to generate the tokens following the initial sequence in PyTorch, and creating a simple loop in TensorFlow.
+Here is an example using the tokenizer and model and leveraging the :func:`~transformers.PreTrainedModel.top_k_top_p_filtering` method to sample the next token following an input sequence of tokens.
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer, top_k_top_p_filtering
+    import torch
+    from torch.nn import functional as F
+
+
+    tokenizer = AutoTokenizer.from_pretrained("gpt2")
+    model = AutoModelWithLMHead.from_pretrained("gpt2")
+
+    sequence = f"Hugging Face is based in DUMBO, New York City, and "
+
+    input_ids = tokenizer.encode(sequence, return_tensors="pt")
+
+    # get logits of last hidden state
+    next_token_logits = model(input_ids)[0][:, -1, :]
+
+    # filter
+    filtered_next_token_logits = top_k_top_p_filtering(next_token_logits, top_k=50, top_p=1.0)
+
+    # sample
+    probs = F.softmax(filtered_next_token_logits, dim=-1)
+    next_token = torch.multinomial(probs, num_samples=1)
+
+    generated = torch.cat([input_ids, next_token], dim=-1)
+
+    resulting_string = tokenizer.decode(generated.tolist()[0])
+    print(resulting_string)
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer, tf_top_k_top_p_filtering
+    import tensorflow as tf
+
+    tokenizer = AutoTokenizer.from_pretrained("gpt2")
+    model = TFAutoModelWithLMHead.from_pretrained("gpt2")
+
+    sequence = f"Hugging Face is based in DUMBO, New York City, and "
+
+    input_ids = tokenizer.encode(sequence, return_tensors="tf")
+
+    # get logits of last hidden state
+    next_token_logits = model(input_ids)[0][:, -1, :]
+
+    # filter
+    filtered_next_token_logits = tf_top_k_top_p_filtering(next_token_logits, top_k=50, top_p=1.0)
+
+    # sample
+    next_token = tf.random.categorical(filtered_next_token_logits, dtype=tf.int32, num_samples=1)
+
+    generated = tf.concat([input_ids, next_token], axis=1)
+
+    resulting_string = tokenizer.decode(generated.numpy().tolist()[0])
+    print(resulting_string)
+
+
+This outputs a (hopefully) coherent next token following the original sequence, which is in our case is the word *has*:
+
+::
+
+    Hugging Face is based in DUMBO, New York City, and has
+
+In the next section, we show how this functionality is leveraged in :func:`~transformers.PreTrainedModel.generate` to generate multiple tokens up to a user-defined length.
+
+Text Generation
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+In text generation (*a.k.a* *open-ended text generation*) the goal is to create a coherent portion of text that is a continuation from the given context. As an example, is it shown how *GPT-2* can be used in pipelines to generate text. As a default all models apply *Top-K* sampling when used in pipelines as configured in their respective configurations (see `gpt-2 config <https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.json>`_ for example).
+
+::
+
+    from transformers import pipeline
+
+    text_generator = pipeline("text-generation")
+    print(text_generator("As far as I am concerned, I will", max_length=50))
+
+
+Here the model generates a random text with a total maximal length of *50* tokens from context *"As far as I am concerned, I will"*.
+The default arguments of ``PreTrainedModel.generate()`` can directly be overriden in the pipeline as is shown above for the argument ``max_length``.
+
+Here is an example for text generation using XLNet and its tokenzier. 
 
 ::
 
     ## PYTORCH CODE
     from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    model = AutoModelWithLMHead.from_pretrained("gpt2")
+    model = AutoModelWithLMHead.from_pretrained("xlnet-base-cased")
+    tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
 
-    sequence = f"Hugging Face is based in DUMBO, New York City, and is"
+    # Padding text helps XLNet with short prompts - proposed by Aman Rusia in https://github.com/rusiaaman/XLNet-gen#methodology
+    PADDING_TEXT = """In 1991, the remains of Russian Tsar Nicholas II and his family
+    (except for Alexei and Maria) are discovered.
+    The voice of Nicholas's young son, Tsarevich Alexei Nikolaevich, narrates the
+    remainder of the story. 1883 Western Siberia,
+    a young Grigori Rasputin is asked by his father and a group of men to perform magic.
+    Rasputin has a vision and denounces one of the men as a horse thief. Although his
+    father initially slaps him for making such an accusation, Rasputin watches as the
+    man is chased outside and beaten. Twenty years later, Rasputin sees a vision of
+    the Virgin Mary, prompting him to become a priest. Rasputin quickly becomes famous,
+    with people, even a bishop, begging for his blessing. <eod> </s> <eos>""" 
 
-    input = tokenizer.encode(sequence, return_tensors="pt")
-    generated = model.generate(input, max_length=50, do_sample=True)
+    prompt = "Today the weather is really nice and I am planning on "
+    inputs = tokenizer.encode(PADDING_TEXT + prompt, add_special_tokens=False, return_tensors="pt")
+    
+    prompt_length = len(tokenizer.decode(inputs[0], skip_special_tokens=True, clean_up_tokenization_spaces=True))
+    outputs = model.generate(inputs, max_length=250, do_sample=True, top_p=0.95, top_k=60)
+    generated = prompt + tokenizer.decode(outputs[0])[prompt_length:]
 
-    resulting_string = tokenizer.decode(generated.tolist()[0])
-    print(resulting_string)
+    print(generated)
     ## TENSORFLOW CODE
     from transformers import TFAutoModelWithLMHead, AutoTokenizer
-    import tensorflow as tf
 
-    tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    model = TFAutoModelWithLMHead.from_pretrained("gpt2")
+    model = TFAutoModelWithLMHead.from_pretrained("xlnet-base-cased")
+    tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
 
-    sequence = f"Hugging Face is based in DUMBO, New York City, and is"
-    input = tokenizer.encode(sequence, return_tensors="tf")
-    generated = model.generate(input, max_length=50, do_sample=True)
+    # Padding text helps XLNet with short prompts - proposed by Aman Rusia in https://github.com/rusiaaman/XLNet-gen#methodology
+    PADDING_TEXT = """In 1991, the remains of Russian Tsar Nicholas II and his family
+    (except for Alexei and Maria) are discovered.
+    The voice of Nicholas's young son, Tsarevich Alexei Nikolaevich, narrates the
+    remainder of the story. 1883 Western Siberia,
+    a young Grigori Rasputin is asked by his father and a group of men to perform magic.
+    Rasputin has a vision and denounces one of the men as a horse thief. Although his
+    father initially slaps him for making such an accusation, Rasputin watches as the
+    man is chased outside and beaten. Twenty years later, Rasputin sees a vision of
+    the Virgin Mary, prompting him to become a priest. Rasputin quickly becomes famous,
+    with people, even a bishop, begging for his blessing. <eod> </s> <eos>""" 
 
-    resulting_string = tokenizer.decode(generated.tolist()[0])
-    print(resulting_string)
+    prompt = "Today the weather is really nice and I am planning on "
+    inputs = tokenizer.encode(PADDING_TEXT + prompt, add_special_tokens=False, return_tensors="tf")
 
+    prompt_length = len(tokenizer.decode(inputs[0], skip_special_tokens=True, clean_up_tokenization_spaces=True))
+    outputs = model.generate(inputs, max_length=250, do_sample=True, top_p=0.95, top_k=60)
+    generated = prompt + tokenizer.decode(outputs[0])[prompt_length:]
 
-This outputs a (hopefully) coherent string from the original sequence, as the
-:func:`~transformers.PreTrainedModel.generate` samples from a top_p/tok_k distribution:
+    print(generated)
 
-::
+Text generation is currently possible with *GPT-2*, *OpenAi-GPT*, *CTRL*, *XLNet*, *Transfo-XL* and *Reformer* in PyTorch and for most models in Tensorflow as well. As can be seen in the example above *XLNet* and *Transfo-xl* often need to be padded to work well.
+GPT-2 is usually a good choice for *open-ended text generation* because it was trained on millions on webpages with a causal language modeling objective.
 
-    Hugging Face is based in DUMBO, New York City, and is a live-action TV series based on the novel by John
-    Carpenter, and its producers, David Kustlin and Steve Pichar. The film is directed by!
+For more information on how to apply different decoding strategies for text generation, please also refer to our generation blog post `here <https://huggingface.co/blog/how-to-generate>`_.
 
 
 Named Entity Recognition

examples/README.md

@@ -14,12 +14,13 @@ This is still a work-in-progress – in particular documentation is still sparse
 
 ## Tasks built on Trainer
 
-| Task | Example datasets | Trainer support | TFTrainer support | pytorch-lightning | Colab | One-click Deploy to Azure (wip) | 
+| Task | Example datasets | Trainer support | TFTrainer support | pytorch-lightning | Colab | One-click Deploy to Azure (wip) |
 |---|---|:---:|:---:|:---:|:---:|:---:|
 | [`language-modeling`](./language-modeling) | Raw text | ✅ | - | - | - | - |
 | [`text-classification`](./text-classification) | GLUE, XNLI | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/trainer/01_text_classification.ipynb) | [![Deploy to Azure](https://aka.ms/deploytoazurebutton)](https://portal.azure.com/#create/Microsoft.Template/uri/https%3A%2F%2Fraw.githubusercontent.com%2FAzure%2Fazure-quickstart-templates%2Fmaster%2F101-storage-account-create%2Fazuredeploy.json) |
 | [`token-classification`](./token-classification) | CoNLL NER | ✅ | ✅ | ✅ | - | - |
-| [`multiple-choice`](./multiple-choice) | SWAG, RACE, ARC | ✅ | - | - | - | - |
+| [`multiple-choice`](./multiple-choice) | SWAG, RACE, ARC | ✅ | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb) | - |
+| [`question-answering`](./question-answering) | SQuAD | - | ✅ | - | - | - |
 
 
 

examples/language-modeling/run_language_modeling.py

@@ -265,7 +265,7 @@ def main():
 
         eval_output = trainer.evaluate()
 
-        perplexity = math.exp(eval_output["loss"])
+        perplexity = math.exp(eval_output["eval_loss"])
         result = {"perplexity": perplexity}
 
         output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")

examples/multiple-choice/README.md

@@ -29,3 +29,28 @@ Training with the defined hyper-parameters yields the following results:
 eval_acc = 0.8338998300509847
 eval_loss = 0.44457291918821606
 ```
+
+
+## Tensorflow
+
+```bash
+export SWAG_DIR=/path/to/swag_data_dir
+python ./examples/multiple-choice/run_tf_multiple_choice.py \
+--task_name swag \
+--model_name_or_path bert-base-cased \
+--do_train \
+--do_eval \
+--data_dir $SWAG_DIR \
+--learning_rate 5e-5 \
+--num_train_epochs 3 \
+--max_seq_length 80 \
+--output_dir models_bert/swag_base \
+--per_gpu_eval_batch_size=16 \
+--per_gpu_train_batch_size=16 \
+--logging-dir logs \
+--gradient_accumulation_steps 2 \
+--overwrite_output
+```
+
+# Run it in colab
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)

examples/multiple-choice/run_tf_multiple_choice.py

@@ -0,0 +1,211 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Finetuning the library models for multiple choice (Bert, Roberta, XLNet)."""
+
+
+import logging
+import os
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+import numpy as np
+
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    EvalPrediction,
+    HfArgumentParser,
+    TFAutoModelForMultipleChoice,
+    TFTrainer,
+    TFTrainingArguments,
+    set_seed,
+)
+from utils_multiple_choice import Split, TFMultipleChoiceDataset, processors
+
+
+logger = logging.getLogger(__name__)
+
+
+def simple_accuracy(preds, labels):
+    return (preds == labels).mean()
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    task_name: str = field(metadata={"help": "The name of the task to train on: " + ", ".join(processors.keys())})
+    data_dir: str = field(metadata={"help": "Should contain the data files for the task."})
+    max_seq_length: int = field(
+        default=128,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.warning(
+        "device: %s, n_gpu: %s, 16-bits training: %s", training_args.device, training_args.n_gpu, training_args.fp16,
+    )
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed
+    set_seed(training_args.seed)
+
+    try:
+        processor = processors[data_args.task_name]()
+        label_list = processor.get_labels()
+        num_labels = len(label_list)
+    except KeyError:
+        raise ValueError("Task not found: %s" % (data_args.task_name))
+
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        num_labels=num_labels,
+        finetuning_task=data_args.task_name,
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    with training_args.strategy.scope():
+        model = TFAutoModelForMultipleChoice.from_pretrained(
+            model_args.model_name_or_path,
+            from_pt=bool(".bin" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+    # Get datasets
+    train_dataset = (
+        TFMultipleChoiceDataset(
+            data_dir=data_args.data_dir,
+            tokenizer=tokenizer,
+            task=data_args.task_name,
+            max_seq_length=data_args.max_seq_length,
+            overwrite_cache=data_args.overwrite_cache,
+            mode=Split.train,
+        )
+        if training_args.do_train
+        else None
+    )
+    eval_dataset = (
+        TFMultipleChoiceDataset(
+            data_dir=data_args.data_dir,
+            tokenizer=tokenizer,
+            task=data_args.task_name,
+            max_seq_length=data_args.max_seq_length,
+            overwrite_cache=data_args.overwrite_cache,
+            mode=Split.dev,
+        )
+        if training_args.do_eval
+        else None
+    )
+
+    def compute_metrics(p: EvalPrediction) -> Dict:
+        preds = np.argmax(p.predictions, axis=1)
+        return {"acc": simple_accuracy(preds, p.label_ids)}
+
+    # Initialize our Trainer
+    trainer = TFTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset.get_dataset() if train_dataset else None,
+        eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
+        compute_metrics=compute_metrics,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train()
+        trainer.save_model()
+        tokenizer.save_pretrained(training_args.output_dir)
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        result = trainer.evaluate()
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results *****")
+            for key, value in result.items():
+                logger.info("  %s = %s", key, value)
+                writer.write("%s = %s\n" % (key, value))
+
+            results.update(result)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/multiple-choice/utils_multiple_choice.py

@@ -25,11 +25,9 @@ from dataclasses import dataclass
 from enum import Enum
 from typing import List, Optional
 
-import torch
 import tqdm
-from torch.utils.data.dataset import Dataset
 
-from transformers import PreTrainedTokenizer, torch_distributed_zero_first
+from transformers import PreTrainedTokenizer, is_tf_available, is_torch_available
 
 
 logger = logging.getLogger(__name__)
@@ -76,66 +74,160 @@ class Split(Enum):
     test = "test"
 
 
-class MultipleChoiceDataset(Dataset):
-    """
-    This will be superseded by a framework-agnostic approach
-    soon.
-    """
+if is_torch_available():
+    import torch
+    from torch.utils.data.dataset import Dataset
+    from transformers import torch_distributed_zero_first
 
-    features: List[InputFeatures]
+    class MultipleChoiceDataset(Dataset):
+        """
+        This will be superseded by a framework-agnostic approach
+        soon.
+        """
 
-    def __init__(
-        self,
-        data_dir: str,
-        tokenizer: PreTrainedTokenizer,
-        task: str,
-        max_seq_length: Optional[int] = None,
-        overwrite_cache=False,
-        mode: Split = Split.train,
-        local_rank=-1,
-    ):
-        processor = processors[task]()
+        features: List[InputFeatures]
 
-        cached_features_file = os.path.join(
-            data_dir,
-            "cached_{}_{}_{}_{}".format(mode.value, tokenizer.__class__.__name__, str(max_seq_length), task,),
-        )
-        with torch_distributed_zero_first(local_rank):
-            # Make sure only the first process in distributed training processes the dataset,
-            # and the others will use the cache.
+        def __init__(
+            self,
+            data_dir: str,
+            tokenizer: PreTrainedTokenizer,
+            task: str,
+            max_seq_length: Optional[int] = None,
+            overwrite_cache=False,
+            mode: Split = Split.train,
+            local_rank=-1,
+        ):
+            processor = processors[task]()
 
-            if os.path.exists(cached_features_file) and not overwrite_cache:
-                logger.info(f"Loading features from cached file {cached_features_file}")
-                self.features = torch.load(cached_features_file)
-            else:
-                logger.info(f"Creating features from dataset file at {data_dir}")
-                label_list = processor.get_labels()
-                if mode == Split.dev:
-                    examples = processor.get_dev_examples(data_dir)
-                elif mode == Split.test:
-                    examples = processor.get_test_examples(data_dir)
+            cached_features_file = os.path.join(
+                data_dir,
+                "cached_{}_{}_{}_{}".format(mode.value, tokenizer.__class__.__name__, str(max_seq_length), task,),
+            )
+            with torch_distributed_zero_first(local_rank):
+                # Make sure only the first process in distributed training processes the dataset,
+                # and the others will use the cache.
+
+                if os.path.exists(cached_features_file) and not overwrite_cache:
+                    logger.info(f"Loading features from cached file {cached_features_file}")
+                    self.features = torch.load(cached_features_file)
                 else:
-                    examples = processor.get_train_examples(data_dir)
-                logger.info("Training examples: %s", len(examples))
-                # TODO clean up all this to leverage built-in features of tokenizers
-                self.features = convert_examples_to_features(
-                    examples,
-                    label_list,
-                    max_seq_length,
-                    tokenizer,
-                    pad_on_left=bool(tokenizer.padding_side == "left"),
-                    pad_token=tokenizer.pad_token_id,
-                    pad_token_segment_id=tokenizer.pad_token_type_id,
-                )
-                if local_rank in [-1, 0]:
-                    logger.info("Saving features into cached file %s", cached_features_file)
-                    torch.save(self.features, cached_features_file)
+                    logger.info(f"Creating features from dataset file at {data_dir}")
+                    label_list = processor.get_labels()
+                    if mode == Split.dev:
+                        examples = processor.get_dev_examples(data_dir)
+                    elif mode == Split.test:
+                        examples = processor.get_test_examples(data_dir)
+                    else:
+                        examples = processor.get_train_examples(data_dir)
+                    logger.info("Training examples: %s", len(examples))
+                    # TODO clean up all this to leverage built-in features of tokenizers
+                    self.features = convert_examples_to_features(
+                        examples,
+                        label_list,
+                        max_seq_length,
+                        tokenizer,
+                        pad_on_left=bool(tokenizer.padding_side == "left"),
+                        pad_token=tokenizer.pad_token_id,
+                        pad_token_segment_id=tokenizer.pad_token_type_id,
+                    )
+                    if local_rank in [-1, 0]:
+                        logger.info("Saving features into cached file %s", cached_features_file)
+                        torch.save(self.features, cached_features_file)
 
-    def __len__(self):
-        return len(self.features)
+        def __len__(self):
+            return len(self.features)
 
-    def __getitem__(self, i) -> InputFeatures:
-        return self.features[i]
+        def __getitem__(self, i) -> InputFeatures:
+            return self.features[i]
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+    class TFMultipleChoiceDataset:
+        """
+        This will be superseded by a framework-agnostic approach
+        soon.
+        """
+
+        features: List[InputFeatures]
+
+        def __init__(
+            self,
+            data_dir: str,
+            tokenizer: PreTrainedTokenizer,
+            task: str,
+            max_seq_length: Optional[int] = 128,
+            overwrite_cache=False,
+            mode: Split = Split.train,
+        ):
+            processor = processors[task]()
+
+            logger.info(f"Creating features from dataset file at {data_dir}")
+            label_list = processor.get_labels()
+            if mode == Split.dev:
+                examples = processor.get_dev_examples(data_dir)
+            elif mode == Split.test:
+                examples = processor.get_test_examples(data_dir)
+            else:
+                examples = processor.get_train_examples(data_dir)
+            logger.info("Training examples: %s", len(examples))
+            # TODO clean up all this to leverage built-in features of tokenizers
+            self.features = convert_examples_to_features(
+                examples,
+                label_list,
+                max_seq_length,
+                tokenizer,
+                pad_on_left=bool(tokenizer.padding_side == "left"),
+                pad_token=tokenizer.pad_token_id,
+                pad_token_segment_id=tokenizer.pad_token_type_id,
+            )
+
+            def gen():
+                for (ex_index, ex) in tqdm.tqdm(enumerate(self.features), desc="convert examples to features"):
+                    if ex_index % 10000 == 0:
+                        logger.info("Writing example %d of %d" % (ex_index, len(examples)))
+
+                    yield (
+                        {
+                            "example_id": 0,
+                            "input_ids": ex.input_ids,
+                            "attention_mask": ex.attention_mask,
+                            "token_type_ids": ex.token_type_ids,
+                        },
+                        ex.label,
+                    )
+
+            self.dataset = tf.data.Dataset.from_generator(
+                gen,
+                (
+                    {
+                        "example_id": tf.int32,
+                        "input_ids": tf.int32,
+                        "attention_mask": tf.int32,
+                        "token_type_ids": tf.int32,
+                    },
+                    tf.int64,
+                ),
+                (
+                    {
+                        "example_id": tf.TensorShape([]),
+                        "input_ids": tf.TensorShape([None, None]),
+                        "attention_mask": tf.TensorShape([None, None]),
+                        "token_type_ids": tf.TensorShape([None, None]),
+                    },
+                    tf.TensorShape([]),
+                ),
+            )
+
+        def get_dataset(self):
+            return self.dataset
+
+        def __len__(self):
+            return len(self.features)
+
+        def __getitem__(self, i) -> InputFeatures:
+            return self.features[i]
 
 
 class DataProcessor:
@@ -225,6 +317,52 @@ class RaceProcessor(DataProcessor):
         return examples
 
 
+class SynonymProcessor(DataProcessor):
+    """Processor for the Synonym data set."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} train".format(data_dir))
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "mctrain.csv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "mchp.csv")), "dev")
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "mctest.csv")), "test")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1", "2", "3", "4"]
+
+    def _read_csv(self, input_file):
+        with open(input_file, "r", encoding="utf-8") as f:
+            return list(csv.reader(f))
+
+    def _create_examples(self, lines: List[List[str]], type: str):
+        """Creates examples for the training and dev sets."""
+
+        examples = [
+            InputExample(
+                example_id=line[0],
+                question="",  # in the swag dataset, the
+                # common beginning of each
+                # choice is stored in "sent2".
+                contexts=[line[1], line[1], line[1], line[1], line[1]],
+                endings=[line[2], line[3], line[4], line[5], line[6]],
+                label=line[7],
+            )
+            for line in lines  # we skip the line with the column names
+        ]
+
+        return examples
+
+
 class SwagProcessor(DataProcessor):
     """Processor for the SWAG data set."""
 
@@ -435,7 +573,5 @@ def convert_examples_to_features(
     return features
 
 
-processors = {"race": RaceProcessor, "swag": SwagProcessor, "arc": ArcProcessor}
-
-
-MULTIPLE_CHOICE_TASKS_NUM_LABELS = {"race", 4, "swag", 4, "arc", 4}
+processors = {"race": RaceProcessor, "swag": SwagProcessor, "arc": ArcProcessor, "syn": SynonymProcessor}
+MULTIPLE_CHOICE_TASKS_NUM_LABELS = {"race", 4, "swag", 4, "arc", 4, "syn", 5}

examples/question-answering/README.md

@@ -157,3 +157,23 @@ Larger batch size may improve the performance while costing more memory.
 }
 ```
 
+## SQuAD with the Tensorflow Trainer
+
+```bash
+python run_tf_squad.py \
+    --model_name_or_path bert-base-uncased \
+    --output_dir model \
+    --max-seq-length 384 \
+    --num_train_epochs 2 \
+    --per_gpu_train_batch_size 8 \
+    --per_gpu_eval_batch_size 16 \
+    --do_train \
+    --logging_dir logs \
+    --mode question-answering \
+    --logging_steps 10 \
+    --learning_rate 3e-5 \
+    --doc_stride 128 \
+    --optimizer_name adamw
+```
+
+For the moment the evaluation is not available in the Tensorflow Trainer only the training.

examples/question-answering/run_tf_squad.py

@@ -0,0 +1,237 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Fine-tuning the library models for question-answering."""
+
+
+import logging
+import os
+from dataclasses import dataclass, field
+from typing import Optional
+
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    HfArgumentParser,
+    TFAutoModelForQuestionAnswering,
+    TFTrainer,
+    TFTrainingArguments,
+    squad_convert_examples_to_features,
+)
+from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    use_fast: bool = field(default=False, metadata={"help": "Set this flag to use fast tokenization."})
+    # If you want to tweak more attributes on your tokenizer, you should do it in a distinct script,
+    # or just modify its tokenizer_config.json.
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    data_dir: Optional[str] = field(
+        default=None, metadata={"help": "The input data dir. Should contain the .json files for the SQuAD task."}
+    )
+    max_seq_length: int = field(
+        default=128,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    doc_stride: int = field(
+        default=128,
+        metadata={"help": "When splitting up a long document into chunks, how much stride to take between chunks."},
+    )
+    max_query_length: int = field(
+        default=64,
+        metadata={
+            "help": "The maximum number of tokens for the question. Questions longer than this will "
+            "be truncated to this length."
+        },
+    )
+    max_answer_length: int = field(
+        default=30,
+        metadata={
+            "help": "The maximum length of an answer that can be generated. This is needed because the start "
+            "and end predictions are not conditioned on one another."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    version_2_with_negative: bool = field(
+        default=False, metadata={"help": "If true, the SQuAD examples contain some that do not have an answer."}
+    )
+    null_score_diff_threshold: float = field(
+        default=0.0, metadata={"help": "If null_score - best_non_null is greater than the threshold predict null."}
+    )
+    n_best_size: int = field(
+        default=20, metadata={"help": "If null_score - best_non_null is greater than the threshold predict null."}
+    )
+    lang_id: int = field(
+        default=0,
+        metadata={
+            "help": "language id of input for language-specific xlm models (see tokenization_xlm.PRETRAINED_INIT_CONFIGURATION)"
+        },
+    )
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(
+        "n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        training_args.n_gpu,
+        bool(training_args.n_gpu > 1),
+        training_args.fp16,
+    )
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Prepare Question-Answering task
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast,
+    )
+
+    with training_args.strategy.scope():
+        model = TFAutoModelForQuestionAnswering.from_pretrained(
+            model_args.model_name_or_path,
+            from_pt=bool(".bin" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+
+    # Get datasets
+    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")
+
+        try:
+            import tensorflow_datasets as tfds
+        except ImportError:
+            raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
+
+        tfds_examples = tfds.load("squad")
+        train_examples = (
+            SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=False)
+            if training_args.do_train
+            else None
+        )
+        eval_examples = (
+            SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=True)
+            if training_args.do_eval
+            else None
+        )
+    else:
+        processor = SquadV2Processor() if data_args.version_2_with_negative else SquadV1Processor()
+        train_examples = processor.get_train_examples(data_args.data_dir) if training_args.do_train else None
+        eval_examples = processor.get_dev_examples(data_args.data_dir) if training_args.do_eval else None
+
+    train_dataset = (
+        squad_convert_examples_to_features(
+            examples=train_examples,
+            tokenizer=tokenizer,
+            max_seq_length=data_args.max_seq_length,
+            doc_stride=data_args.doc_stride,
+            max_query_length=data_args.max_query_length,
+            is_training=True,
+            return_dataset="tf",
+        )
+        if training_args.do_train
+        else None
+    )
+
+    eval_dataset = (
+        squad_convert_examples_to_features(
+            examples=eval_examples,
+            tokenizer=tokenizer,
+            max_seq_length=data_args.max_seq_length,
+            doc_stride=data_args.doc_stride,
+            max_query_length=data_args.max_query_length,
+            is_training=False,
+            return_dataset="tf",
+        )
+        if training_args.do_eval
+        else None
+    )
+
+    # Initialize our Trainer
+    trainer = TFTrainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)
+
+    # Training
+    if training_args.do_train:
+        trainer.train()
+        trainer.save_model()
+        tokenizer.save_pretrained(training_args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

examples/test_examples.py

@@ -72,7 +72,7 @@ class ExamplesTests(unittest.TestCase):
             """.split()
         with patch.object(sys, "argv", testargs):
             result = run_glue.main()
-            del result["loss"]
+            del result["eval_loss"]
             for value in result.values():
                 self.assertGreaterEqual(value, 0.75)
 

examples/token-classification/test_ner_examples.py

@@ -6,7 +6,7 @@ from unittest.mock import patch
 import run_ner
 
 
-logging.basicConfig(level=logging.DEBUG)
+logging.basicConfig(level=logging.INFO)
 
 logger = logging.getLogger()
 
@@ -30,4 +30,4 @@ class ExamplesTests(unittest.TestCase):
             """.split()
         with patch.object(sys, "argv", ["run.py"] + testargs):
             result = run_ner.main()
-            self.assertLess(result["loss"], 1.5)
+            self.assertLess(result["eval_loss"], 1.5)

model_cards/ViktorAlm/electra-base-norwegian-uncased-discriminator/README.md

@@ -0,0 +1,19 @@
+# Norwegian Electra
+Image incoming, im going to have som fun with this one.
+
+Trained on Oscar + wikipedia + opensubtitles + some other data I had with the awesome power of TPUs(V3-8)
+
+Use with caution. I have no downstream tasks in Norwegian to test on so I have no idea of its performance yet.
+
+# Acknowledgments
+
+### TensorFlow Research Cloud
+Research supported with Cloud TPUs from Google's TensorFlow Research Cloud (TFRC). Thanks for providing access to the TFRC ❤️
+- https://www.tensorflow.org/tfrc
+
+#### OSCAR corpus
+- https://oscar-corpus.com/
+
+#### OPUS
+- http://opus.nlpl.eu/
+- http://www.opensubtitles.org/

model_cards/dbmdz/electra-base-turkish-cased-discriminator/README.md

@@ -0,0 +1,79 @@
+---
+language: turkish
+license: mit
+---
+
+# 🤗 + 📚 dbmdz Turkish ELECTRA model
+
+In this repository the MDZ Digital Library team (dbmdz) at the Bavarian State
+Library open sources a cased ELECTRA base model for Turkish 🎉
+
+# Turkish ELECTRA model
+
+We release a base ELEC**TR**A model for Turkish, that was trained on the same data as *BERTurk*.
+
+> ELECTRA is a new method for self-supervised language representation learning. It can be used to
+> pre-train transformer networks using relatively little compute. ELECTRA models are trained to
+> distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to
+> the discriminator of a GAN.
+
+More details about ELECTRA can be found in the [ICLR paper](https://openreview.net/forum?id=r1xMH1BtvB)
+or in the [official ELECTRA repository](https://github.com/google-research/electra) on GitHub.
+
+## Stats
+
+The current version of the model is trained on a filtered and sentence
+segmented version of the Turkish [OSCAR corpus](https://traces1.inria.fr/oscar/),
+a recent Wikipedia dump, various [OPUS corpora](http://opus.nlpl.eu/) and a
+special corpus provided by [Kemal Oflazer](http://www.andrew.cmu.edu/user/ko/).
+
+The final training corpus has a size of 35GB and 44,04,976,662 tokens.
+
+Thanks to Google's TensorFlow Research Cloud (TFRC) we could train a cased model
+on a TPU v3-8 for 1M steps.
+
+## Model weights
+
+[Transformers](https://github.com/huggingface/transformers)
+compatible weights for both PyTorch and TensorFlow are available.
+
+| Model                                            | Downloads
+| ------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------
+| `dbmdz/electra-base-turkish-cased-discriminator` | [`config.json`](https://cdn.huggingface.co/dbmdz/electra-base-turkish-cased-discriminator/config.json) • [`pytorch_model.bin`](https://cdn.huggingface.co/dbmdz/electra-base-turkish-cased-discriminator/pytorch_model.bin) • [`vocab.txt`](https://cdn.huggingface.co/dbmdz/electra-base-turkish-cased-discriminator/vocab.txt)
+
+## Usage
+
+With Transformers >= 2.8 our ELECTRA base cased model can be loaded like:
+
+```python
+from transformers import AutoModelWithLMHead, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("dbmdz/electra-base-turkish-cased-discriminator")
+model = AutoModelWithLMHead.from_pretrained("dbmdz/electra-base-turkish-cased-discriminator")
+```
+
+## Results
+
+For results on PoS tagging or NER tasks, please refer to
+[this repository](https://github.com/stefan-it/turkish-bert/electra).
+
+# Huggingface model hub
+
+All models are available on the [Huggingface model hub](https://huggingface.co/dbmdz).
+
+# Contact (Bugs, Feedback, Contribution and more)
+
+For questions about our ELECTRA models just open an issue
+[here](https://github.com/dbmdz/berts/issues/new) 🤗
+
+# Acknowledgments
+
+Thanks to [Kemal Oflazer](http://www.andrew.cmu.edu/user/ko/) for providing us
+additional large corpora for Turkish. Many thanks to Reyyan Yeniterzi for providing
+us the Turkish NER dataset for evaluation.
+
+Research supported with Cloud TPUs from Google's TensorFlow Research Cloud (TFRC).
+Thanks for providing access to the TFRC ❤️
+
+Thanks to the generous support from the [Hugging Face](https://huggingface.co/) team,
+it is possible to download both cased and uncased models from their S3 storage 🤗

model_cards/dbmdz/electra-small-turkish-cased-discriminator/README.md

@@ -0,0 +1,79 @@
+---
+language: turkish
+license: mit
+---
+
+# 🤗 + 📚 dbmdz Turkish ELECTRA model
+
+In this repository the MDZ Digital Library team (dbmdz) at the Bavarian State
+Library open sources a cased ELECTRA small model for Turkish 🎉
+
+# Turkish ELECTRA model
+
+We release a small ELEC**TR**A model for Turkish, that was trained on the same data as *BERTurk*.
+
+> ELECTRA is a new method for self-supervised language representation learning. It can be used to
+> pre-train transformer networks using relatively little compute. ELECTRA models are trained to
+> distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to
+> the discriminator of a GAN.
+
+More details about ELECTRA can be found in the [ICLR paper](https://openreview.net/forum?id=r1xMH1BtvB)
+or in the [official ELECTRA repository](https://github.com/google-research/electra) on GitHub.
+
+## Stats
+
+The current version of the model is trained on a filtered and sentence
+segmented version of the Turkish [OSCAR corpus](https://traces1.inria.fr/oscar/),
+a recent Wikipedia dump, various [OPUS corpora](http://opus.nlpl.eu/) and a
+special corpus provided by [Kemal Oflazer](http://www.andrew.cmu.edu/user/ko/).
+
+The final training corpus has a size of 35GB and 44,04,976,662 tokens.
+
+Thanks to Google's TensorFlow Research Cloud (TFRC) we could train a cased model
+on a TPU v3-8 for 1M steps.
+
+## Model weights
+
+[Transformers](https://github.com/huggingface/transformers)
+compatible weights for both PyTorch and TensorFlow are available.
+
+| Model                                             | Downloads
+| ------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------
+| `dbmdz/electra-small-turkish-cased-discriminator` | [`config.json`](https://cdn.huggingface.co/dbmdz/electra-small-turkish-cased-discriminator/config.json) • [`pytorch_model.bin`](https://cdn.huggingface.co/dbmdz/electra-small-turkish-cased-discriminator/pytorch_model.bin) • [`vocab.txt`](https://cdn.huggingface.co/dbmdz/electra-small-turkish-cased-discriminator/vocab.txt)
+
+## Usage
+
+With Transformers >= 2.8 our ELECTRA small cased model can be loaded like:
+
+```python
+from transformers import AutoModelWithLMHead, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("dbmdz/electra-small-turkish-cased-discriminator")
+model = AutoModelWithLMHead.from_pretrained("dbmdz/electra-small-turkish-cased-discriminator")
+```
+
+## Results
+
+For results on PoS tagging or NER tasks, please refer to
+[this repository](https://github.com/stefan-it/turkish-bert/electra).
+
+# Huggingface model hub
+
+All models are available on the [Huggingface model hub](https://huggingface.co/dbmdz).
+
+# Contact (Bugs, Feedback, Contribution and more)
+
+For questions about our ELECTRA models just open an issue
+[here](https://github.com/dbmdz/berts/issues/new) 🤗
+
+# Acknowledgments
+
+Thanks to [Kemal Oflazer](http://www.andrew.cmu.edu/user/ko/) for providing us
+additional large corpora for Turkish. Many thanks to Reyyan Yeniterzi for providing
+us the Turkish NER dataset for evaluation.
+
+Research supported with Cloud TPUs from Google's TensorFlow Research Cloud (TFRC).
+Thanks for providing access to the TFRC ❤️
+
+Thanks to the generous support from the [Hugging Face](https://huggingface.co/) team,
+it is possible to download both cased and uncased models from their S3 storage 🤗

model_cards/google/electra-base-generator/README.md

@@ -25,7 +25,7 @@ fill_mask = pipeline(
 )
 
 print(
-	fill_mask(f"HuggingFace is creating a {nlp.tokenizer.mask_token} that the community uses to solve NLP tasks.")
+	fill_mask(f"HuggingFace is creating a {fill_mask.tokenizer.mask_token} that the community uses to solve NLP tasks.")
 )
 
 ```

model_cards/lserinol/bert-turkish-question-answering/README.md

@@ -17,6 +17,7 @@ nlp({
 
 ```python
 from transformers import AutoTokenizer, AutoModelForQuestionAnswering
+import torch
 
 tokenizer = AutoTokenizer.from_pretrained("lserinol/bert-turkish-question-answering")
 model = AutoModelForQuestionAnswering.from_pretrained("lserinol/bert-turkish-question-answering")

model_cards/savasy/bert-base-turkish-sentiment-cased/README.md

@@ -38,7 +38,7 @@ The dataset is used by following papers
 * 1 Yildirim, Savaş. (2020). Comparing Deep Neural Networks to Traditional Models for Sentiment Analysis in Turkish Language. 10.1007/978-981-15-1216-2_12. 
 * 2 Demirtas, Erkin and Mykola Pechenizkiy. 2013. Cross-lingual polarity detection with machine translation. In Proceedings of the Second International Workshop on Issues of Sentiment
 Discovery and Opinion Mining (WISDOM ’13)
-* Hayran, A.,   Sert, M. (2017), "Sentiment Analysis on Microblog Data based on Word Embedding and Fusion Techniques", IEEE 25th Signal Processing and Communications Applications Conference (SIU 2017), Belek, Turkey
+* [3] Hayran, A.,   Sert, M. (2017), "Sentiment Analysis on Microblog Data based on Word Embedding and Fusion Techniques", IEEE 25th Signal Processing and Communications Applications Conference (SIU 2017), Belek, Turkey
 
 # Training
 
@@ -121,8 +121,8 @@ Suppose your file has lots of lines of comment and label (1 or 0) at the end  (t
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
 
 f="/path/to/your/file/yourfile.tsv"
-model = AutoModelForSequenceClassification.from_pretrained(folder)
-tokenizer = AutoTokenizer.from_pretrained(folder)
+model = AutoModelForSequenceClassification.from_pretrained("savasy/bert-base-turkish-sentiment-cased")
+tokenizer = AutoTokenizer.from_pretrained("savasy/bert-base-turkish-sentiment-cased")
 sa= pipeline("sentiment-analysis", tokenizer=tokenizer, model=model)
 
 i,crr=0,0

notebooks/03-pipelines.ipynb

@@ -30,7 +30,8 @@
     },
     "colab": {
       "name": "03-pipelines.ipynb",
-      "provenance": []
+      "provenance": [],
+      "include_colab_link": true
     },
     "widgets": {
       "application/vnd.jupyter.widget-state+json": {
@@ -1504,6 +1505,251 @@
             "left": null
           }
         },
+        "3c86415352574190b71e1fe5a15d36f1": {
+          "model_module": "@jupyter-widgets/controls",
+          "model_name": "HBoxModel",
+          "state": {
+            "_view_name": "HBoxView",
+            "_dom_classes": [],
+            "_model_name": "HBoxModel",
+            "_view_module": "@jupyter-widgets/controls",
+            "_model_module_version": "1.5.0",
+            "_view_count": null,
+            "_view_module_version": "1.5.0",
+            "box_style": "",
+            "layout": "IPY_MODEL_dd2c9dd935754cf2802233053554c21c",
+            "_model_module": "@jupyter-widgets/controls",
+            "children": [
+              "IPY_MODEL_8ae3be32d9c845e59fdb1c47884d48aa",
+              "IPY_MODEL_4dea0031f3554752ad5aad01fe516a60"
+            ]
+          }
+        },
+        "dd2c9dd935754cf2802233053554c21c": {
+          "model_module": "@jupyter-widgets/base",
+          "model_name": "LayoutModel",
+          "state": {
+            "_view_name": "LayoutView",
+            "grid_template_rows": null,
+            "right": null,
+            "justify_content": null,
+            "_view_module": "@jupyter-widgets/base",
+            "overflow": null,
+            "_model_module_version": "1.2.0",
+            "_view_count": null,
+            "flex_flow": null,
+            "width": null,
+            "min_width": null,
+            "border": null,
+            "align_items": null,
+            "bottom": null,
+            "_model_module": "@jupyter-widgets/base",
+            "top": null,
+            "grid_column": null,
+            "overflow_y": null,
+            "overflow_x": null,
+            "grid_auto_flow": null,
+            "grid_area": null,
+            "grid_template_columns": null,
+            "flex": null,
+            "_model_name": "LayoutModel",
+            "justify_items": null,
+            "grid_row": null,
+            "max_height": null,
+            "align_content": null,
+            "visibility": null,
+            "align_self": null,
+            "height": null,
+            "min_height": null,
+            "padding": null,
+            "grid_auto_rows": null,
+            "grid_gap": null,
+            "max_width": null,
+            "order": null,
+            "_view_module_version": "1.2.0",
+            "grid_template_areas": null,
+            "object_position": null,
+            "object_fit": null,
+            "grid_auto_columns": null,
+            "margin": null,
+            "display": null,
+            "left": null
+          }
+        },
+        "8ae3be32d9c845e59fdb1c47884d48aa": {
+          "model_module": "@jupyter-widgets/controls",
+          "model_name": "FloatProgressModel",
+          "state": {
+            "_view_name": "ProgressView",
+            "style": "IPY_MODEL_1efb96d931a446de92f1930b973ae846",
+            "_dom_classes": [],
+            "description": "Downloading: 100%",
+            "_model_name": "FloatProgressModel",
+            "bar_style": "success",
+            "max": 230,
+            "_view_module": "@jupyter-widgets/controls",
+            "_model_module_version": "1.5.0",
+            "value": 230,
+            "_view_count": null,
+            "_view_module_version": "1.5.0",
+            "orientation": "horizontal",
+            "min": 0,
+            "description_tooltip": null,
+            "_model_module": "@jupyter-widgets/controls",
+            "layout": "IPY_MODEL_6a4f5aab5ba949fd860b5a35bba7db9c"
+          }
+        },
+        "4dea0031f3554752ad5aad01fe516a60": {
+          "model_module": "@jupyter-widgets/controls",
+          "model_name": "HTMLModel",
+          "state": {
+            "_view_name": "HTMLView",
+            "style": "IPY_MODEL_4b02b2e964ad49af9f7ce7023131ceb8",
+            "_dom_classes": [],
+            "description": "",
+            "_model_name": "HTMLModel",
+            "placeholder": "​",
+            "_view_module": "@jupyter-widgets/controls",
+            "_model_module_version": "1.5.0",
+            "value": " 230/230 [00:00<00:00, 8.69kB/s]",
+            "_view_count": null,
+            "_view_module_version": "1.5.0",
+            "description_tooltip": null,
+            "_model_module": "@jupyter-widgets/controls",
+            "layout": "IPY_MODEL_0ae8a68c3668401da8d8a6d5ec9cac8f"
+          }
+        },
+        "1efb96d931a446de92f1930b973ae846": {
+          "model_module": "@jupyter-widgets/controls",
+          "model_name": "ProgressStyleModel",
+          "state": {
+            "_view_name": "StyleView",
+            "_model_name": "ProgressStyleModel",
+            "description_width": "initial",
+            "_view_module": "@jupyter-widgets/base",
+            "_model_module_version": "1.5.0",
+            "_view_count": null,
+            "_view_module_version": "1.2.0",
+            "bar_color": null,
+            "_model_module": "@jupyter-widgets/controls"
+          }
+        },
+        "6a4f5aab5ba949fd860b5a35bba7db9c": {
+          "model_module": "@jupyter-widgets/base",
+          "model_name": "LayoutModel",
+          "state": {
+            "_view_name": "LayoutView",
+            "grid_template_rows": null,
+            "right": null,
+            "justify_content": null,
+            "_view_module": "@jupyter-widgets/base",
+            "overflow": null,
+            "_model_module_version": "1.2.0",
+            "_view_count": null,
+            "flex_flow": null,
+            "width": null,
+            "min_width": null,
+            "border": null,
+            "align_items": null,
+            "bottom": null,
+            "_model_module": "@jupyter-widgets/base",
+            "top": null,
+            "grid_column": null,
+            "overflow_y": null,
+            "overflow_x": null,
+            "grid_auto_flow": null,
+            "grid_area": null,
+            "grid_template_columns": null,
+            "flex": null,
+            "_model_name": "LayoutModel",
+            "justify_items": null,
+            "grid_row": null,
+            "max_height": null,
+            "align_content": null,
+            "visibility": null,
+            "align_self": null,
+            "height": null,
+            "min_height": null,
+            "padding": null,
+            "grid_auto_rows": null,
+            "grid_gap": null,
+            "max_width": null,
+            "order": null,
+            "_view_module_version": "1.2.0",
+            "grid_template_areas": null,
+            "object_position": null,
+            "object_fit": null,
+            "grid_auto_columns": null,
+            "margin": null,
+            "display": null,
+            "left": null
+          }
+        },
+        "4b02b2e964ad49af9f7ce7023131ceb8": {
+          "model_module": "@jupyter-widgets/controls",
+          "model_name": "DescriptionStyleModel",
+          "state": {
+            "_view_name": "StyleView",
+            "_model_name": "DescriptionStyleModel",
+            "description_width": "",
+            "_view_module": "@jupyter-widgets/base",
+            "_model_module_version": "1.5.0",
+            "_view_count": null,
+            "_view_module_version": "1.2.0",
+            "_model_module": "@jupyter-widgets/controls"
+          }
+        },
+        "0ae8a68c3668401da8d8a6d5ec9cac8f": {
+          "model_module": "@jupyter-widgets/base",
+          "model_name": "LayoutModel",
+          "state": {
+            "_view_name": "LayoutView",
+            "grid_template_rows": null,
+            "right": null,
+            "justify_content": null,
+            "_view_module": "@jupyter-widgets/base",
+            "overflow": null,
+            "_model_module_version": "1.2.0",
+            "_view_count": null,
+            "flex_flow": null,
+            "width": null,
+            "min_width": null,
+            "border": null,
+            "align_items": null,
+            "bottom": null,
+            "_model_module": "@jupyter-widgets/base",
+            "top": null,
+            "grid_column": null,
+            "overflow_y": null,
+            "overflow_x": null,
+            "grid_auto_flow": null,
+            "grid_area": null,
+            "grid_template_columns": null,
+            "flex": null,
+            "_model_name": "LayoutModel",
+            "justify_items": null,
+            "grid_row": null,
+            "max_height": null,
+            "align_content": null,
+            "visibility": null,
+            "align_self": null,
+            "height": null,
+            "min_height": null,
+            "padding": null,
+            "grid_auto_rows": null,
+            "grid_gap": null,
+            "max_width": null,
+            "order": null,
+            "_view_module_version": "1.2.0",
+            "grid_template_areas": null,
+            "object_position": null,
+            "object_fit": null,
+            "grid_auto_columns": null,
+            "margin": null,
+            "display": null,
+            "left": null
+          }
+        },
         "fd44cf6ab17e4b768b2e1d5cb8ce5af9": {
           "model_module": "@jupyter-widgets/controls",
           "model_name": "HBoxModel",
@@ -2105,6 +2351,16 @@
     }
   },
   "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "view-in-github",
+        "colab_type": "text"
+      },
+      "source": [
+        "<a href=\"https://colab.research.google.com/github/huggingface/transformers/blob/generation_pipeline_docs/notebooks/03-pipelines.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+      ]
+    },
     {
       "cell_type": "markdown",
       "metadata": {
@@ -2170,13 +2426,29 @@
         },
         "id": "4maAknWNrl_N",
         "colab_type": "code",
-        "colab": {}
+        "colab": {
+          "base_uri": "https://localhost:8080/",
+          "height": 102
+        },
+        "outputId": "467e3cc8-a069-47da-8029-86e4142c7dde"
       },
       "source": [
         "!pip install -q transformers"
       ],
-      "execution_count": 0,
-      "outputs": []
+      "execution_count": 2,
+      "outputs": [
+        {
+          "output_type": "stream",
+          "text": [
+            "\u001b[K     |████████████████████████████████| 645kB 4.4MB/s \n",
+            "\u001b[K     |████████████████████████████████| 3.8MB 11.7MB/s \n",
+            "\u001b[K     |████████████████████████████████| 890kB 51.5MB/s \n",
+            "\u001b[K     |████████████████████████████████| 1.0MB 46.0MB/s \n",
+            "\u001b[?25h  Building wheel for sacremoses (setup.py) ... \u001b[?25l\u001b[?25hdone\n"
+          ],
+          "name": "stdout"
+        }
+      ]
     },
     {
       "cell_type": "code",
@@ -2219,6 +2491,7 @@
         },
         "id": "AMRXHQw9rl_d",
         "colab_type": "code",
+        "outputId": "a7a10851-b71e-4553-9afc-04066120410d",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 83,
@@ -2232,14 +2505,13 @@
             "ad84da685cf44abb90d17d9d2e023b48",
             "a246f9eea2d7440cb979e728741d2e32"
           ]
-        },
-        "outputId": "a7a10851-b71e-4553-9afc-04066120410d"
+        }
       },
       "source": [
         "nlp_sentence_classif = pipeline('sentiment-analysis')\n",
         "nlp_sentence_classif('Such a nice weather outside !')"
       ],
-      "execution_count": 3,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2300,6 +2572,7 @@
         },
         "id": "B3BDRX_Krl_n",
         "colab_type": "code",
+        "outputId": "a6b90b11-a272-4ecb-960d-4c682551b399",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 185,
@@ -2313,14 +2586,13 @@
             "405afa5bb8b840d8bc0850e02f593ce4",
             "78c718e3d5fa4cb892217260bea6d540"
           ]
-        },
-        "outputId": "a6b90b11-a272-4ecb-960d-4c682551b399"
+        }
       },
       "source": [
         "nlp_token_class = pipeline('ner')\n",
         "nlp_token_class('Hugging Face is a French company based in New-York.')"
       ],
-      "execution_count": 4,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2384,6 +2656,7 @@
         },
         "id": "ND_8LzQKrl_u",
         "colab_type": "code",
+        "outputId": "c59ae695-c465-4de6-fa6e-181d8f1a3992",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 117,
@@ -2397,14 +2670,13 @@
             "cd64e3f20b23483daa79712bde6622ea",
             "67cbaa1f55d24e62ad6b022af36bca56"
           ]
-        },
-        "outputId": "c59ae695-c465-4de6-fa6e-181d8f1a3992"
+        }
       },
       "source": [
         "nlp_qa = pipeline('question-answering')\n",
         "nlp_qa(context='Hugging Face is a French company based in New-York.', question='Where is based Hugging Face ?')"
       ],
-      "execution_count": 5,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2470,6 +2742,7 @@
         },
         "id": "zpJQ2HXNrl_4",
         "colab_type": "code",
+        "outputId": "3fb62e7a-25a6-4b06-ced8-51eb8aa6bf33",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 321,
@@ -2483,14 +2756,13 @@
             "a35703cc8ff44e93a8c0eb413caddc40",
             "9df7014c99b343f3b178fa020ff56010"
           ]
-        },
-        "outputId": "3fb62e7a-25a6-4b06-ced8-51eb8aa6bf33"
+        }
       },
       "source": [
         "nlp_fill = pipeline('fill-mask')\n",
         "nlp_fill('Hugging Face is a French company based in ' + nlp_fill.tokenizer.mask_token)"
       ],
-      "execution_count": 6,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2560,11 +2832,11 @@
       "metadata": {
         "id": "8BaOgzi1u1Yc",
         "colab_type": "code",
+        "outputId": "2168e437-cfba-4247-a38c-07f02f555c6e",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 88
-        },
-        "outputId": "2168e437-cfba-4247-a38c-07f02f555c6e"
+        }
       },
       "source": [
         "TEXT_TO_SUMMARIZE = \"\"\" \n",
@@ -2590,7 +2862,7 @@
         "summarizer = pipeline('summarization')\n",
         "summarizer(TEXT_TO_SUMMARIZE)"
       ],
-      "execution_count": 7,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "stream",
@@ -2631,6 +2903,7 @@
       "metadata": {
         "id": "8FwayP4nwV3Z",
         "colab_type": "code",
+        "outputId": "66956816-c924-4718-fe58-cabef7d51974",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 83,
@@ -2644,15 +2917,14 @@
             "ad78042ee71a41fd989e4b4ce9d2e3c1",
             "40c8d2617f3d4c84b923b140456fa5da"
           ]
-        },
-        "outputId": "66956816-c924-4718-fe58-cabef7d51974"
+        }
       },
       "source": [
         "# English to French\n",
         "translator = pipeline('translation_en_to_fr')\n",
         "translator(\"HuggingFace is a French company that is based in New York City. HuggingFace's mission is to solve NLP one commit at a time\")"
       ],
-      "execution_count": 8,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2696,6 +2968,7 @@
       "metadata": {
         "colab_type": "code",
         "id": "ra0-WfznwoIW",
+        "outputId": "278a3d5f-cc42-40bc-a9db-c92ec5a3a2f0",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 83,
@@ -2709,15 +2982,14 @@
             "4486f8a2efc34b9aab3864eb5ad2ba48",
             "d6228324f3444aa6bd1323d65ae4ff75"
           ]
-        },
-        "outputId": "278a3d5f-cc42-40bc-a9db-c92ec5a3a2f0"
+        }
       },
       "source": [
         "# English to German\n",
         "translator = pipeline('translation_en_to_de')\n",
         "translator(\"The history of natural language processing (NLP) generally started in the 1950s, although work can be found from earlier periods.\")"
       ],
-      "execution_count": 9,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2756,6 +3028,89 @@
         }
       ]
     },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "qPUpg0M8hCtB",
+        "colab_type": "text"
+      },
+      "source": [
+        "## 7. Text Generation\n",
+        "\n",
+        "Text generation is currently supported by GPT-2, OpenAi-GPT, TransfoXL, XLNet, CTRL and Reformer."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "5pKfxTxohXuZ",
+        "colab_type": "code",
+        "colab": {
+          "base_uri": "https://localhost:8080/",
+          "height": 120,
+          "referenced_widgets": [
+            "3c86415352574190b71e1fe5a15d36f1",
+            "dd2c9dd935754cf2802233053554c21c",
+            "8ae3be32d9c845e59fdb1c47884d48aa",
+            "4dea0031f3554752ad5aad01fe516a60",
+            "1efb96d931a446de92f1930b973ae846",
+            "6a4f5aab5ba949fd860b5a35bba7db9c",
+            "4b02b2e964ad49af9f7ce7023131ceb8",
+            "0ae8a68c3668401da8d8a6d5ec9cac8f"
+          ]
+        },
+        "outputId": "8705f6b4-2413-4ac6-f72d-e5ecce160662"
+      },
+      "source": [
+        "text_generator = pipeline(\"text-generation\")\n",
+        "text_generator(\"Today is a beautiful day and I will\")"
+      ],
+      "execution_count": 5,
+      "outputs": [
+        {
+          "output_type": "display_data",
+          "data": {
+            "application/vnd.jupyter.widget-view+json": {
+              "model_id": "3c86415352574190b71e1fe5a15d36f1",
+              "version_minor": 0,
+              "version_major": 2
+            },
+            "text/plain": [
+              "HBox(children=(FloatProgress(value=0.0, description='Downloading', max=230.0, style=ProgressStyle(description_…"
+            ]
+          },
+          "metadata": {
+            "tags": []
+          }
+        },
+        {
+          "output_type": "stream",
+          "text": [
+            "\n"
+          ],
+          "name": "stdout"
+        },
+        {
+          "output_type": "stream",
+          "text": [
+            "Setting `pad_token_id` to 50256 (first `eos_token_id`) to generate sequence\n"
+          ],
+          "name": "stderr"
+        },
+        {
+          "output_type": "execute_result",
+          "data": {
+            "text/plain": [
+              "[{'generated_text': 'Today is a beautiful day and I will celebrate my birthday!\"\\n\\nThe mother told CNN the two had planned their meal together. After dinner, she added that she and I walked down the street and stopped at a diner near her home. \"He'}]"
+            ]
+          },
+          "metadata": {
+            "tags": []
+          },
+          "execution_count": 5
+        }
+      ]
+    },
     {
       "cell_type": "markdown",
       "metadata": {
@@ -2763,7 +3118,7 @@
         "colab_type": "text"
       },
       "source": [
-        "## 7. Projection - Features Extraction "
+        "## 8. Projection - Features Extraction "
       ]
     },
     {
@@ -2775,6 +3130,7 @@
         },
         "id": "O4SjR1QQrl__",
         "colab_type": "code",
+        "outputId": "2ce966d5-7a89-4488-d48f-626d1c2a8222",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 83,
@@ -2788,8 +3144,7 @@
             "31d97ecf78fa412c99e6659196d82828",
             "c6be5d48ec3c4c799d1445607e5f1ac6"
           ]
-        },
-        "outputId": "2ce966d5-7a89-4488-d48f-626d1c2a8222"
+        }
       },
       "source": [
         "import numpy as np\n",
@@ -2797,7 +3152,7 @@
         "output = nlp_features('Hugging Face is a French company based in Paris')\n",
         "np.array(output).shape   # (Samples, Tokens, Vector Size)\n"
       ],
-      "execution_count": 10,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2861,6 +3216,7 @@
         },
         "id": "yFlBPQHtrmAH",
         "colab_type": "code",
+        "outputId": "03cc3207-a7e8-49fd-904a-63a7a1d0eb7a",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 116,
@@ -2872,8 +3228,7 @@
             "62b10ca525cc4ac68f3a006434eb7416",
             "211109537fbe4e60b89a238c89db1346"
           ]
-        },
-        "outputId": "03cc3207-a7e8-49fd-904a-63a7a1d0eb7a"
+        }
       },
       "source": [
         "task = widgets.Dropdown(\n",
@@ -2906,7 +3261,7 @@
         "input.on_submit(forward)\n",
         "display(task, input)"
       ],
-      "execution_count": 11,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",
@@ -2958,6 +3313,7 @@
         },
         "id": "GCoKbBTYrmAN",
         "colab_type": "code",
+        "outputId": "57c3a647-160a-4b3a-e852-e7a1daf1294a",
         "colab": {
           "base_uri": "https://localhost:8080/",
           "height": 143,
@@ -2969,8 +3325,7 @@
             "d305ba1662e3466c93ab5cca7ebf8f33",
             "879f7a3747ad455d810c7a29918648ee"
           ]
-        },
-        "outputId": "57c3a647-160a-4b3a-e852-e7a1daf1294a"
+        }
       },
       "source": [
         "context = widgets.Textarea(\n",
@@ -2995,7 +3350,7 @@
         "query.on_submit(forward)\n",
         "display(context, query)"
       ],
-      "execution_count": 12,
+      "execution_count": 0,
       "outputs": [
         {
           "output_type": "display_data",

setup.py

@@ -67,8 +67,8 @@ extras = {}
 
 extras["mecab"] = ["mecab-python3"]
 extras["sklearn"] = ["scikit-learn"]
-extras["tf"] = ["tensorflow <= 2.1"]
-extras["tf-cpu"] = ["tensorflow-cpu  <= 2.1"]
+extras["tf"] = ["tensorflow"]
+extras["tf-cpu"] = ["tensorflow-cpu"]
 extras["torch"] = ["torch"]
 
 extras["serving"] = ["pydantic", "uvicorn", "fastapi", "starlette"]
@@ -78,14 +78,14 @@ extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator"]
 extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme"]
 extras["quality"] = [
     "black",
-    "isort @ git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort",
+    "isort",
     "flake8==3.7.9",
 ]
-extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3", "scikit-learn", "tensorflow <= 2.1", "torch"]
+extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3", "scikit-learn", "tensorflow", "torch"]
 
 setup(
     name="transformers",
-    version="2.9.0",
+    version="2.9.1",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",

src/transformers/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-__version__ = "2.9.0"
+__version__ = "2.9.1"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -359,6 +359,7 @@ if is_tf_available():
     from .modeling_tf_auto import (
         TFAutoModel,
         TFAutoModelForPreTraining,
+        TFAutoModelForMultipleChoice,
         TFAutoModelForSequenceClassification,
         TFAutoModelForQuestionAnswering,
         TFAutoModelWithLMHead,
@@ -493,6 +494,7 @@ if is_tf_available():
         TFAlbertModel,
         TFAlbertForPreTraining,
         TFAlbertForMaskedLM,
+        TFAlbertForMultipleChoice,
         TFAlbertForSequenceClassification,
         TFAlbertForQuestionAnswering,
         TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,

src/transformers/convert_marian_to_pytorch.py

@@ -95,6 +95,97 @@ def find_model_file(dest_dir):  # this one better
     return model_file
 
 
+# Group Names Logic: change long opus model names to something shorter, like opus-mt-en-ROMANCE
+ROM_GROUP = "fr+fr_BE+fr_CA+fr_FR+wa+frp+oc+ca+rm+lld+fur+lij+lmo+es+es_AR+es_CL+es_CO+es_CR+es_DO+es_EC+es_ES+es_GT+es_HN+es_MX+es_NI+es_PA+es_PE+es_PR+es_SV+es_UY+es_VE+pt+pt_br+pt_BR+pt_PT+gl+lad+an+mwl+it+it_IT+co+nap+scn+vec+sc+ro+la"
+GROUPS = [
+    ("cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh", "ZH"),
+    (ROM_GROUP, "ROMANCE"),
+    ("de+nl+fy+af+da+fo+is+no+nb+nn+sv", "NORTH_EU"),
+    ("da+fo+is+no+nb+nn+sv", "SCANDINAVIA"),
+    ("se+sma+smj+smn+sms", "SAMI"),
+    ("nb_NO+nb+nn_NO+nn+nog+no_nb+no", "NORWAY"),
+    ("ga+cy+br+gd+kw+gv", "CELTIC"),  # https://en.wikipedia.org/wiki/Insular_Celtic_languages
+]
+GROUP_TO_OPUS_NAME = {
+    "opus-mt-ZH-de": "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-de",
+    "opus-mt-ZH-fi": "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-fi",
+    "opus-mt-ZH-sv": "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-sv",
+    "opus-mt-SCANDINAVIA-SCANDINAVIA": "da+fo+is+no+nb+nn+sv-da+fo+is+no+nb+nn+sv",
+    "opus-mt-NORTH_EU-NORTH_EU": "de+nl+fy+af+da+fo+is+no+nb+nn+sv-de+nl+fy+af+da+fo+is+no+nb+nn+sv",
+    "opus-mt-de-ZH": "de-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh",
+    "opus-mt-en_el_es_fi-en_el_es_fi": "en+el+es+fi-en+el+es+fi",
+    "opus-mt-en-ROMANCE": "en-fr+fr_BE+fr_CA+fr_FR+wa+frp+oc+ca+rm+lld+fur+lij+lmo+es+es_AR+es_CL+es_CO+es_CR+es_DO"
+    "+es_EC+es_ES+es_GT+es_HN+es_MX+es_NI+es_PA+es_PE+es_PR+es_SV+es_UY+es_VE+pt+pt_br+pt_BR"
+    "+pt_PT+gl+lad+an+mwl+it+it_IT+co+nap+scn+vec+sc+ro+la",
+    "opus-mt-en-CELTIC": "en-ga+cy+br+gd+kw+gv",
+    "opus-mt-es-NORWAY": "es-nb_NO+nb+nn_NO+nn+nog+no_nb+no",
+    "opus-mt-fi_nb_no_nn_ru_sv_en-SAMI": "fi+nb+no+nn+ru+sv+en-se+sma+smj+smn+sms",
+    "opus-mt-fi-ZH": "fi-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh",
+    "opus-mt-fi-NORWAY": "fi-nb_NO+nb+nn_NO+nn+nog+no_nb+no",
+    "opus-mt-ROMANCE-en": "fr+fr_BE+fr_CA+fr_FR+wa+frp+oc+ca+rm+lld+fur+lij+lmo+es+es_AR+es_CL+es_CO+es_CR+es_DO"
+    "+es_EC+es_ES+es_GT+es_HN+es_MX+es_NI+es_PA+es_PE+es_PR+es_SV+es_UY+es_VE+pt+pt_br+pt_BR"
+    "+pt_PT+gl+lad+an+mwl+it+it_IT+co+nap+scn+vec+sc+ro+la-en",
+    "opus-mt-CELTIC-en": "ga+cy+br+gd+kw+gv-en",
+    "opus-mt-sv-ZH": "sv-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh",
+    "opus-mt-sv-NORWAY": "sv-nb_NO+nb+nn_NO+nn+nog+no_nb+no",
+}
+OPUS_GITHUB_URL = "https://github.com/Helsinki-NLP/OPUS-MT-train/blob/master/models/"
+ORG_NAME = "Helsinki-NLP/"
+
+
+def convert_opus_name_to_hf_name(x):
+    for substr, grp_name in GROUPS:
+        x = x.replace(substr, grp_name)
+    return x.replace("+", "_")
+
+
+def convert_hf_name_to_opus_name(hf_model_name):
+    """Relies on the assumption that there are no language codes like pt_br in models that are not in GROUP_TO_OPUS_NAME."""
+    hf_model_name = remove_prefix(hf_model_name, ORG_NAME)
+    if hf_model_name in GROUP_TO_OPUS_NAME:
+        opus_w_prefix = GROUP_TO_OPUS_NAME[hf_model_name]
+    else:
+        opus_w_prefix = hf_model_name.replace("_", "+")
+    return remove_prefix(opus_w_prefix, "opus-mt-")
+
+
+def write_model_card(
+    hf_model_name: str,
+    repo_path="OPUS-MT-train/models/",
+    dry_run=False,
+    model_card_dir=Path("marian_converted/model_cards/Helsinki-NLP/"),
+) -> str:
+    """Copy the most recent model's readme section from opus, and add metadata.
+    upload command: s3cmd sync --recursive model_card_dir s3://models.huggingface.co/bert/Helsinki-NLP/
+    """
+    hf_model_name = remove_prefix(hf_model_name, ORG_NAME)
+    opus_name: str = convert_hf_name_to_opus_name(hf_model_name)
+    opus_src, opus_tgt = [x.split("+") for x in opus_name.split("-")]
+    readme_url = OPUS_GITHUB_URL + f"{opus_name}/README.md"
+    s, t = ",".join(opus_src), ",".join(opus_tgt)
+    extra_markdown = f"### {hf_model_name}\n\n* source languages: {s}\n* target languages: {t}\n*  OPUS readme: [{opus_name}]({readme_url})\n"
+    # combine with opus markdown
+    opus_readme_path = Path(f"{repo_path}{opus_name}/README.md")
+    assert opus_readme_path.exists(), opus_readme_path
+    content = opus_readme_path.open().read()
+    content = content.split("\n# ")[-1]  # Get the lowest level 1 header in the README -- the most recent model.
+    content = "*".join(content.split("*")[1:])
+    content = extra_markdown + "\n* " + content.replace("download", "download original weights")
+    if dry_run:
+        return content
+    # Save string to model_cards/hf_model_name/readme.md
+    model_card_dir.mkdir(exist_ok=True)
+    sub_dir = model_card_dir / hf_model_name
+    sub_dir.mkdir(exist_ok=True)
+    dest = sub_dir / "README.md"
+    dest.open("w").write(content)
+    return content
+
+
+def get_clean_model_id_mapping(multiling_model_ids):
+    return {x: convert_opus_name_to_hf_name(x) for x in multiling_model_ids}
+
+
 def make_registry(repo_path="Opus-MT-train/models"):
     if not (Path(repo_path) / "fr-en" / "README.md").exists():
         raise ValueError(
@@ -109,10 +200,7 @@ def make_registry(repo_path="Opus-MT-train/models"):
         else:
             lns = list(open(p / "README.md").readlines())
             results[p.name] = _parse_readme(lns)
-    return [(k, v["pre-processing"], v["download"]) for k, v in results.items()]
-
-
-CH_GROUP = "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh"
+    return [(k, v["pre-processing"], v["download"], v["download"][:-4] + ".test.txt") for k, v in results.items()]
 
 
 def convert_all_sentencepiece_models(model_list=None, repo_path=None):
@@ -122,12 +210,12 @@ def convert_all_sentencepiece_models(model_list=None, repo_path=None):
     dest_dir.mkdir(exist_ok=True)
     if model_list is None:
         model_list: list = make_registry(repo_path=repo_path)
-    for k, prepro, download in tqdm(model_list):
+    for k, prepro, download, test_set_url in tqdm(model_list):
         if "SentencePiece" not in prepro:  # dont convert BPE models.
             continue
         if not os.path.exists(save_dir / k / "pytorch_model.bin"):
             download_and_unzip(download, save_dir / k)
-        pair_name = k.replace(CH_GROUP, "ch_group")
+        pair_name = convert_opus_name_to_hf_name(k)
         convert(save_dir / k, dest_dir / f"opus-mt-{pair_name}")
 
 
@@ -135,12 +223,10 @@ def lmap(f, x) -> List:
     return list(map(f, x))
 
 
-def fetch_test_set(readmes_raw, pair):
+def fetch_test_set(test_set_url):
     import wget
 
-    download_url = readmes_raw[pair]["download"]
-    test_set_url = download_url[:-4] + ".test.txt"
-    fname = wget.download(test_set_url, f"opus_test_{pair}.txt")
+    fname = wget.download(test_set_url, f"opus_test.txt")
     lns = Path(fname).open().readlines()
     src = lmap(str.strip, lns[::4])
     gold = lmap(str.strip, lns[1::4])

src/transformers/modeling_albert.py

@@ -175,7 +175,7 @@ class AlbertEmbeddings(BertEmbeddings):
     def __init__(self, config):
         super().__init__(config)
 
-        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=0)
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
         self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
         self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
         self.LayerNorm = torch.nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)

src/transformers/modeling_bart.py

@@ -886,7 +886,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
         if new_num_tokens <= old_num_tokens:
             new_bias = self.final_logits_bias[:, :new_num_tokens]
         else:
-            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens))
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
             new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
         self.register_buffer("final_logits_bias", new_bias)
 
@@ -980,12 +980,12 @@ class BartForConditionalGeneration(PretrainedBartModel):
             "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
         }
 
-    def prepare_scores_for_generation(self, scores, cur_len, max_length):
+    def prepare_logits_for_generation(self, logits, cur_len, max_length):
         if cur_len == 1:
-            self._force_token_ids_generation(scores, self.config.bos_token_id)
+            self._force_token_ids_generation(logits, self.config.bos_token_id)
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:
-            self._force_token_ids_generation(scores, self.config.eos_token_id)
-        return scores
+            self._force_token_ids_generation(logits, self.config.eos_token_id)
+        return logits
 
     def _force_token_ids_generation(self, scores, token_ids) -> None:
         """force one of token_ids to be generated by setting prob of all other tokens to 0"""

src/transformers/modeling_distilbert.py

@@ -61,7 +61,7 @@ def create_sinusoidal_embeddings(n_pos, dim, out):
 class Embeddings(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.word_embeddings = nn.Embedding(config.vocab_size, config.dim, padding_idx=0)
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.dim, padding_idx=config.pad_token_id)
         self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.dim)
         if config.sinusoidal_pos_embds:
             create_sinusoidal_embeddings(

src/transformers/modeling_marian.py

@@ -31,7 +31,7 @@ class MarianMTModel(BartForConditionalGeneration):
         src = 'fr'  # source language
         trg = 'en'  # target language
         sample_text = "où est l'arrêt de bus ?"
-        mname = f'Helsinki-NLP/opus-mt-{src}-{trg}'   # `Model List`__
+        mname = f'Helsinki-NLP/opus-mt-{src}-{trg}'
 
         model = MarianMTModel.from_pretrained(mname)
         tok = MarianTokenizer.from_pretrained(mname)
@@ -43,7 +43,8 @@ class MarianMTModel(BartForConditionalGeneration):
 
     pretrained_model_archive_map = {}  # see https://huggingface.co/models?search=Helsinki-NLP
 
-    def prepare_scores_for_generation(self, scores, cur_len, max_length):
+    def prepare_logits_for_generation(self, logits, cur_len, max_length):
+        logits[:, self.config.pad_token_id] = float("-inf")
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:
-            self._force_token_ids_generation(scores, self.config.eos_token_id)
-        return scores
+            self._force_token_ids_generation(logits, self.config.eos_token_id)
+        return logits

src/transformers/modeling_tf_albert.py

@@ -21,7 +21,7 @@ import logging
 import tensorflow as tf
 
 from .configuration_albert import AlbertConfig
-from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from .file_utils import MULTIPLE_CHOICE_DUMMY_INPUTS, add_start_docstrings, add_start_docstrings_to_callable
 from .modeling_tf_bert import ACT2FN, TFBertSelfAttention
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer, keras_serializable, shape_list
 from .tokenization_utils import BatchEncoding
@@ -957,3 +957,127 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel):
         outputs = (start_logits, end_logits,) + outputs[2:]
 
         return outputs  # start_logits, end_logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Albert Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel):
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(
+            1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+
+    @property
+    def dummy_inputs(self):
+        """ Dummy inputs to build the network.
+
+        Returns:
+            tf.Tensor with dummy inputs
+        """
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    def call(
+        self,
+        inputs,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        training=False,
+    ):
+        r"""
+    Return:
+        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
+            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when :obj:`config.output_hidden_states=True`):
+            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_attentions=True``):
+            tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForMultipleChoice
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForMultipleChoice.from_pretrained('albert-base-v2')
+
+        example1 = ["This is a context", "Is it a context? Yes"]
+        example2 = ["This is a context", "Is it a context? No"]
+        encoding = tokenizer.batch_encode_plus([example1, example2], return_tensors='tf', truncation_strategy="only_first", pad_to_max_length=True, max_length=128)
+        outputs = model(encoding["input_ids"][None, :])
+        logits = outputs[0]
+
+        """
+        if isinstance(inputs, (tuple, list)):
+            input_ids = inputs[0]
+            attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
+            token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
+            position_ids = inputs[3] if len(inputs) > 3 else position_ids
+            head_mask = inputs[4] if len(inputs) > 4 else head_mask
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
+        elif isinstance(inputs, dict):
+            print("isdict(1)")
+            input_ids = inputs.get("input_ids")
+            print(input_ids)
+
+            attention_mask = inputs.get("attention_mask", attention_mask)
+            token_type_ids = inputs.get("token_type_ids", token_type_ids)
+            position_ids = inputs.get("position_ids", position_ids)
+            head_mask = inputs.get("head_mask", head_mask)
+            inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
+        else:
+            input_ids = inputs
+
+        if input_ids is not None:
+            num_choices = shape_list(input_ids)[1]
+            seq_length = shape_list(input_ids)[2]
+        else:
+            num_choices = shape_list(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[2]
+
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
+        flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
+        flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None
+        flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None
+
+        flat_inputs = [
+            flat_input_ids,
+            flat_attention_mask,
+            flat_token_type_ids,
+            flat_position_ids,
+            head_mask,
+            inputs_embeds,
+        ]
+
+        outputs = self.albert(flat_inputs, training=training)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output, training=training)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = tf.reshape(logits, (-1, num_choices))
+
+        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # reshaped_logits, (hidden_states), (attentions)

src/transformers/modeling_tf_auto.py

@@ -36,6 +36,7 @@ from .configuration_utils import PretrainedConfig
 from .modeling_tf_albert import (
     TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFAlbertForMaskedLM,
+    TFAlbertForMultipleChoice,
     TFAlbertForPreTraining,
     TFAlbertForQuestionAnswering,
     TFAlbertForSequenceClassification,
@@ -44,6 +45,7 @@ from .modeling_tf_albert import (
 from .modeling_tf_bert import (
     TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFBertForMaskedLM,
+    TFBertForMultipleChoice,
     TFBertForPreTraining,
     TFBertForQuestionAnswering,
     TFBertForSequenceClassification,
@@ -172,6 +174,10 @@ TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
     ]
 )
 
+TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
+    [(BertConfig, TFBertForMultipleChoice), (AlbertConfig, TFAlbertForMultipleChoice)]
+)
+
 TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
     [
         (DistilBertConfig, TFDistilBertForQuestionAnswering),
@@ -662,6 +668,153 @@ class TFAutoModelWithLMHead(object):
         )
 
 
+class TFAutoModelForMultipleChoice:
+    r"""
+        :class:`~transformers.TFAutoModelForMultipleChoice` is a generic model class
+        that will be instantiated as one of the multiple choice model classes of the library
+        when created with the `TFAutoModelForMultipleChoice.from_pretrained(pretrained_model_name_or_path)`
+        class method.
+
+        The `from_pretrained()` method takes care of returning the correct model class instance
+        based on the `model_type` property of the config object, or when it's missing,
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `albert`: TFAlbertForMultipleChoice (Albert model)
+            - contains `bert`: TFBertForMultipleChoice (Bert model)
+
+        This class cannot be instantiated using `__init__()` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "TFAutoModelForMultipleChoice is designed to be instantiated "
+            "using the `TFAutoModelForMultipleChoice.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelForMultipleChoice.from_config(config)` methods."
+        )
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `albert` configuration class: AlbertModel (Albert model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForMulitpleChoice.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        for config_class, model_class in TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING.items():
+            if isinstance(config, config_class):
+                return model_class(config)
+        raise ValueError(
+            "Unrecognized configuration class {} for this kind of TFAutoModel: {}.\n"
+            "Model type should be one of {}.".format(
+                config.__class__,
+                cls.__name__,
+                ", ".join(c.__name__ for c in TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING.keys()),
+            )
+        )
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        r""" Instantiates one of the multiple choice model classes of the library
+        from a pre-trained model configuration.
+
+        The `from_pretrained()` method takes care of returning the correct model class instance
+        based on the `model_type` property of the config object, or when it's missing,
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `albert`: TFRobertaForMultiple (Albert model)
+            - contains `bert`: TFBertForMultipleChoice (Bert model)
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
+        To train the model, you should first set it back in training mode with `model.train()`
+
+        Params:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
+                - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
+                - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
+
+            from_pt: (`Optional`) Boolean
+                Set to True if the Checkpoint is a PyTorch checkpoint.
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+
+                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
+                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments:
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
+
+        Examples::
+
+            model = TFAutoModelFormultipleChoice.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = TFAutoModelFormultipleChoice.from_pretrained('./test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = TFAutoModelFormultipleChoice.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+            model = TFAutoModelFormultipleChoice.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
+
+        """
+        config = kwargs.pop("config", None)
+        if not isinstance(config, PretrainedConfig):
+            config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+
+        for config_class, model_class in TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING.items():
+            if isinstance(config, config_class):
+                return model_class.from_pretrained(pretrained_model_name_or_path, *model_args, config=config, **kwargs)
+        raise ValueError(
+            "Unrecognized configuration class {} for this kind of TFAutoModel: {}.\n"
+            "Model type should be one of {}.".format(
+                config.__class__,
+                cls.__name__,
+                ", ".join(c.__name__ for c in TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING.keys()),
+            )
+        )
+
+
 class TFAutoModelForSequenceClassification(object):
     r"""
         :class:`~transformers.TFAutoModelForSequenceClassification` is a generic model class

src/transformers/modeling_utils.py

@@ -744,8 +744,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
     def prepare_inputs_for_generation(self, input_ids, **kwargs):
         return {"input_ids": input_ids}
 
-    def prepare_scores_for_generation(self, scores, **kwargs):
-        return scores
+    def prepare_logits_for_generation(self, logits, **kwargs):
+        return logits
 
     def _use_cache(self, outputs, use_cache):
         """During generation, decide whether to pass the `past` variable to the next forward pass."""
@@ -857,7 +857,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                 ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
                 Defaults to `None`.
 
-            `What are attention masks? <../glossary.html#attention-mask>`__
+                `What are attention masks? <../glossary.html#attention-mask>`__
 
             decoder_start_token_id=None: (`optional`) int
                 If an encoder-decoder model starts decoding with a different token than BOS.
@@ -1342,10 +1342,13 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
             if temperature != 1.0:
                 next_token_logits = next_token_logits / temperature
 
-            scores = F.log_softmax(next_token_logits, dim=-1)  # (batch_size * num_beams, vocab_size)
             if self.config.is_encoder_decoder and do_sample is False:
-                # TODO (PVP) still a bit hacky here - there might be a better solutino
-                scores = self.prepare_scores_for_generation(scores, cur_len=cur_len, max_length=max_length)
+                # TODO (PVP) still a bit hacky here - there might be a better solution
+                next_token_logits = self.prepare_logits_for_generation(
+                    next_token_logits, cur_len=cur_len, max_length=max_length
+                )
+
+            scores = F.log_softmax(next_token_logits, dim=-1)  # (batch_size * num_beams, vocab_size)
 
             # set eos token prob to zero if min_length is not reached
             if eos_token_id is not None and cur_len < min_length:

src/transformers/tokenization_marian.py

@@ -1,4 +1,5 @@
 import json
+import re
 import warnings
 from typing import Dict, List, Optional, Union
 
@@ -14,7 +15,7 @@ vocab_files_names = {
     "vocab": "vocab.json",
     "tokenizer_config_file": "tokenizer_config.json",
 }
-MODEL_NAMES = ("opus-mt-en-de",)
+MODEL_NAMES = ("opus-mt-en-de",)  # TODO(SS): the only required constant is vocab_files_names
 PRETRAINED_VOCAB_FILES_MAP = {
     k: {m: f"{S3_BUCKET_PREFIX}/Helsinki-NLP/{m}/{fname}" for m in MODEL_NAMES}
     for k, fname in vocab_files_names.items()
@@ -41,6 +42,7 @@ class MarianTokenizer(PreTrainedTokenizer):
     pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
     max_model_input_sizes = {m: 512 for m in MODEL_NAMES}
     model_input_names = ["attention_mask"]  # actually attention_mask, decoder_attention_mask
+    language_code_re = re.compile(">>.+<<")  # type: re.Pattern
 
     def __init__(
         self,
@@ -72,8 +74,6 @@ class MarianTokenizer(PreTrainedTokenizer):
         self.target_lang = target_lang
 
         # load SentencePiece model for pre-processing
-        self.paths = {}
-
         self.spm_source = sentencepiece.SentencePieceProcessor()
         self.spm_source.Load(source_spm)
 
@@ -82,9 +82,7 @@ class MarianTokenizer(PreTrainedTokenizer):
 
         # Multilingual target side: default to using first supported language code.
         self.supported_language_codes: list = [k for k in self.encoder if k.startswith(">>") and k.endswith("<<")]
-        self.tgt_lang_id = None  # will not be used unless it is set through prepare_translation_batch
 
-        # Note(SS): sentence_splitter would require lots of book-keeping.
         try:
             from mosestokenizer import MosesPunctuationNormalizer
 
@@ -93,11 +91,23 @@ class MarianTokenizer(PreTrainedTokenizer):
             warnings.warn("Recommended: pip install mosestokenizer")
             self.punc_normalizer = lambda x: x
 
+    def normalize(self, x: str) -> str:
+        """Cover moses empty string edge case. They return empty list for '' input!"""
+        return self.punc_normalizer(x) if x else ""
+
     def _convert_token_to_id(self, token):
         return self.encoder.get(token, self.encoder[self.unk_token])
 
+    def remove_language_code(self, text: str):
+        """Remove language codes like <<fr>> before sentencepiece"""
+        match = self.language_code_re.match(text)
+        code: list = [match.group(0)] if match else []
+        return code, self.language_code_re.sub("", text)
+
     def _tokenize(self, text: str) -> List[str]:
-        return self.current_spm.EncodeAsPieces(text)
+        code, text = self.remove_language_code(text)
+        pieces = self.current_spm.EncodeAsPieces(text)
+        return code + pieces
 
     def _convert_id_to_token(self, index: int) -> str:
         """Converts an index (integer) in a token (str) using the encoder."""
@@ -125,7 +135,7 @@ class MarianTokenizer(PreTrainedTokenizer):
         pad_to_max_length: bool = True,
         return_tensors: str = "pt",
     ) -> BatchEncoding:
-        """
+        """Prepare model inputs for translation. For best performance, translate one sentence at a time.
         Arguments:
             src_texts: list of src language texts
             tgt_texts: list of tgt language texts
@@ -138,7 +148,10 @@ class MarianTokenizer(PreTrainedTokenizer):
             all shaped bs, seq_len. (BatchEncoding is a dict of string -> tensor or lists).
             If no tgt_text is specified, the only keys will be input_ids and attention_mask.
         """
+        if "" in src_texts:
+            raise ValueError(f"found empty string in src_texts: {src_texts}")
         self.current_spm = self.spm_source
+        src_texts = [self.normalize(t) for t in src_texts]  # this does not appear to do much
         model_inputs: BatchEncoding = self.batch_encode_plus(
             src_texts,
             add_special_tokens=True,

src/transformers/tokenization_utils.py

@@ -173,7 +173,11 @@ class BatchEncoding(UserDict):
 
     """
 
-    def __init__(self, data: Dict[str, Any], encoding: Optional[Union[EncodingFast, Sequence[EncodingFast]]] = None):
+    def __init__(
+        self,
+        data: Optional[Dict[str, Any]] = None,
+        encoding: Optional[Union[EncodingFast, Sequence[EncodingFast]]] = None,
+    ):
         super().__init__(data)
 
         if isinstance(encoding, EncodingFast):
@@ -2191,7 +2195,6 @@ class PreTrainedTokenizer(SpecialTokensMixin):
             .replace(" ' ", "'")
             .replace(" n't", "n't")
             .replace(" 'm", "'m")
-            .replace(" do not", " don't")
             .replace(" 's", "'s")
             .replace(" 've", "'ve")
             .replace(" 're", "'re")

src/transformers/trainer.py

@@ -61,7 +61,12 @@ def is_tensorboard_available():
 try:
     import wandb
 
-    _has_wandb = True
+    wandb.ensure_configured()
+    if wandb.api.api_key is None:
+        _has_wandb = False
+        wandb.termwarn("W&B installed but not logged in.  Run `wandb login` or set the WANDB_API_KEY env variable.")
+    else:
+        _has_wandb = False if os.getenv("WANDB_DISABLED") else True
 except ImportError:
     _has_wandb = False
 
@@ -114,6 +119,8 @@ class Trainer:
     prediction_loss_only: bool
     tb_writer: Optional["SummaryWriter"] = None
     optimizers: Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = None
+    global_step: Optional[int] = None
+    epoch: Optional[float] = None
 
     def __init__(
         self,
@@ -154,9 +161,12 @@ class Trainer:
             logger.warning(
                 "You are instantiating a Trainer but Tensorboard is not installed. You should consider installing it."
             )
-        if not is_wandb_available():
+        if is_wandb_available():
+            self._setup_wandb()
+        else:
             logger.info(
-                "You are instantiating a Trainer but wandb is not installed. Install it to use Weights & Biases logging."
+                "You are instantiating a Trainer but W&B is not installed. To use wandb logging, "
+                "run `pip install wandb; wandb login` see https://docs.wandb.com/huggingface."
             )
         set_seed(self.args.seed)
         # Create output directory if needed
@@ -263,11 +273,25 @@ class Trainer:
         """
         Setup the optional Weights & Biases (`wandb`) integration.
 
-        One can override this method to customize the setup if needed.
+        One can override this method to customize the setup if needed.  Find more information at https://docs.wandb.com/huggingface
+        You can also override the following environment variables:
+
+        Environment:
+            WANDB_WATCH:
+                (Optional, ["gradients", "all", "false"]) "gradients" by default, set to "false" to disable gradient logging
+                or "all" to log gradients and parameters
+            WANDB_PROJECT:
+                (Optional): str - "huggingface" by default, set this to a custom string to store results in a different project
+            WANDB_DISABLED:
+                (Optional): boolean - defaults to false, set to "true" to disable wandb entirely
         """
-        wandb.init(name=self.args.logging_dir, config=vars(self.args))
+        logger.info('Automatic Weights & Biases logging enabled, to disable set os.environ["WANDB_DISABLED"] = "true"')
+        wandb.init(project=os.getenv("WANDB_PROJECT", "huggingface"), config=vars(self.args))
         # keep track of model topology and gradients
-        wandb.watch(self.model)
+        if os.getenv("WANDB_WATCH") != "false":
+            wandb.watch(
+                self.model, log=os.getenv("WANDB_WATCH", "gradients"), log_freq=max(100, self.args.logging_steps)
+            )
 
     def num_examples(self, dataloader: Union[DataLoader, "pl.PerDeviceLoader"]) -> int:
         """
@@ -333,8 +357,6 @@ class Trainer:
         if self.tb_writer is not None:
             self.tb_writer.add_text("args", self.args.to_json_string())
             self.tb_writer.add_hparams(self.args.to_sanitized_dict(), metric_dict={})
-        if is_wandb_available():
-            self._setup_wandb()
 
         # Train!
         if is_tpu_available():
@@ -353,25 +375,26 @@ class Trainer:
         logger.info("  Gradient Accumulation steps = %d", self.args.gradient_accumulation_steps)
         logger.info("  Total optimization steps = %d", t_total)
 
-        global_step = 0
+        self.global_step = 0
+        self.epoch = 0
         epochs_trained = 0
         steps_trained_in_current_epoch = 0
         # Check if continuing training from a checkpoint
         if model_path is not None:
             # set global_step to global_step of last saved checkpoint from model path
             try:
-                global_step = int(model_path.split("-")[-1].split("/")[0])
-                epochs_trained = global_step // (len(train_dataloader) // self.args.gradient_accumulation_steps)
-                steps_trained_in_current_epoch = global_step % (
+                self.global_step = int(model_path.split("-")[-1].split("/")[0])
+                epochs_trained = self.global_step // (len(train_dataloader) // self.args.gradient_accumulation_steps)
+                steps_trained_in_current_epoch = self.global_step % (
                     len(train_dataloader) // self.args.gradient_accumulation_steps
                 )
 
                 logger.info("  Continuing training from checkpoint, will skip to saved global_step")
                 logger.info("  Continuing training from epoch %d", epochs_trained)
-                logger.info("  Continuing training from global step %d", global_step)
+                logger.info("  Continuing training from global step %d", self.global_step)
                 logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
             except ValueError:
-                global_step = 0
+                self.global_step = 0
                 logger.info("  Starting fine-tuning.")
 
         tr_loss = 0.0
@@ -408,34 +431,24 @@ class Trainer:
 
                     scheduler.step()
                     model.zero_grad()
-                    global_step += 1
+                    self.global_step += 1
+                    self.epoch = epoch + (step + 1) / len(epoch_iterator)
 
                     if self.is_local_master():
-                        if (self.args.logging_steps > 0 and global_step % self.args.logging_steps == 0) or (
-                            global_step == 1 and self.args.logging_first_step
+                        if (self.args.logging_steps > 0 and self.global_step % self.args.logging_steps == 0) or (
+                            self.global_step == 1 and self.args.logging_first_step
                         ):
-                            logs = {}
-                            if self.args.evaluate_during_training:
-                                results = self.evaluate()
-                                for key, value in results.items():
-                                    eval_key = "eval_{}".format(key)
-                                    logs[eval_key] = value
-
-                            loss_scalar = (tr_loss - logging_loss) / self.args.logging_steps
-                            learning_rate_scalar = scheduler.get_last_lr()[0]
-                            logs["learning_rate"] = learning_rate_scalar
-                            logs["loss"] = loss_scalar
+                            logs: Dict[str, float] = {}
+                            logs["loss"] = (tr_loss - logging_loss) / self.args.logging_steps
+                            logs["learning_rate"] = scheduler.get_last_lr()[0]
                             logging_loss = tr_loss
 
-                            if self.tb_writer:
-                                for k, v in logs.items():
-                                    self.tb_writer.add_scalar(k, v, global_step)
-                            if is_wandb_available():
-                                wandb.log(logs, step=global_step)
+                            self._log(logs)
 
-                            epoch_iterator.write(json.dumps({**logs, **{"step": global_step}}))
+                            if self.args.evaluate_during_training:
+                                self.evaluate()
 
-                        if self.args.save_steps > 0 and global_step % self.args.save_steps == 0:
+                        if self.args.save_steps > 0 and self.global_step % self.args.save_steps == 0:
                             # In all cases (even distributed/parallel), self.model is always a reference
                             # to the model we want to save.
                             if hasattr(model, "module"):
@@ -443,7 +456,9 @@ class Trainer:
                             else:
                                 assert model is self.model
                             # Save model checkpoint
-                            output_dir = os.path.join(self.args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{global_step}")
+                            output_dir = os.path.join(
+                                self.args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}"
+                            )
 
                             self.save_model(output_dir)
                             self._rotate_checkpoints()
@@ -451,10 +466,10 @@ class Trainer:
                             torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
                             logger.info("Saving optimizer and scheduler states to %s", output_dir)
 
-                if self.args.max_steps > 0 and global_step > self.args.max_steps:
+                if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
                     epoch_iterator.close()
                     break
-            if self.args.max_steps > 0 and global_step > self.args.max_steps:
+            if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
                 train_iterator.close()
                 break
             if self.args.tpu_metrics_debug:
@@ -465,7 +480,21 @@ class Trainer:
             self.tb_writer.close()
 
         logger.info("\n\nTraining completed. Do not forget to share your model on huggingface.co/models =)\n\n")
-        return TrainOutput(global_step, tr_loss / global_step)
+        return TrainOutput(self.global_step, tr_loss / self.global_step)
+
+    def _log(self, logs: Dict[str, float], iterator: Optional[tqdm] = None) -> None:
+        if self.epoch is not None:
+            logs["epoch"] = self.epoch
+        if self.tb_writer:
+            for k, v in logs.items():
+                self.tb_writer.add_scalar(k, v, self.global_step)
+        if is_wandb_available():
+            wandb.log(logs, step=self.global_step)
+        output = json.dumps({**logs, **{"step": self.global_step}})
+        if iterator is not None:
+            iterator.write(output)
+        else:
+            print(output)
 
     def _training_step(
         self, model: nn.Module, inputs: Dict[str, torch.Tensor], optimizer: torch.optim.Optimizer
@@ -582,6 +611,8 @@ class Trainer:
 
         output = self._prediction_loop(eval_dataloader, description="Evaluation")
 
+        self._log(output.metrics)
+
         if self.args.tpu_metrics_debug:
             # tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
             xm.master_print(met.metrics_report())
@@ -663,6 +694,11 @@ class Trainer:
         else:
             metrics = {}
         if len(eval_losses) > 0:
-            metrics["loss"] = np.mean(eval_losses)
+            metrics["eval_loss"] = np.mean(eval_losses)
+
+        # Prefix all keys with eval_
+        for key in list(metrics.keys()):
+            if not key.startswith("eval_"):
+                metrics[f"eval_{key}"] = metrics.pop(key)
 
         return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics)

src/transformers/trainer_tf.py

@@ -125,7 +125,9 @@ class TFTrainer:
         in the Tensorflow documentation and those contained in the transformers library.
         """
         if self.args.optimizer_name == "adamw":
-            self.optimizer = create_optimizer(self.args.learning_rate, self.train_steps, self.args.warmup_steps)
+            self.optimizer = create_optimizer(
+                self.args.learning_rate, self.train_steps, self.args.warmup_steps, self.args.end_lr
+            )
         else:
             try:
                 self.optimizer = tf.keras.optimizers.get(
@@ -139,6 +141,7 @@ class TFTrainer:
                 self.optimizer = tf.keras.optimizers.get(
                     {"class_name": self.args.optimizer_name, "config": {"learning_rate": self.args.learning_rate}}
                 )
+        logger.info("Created an/a {} optimizer".format(self.optimizer))
 
     def _create_checkpoint_manager(self, max_to_keep: int = 5, load_model: bool = True) -> None:
         """
@@ -149,6 +152,7 @@ class TFTrainer:
           load_model: if we want to start the training from the latest checkpoint.
         """
         ckpt = tf.train.Checkpoint(optimizer=self.optimizer, model=self.model)
+
         self.model.ckpt_manager = tf.train.CheckpointManager(ckpt, PREFIX_CHECKPOINT_DIR, max_to_keep=max_to_keep)
 
         if load_model:
@@ -222,7 +226,11 @@ class TFTrainer:
         else:
             metrics = {}
 
-        metrics["loss"] = loss.numpy()
+        metrics["eval_loss"] = loss.numpy()
+
+        for key in list(metrics.keys()):
+            if not key.startswith("eval_"):
+                metrics[f"eval_{key}"] = metrics.pop(key)
 
         return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics)
 
@@ -329,7 +337,7 @@ class TFTrainer:
         gradients = [(tf.clip_by_value(grad, -self.args.max_grad_norm, self.args.max_grad_norm)) for grad in gradients]
         vars = self.model.trainable_variables
 
-        if self.args.mode == "token-classification":
+        if self.args.mode in ["token-classification", "question-answering"]:
             vars = [var for var in self.model.trainable_variables if "pooler" not in var.name]
 
         self.optimizer.apply_gradients(list(zip(gradients, vars)))
@@ -369,7 +377,7 @@ class TFTrainer:
         per_example_loss, _ = self._run_model(features, labels, True)
         vars = self.model.trainable_variables
 
-        if self.args.mode == "token-classification":
+        if self.args.mode in ["token-classification", "question-answering"]:
             vars = [var for var in self.model.trainable_variables if "pooler" not in var.name]
 
         gradients = self.optimizer.get_gradients(per_example_loss, vars)
@@ -386,7 +394,7 @@ class TFTrainer:
           labels: the batched labels.
           training: run the model in training mode or not
         """
-        if self.args.mode == "sequence-classification" or self.args.mode == "token-classification":
+        if self.args.mode == "text-classification" or self.args.mode == "token-classification":
             logits = self.model(features, training=training)[0]
         else:
             logits = self.model(features, training=training)
@@ -396,6 +404,10 @@ class TFTrainer:
             reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, shape_list(logits)[2])), active_loss)
             labels = tf.boolean_mask(tf.reshape(labels, (-1,)), active_loss)
             loss = self.loss(labels, reduced_logits)
+        elif self.args.mode == "question-answering":
+            start_loss = self.loss(labels["start_position"], logits[0])
+            end_loss = self.loss(labels["end_position"], logits[1])
+            loss = (start_loss + end_loss) / 2.0
         else:
             loss = self.loss(labels, logits)
 
@@ -425,5 +437,6 @@ class TFTrainer:
 
         path = os.path.join(self.args.output_dir, "saved_model")
 
+        logger.info("Saving model in {}".format(path))
         os.makedirs(path, exist_ok=True)
         self.model.save_pretrained(self.args.output_dir)

src/transformers/training_args_tf.py

@@ -21,8 +21,8 @@ class TFTrainingArguments(TrainingArguments):
         },
     )
     mode: str = field(
-        default="sequence-classification",
-        metadata={"help": 'Type of task, one of "sequence-classification", "token-classification" '},
+        default="text-classification",
+        metadata={"help": 'Type of task, one of "text-classification", "token-classification", "question-answering"'},
     )
     loss_name: str = field(
         default="SparseCategoricalCrossentropy",
@@ -30,6 +30,12 @@ class TFTrainingArguments(TrainingArguments):
             "help": "Name of a Tensorflow loss. For the list see: https://www.tensorflow.org/api_docs/python/tf/keras/losses"
         },
     )
+    tpu_name: str = field(
+        default=None, metadata={"help": "Name of TPU"},
+    )
+    end_lr: float = field(
+        default=0, metadata={"help": "End learning rate for optimizer"},
+    )
     eval_steps: int = field(default=1000, metadata={"help": "Run an evaluation every X steps."})
     debug: bool = field(
         default=False, metadata={"help": "Activate the trace to record computation graphs and profiling information"}
@@ -45,7 +51,10 @@ class TFTrainingArguments(TrainingArguments):
             strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
         else:
             try:
-                tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
+                if self.tpu_name:
+                    tpu = tf.distribute.cluster_resolver.TPUClusterResolver(self.tpu_name)
+                else:
+                    tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
             except ValueError:
                 tpu = None
 

tests/test_modeling_bart.py

@@ -690,4 +690,8 @@ class TestSinusoidalPositionalEmbeddings(unittest.TestCase):
         # test that forward pass is just a lookup, there is no ignore padding logic
         input_ids = torch.tensor([[4, 10, pad, pad, pad]], dtype=torch.long, device=torch_device)
         no_cache_pad_zero = emb1(input_ids)
-        self.assertTrue(torch.allclose(torch.Tensor(self.desired_weights), no_cache_pad_zero[:3, :5], atol=1e-3))
+        self.assertTrue(
+            torch.allclose(
+                torch.tensor(self.desired_weights, device=torch_device), no_cache_pad_zero[:3, :5], atol=1e-3
+            )
+        )

tests/test_modeling_marian.py

@@ -33,15 +33,21 @@ if is_torch_available():
         MarianTokenizer,
         MarianMTModel,
     )
+    from transformers.convert_marian_to_pytorch import (
+        convert_hf_name_to_opus_name,
+        convert_opus_name_to_hf_name,
+        ORG_NAME,
+    )
 
 
 class ModelManagementTests(unittest.TestCase):
     @slow
-    def test_model_count(self):
+    def test_model_names(self):
         model_list = HfApi().model_list()
-        expected_num_models = 1011
-        actual_num_models = len([x for x in model_list if x.modelId.startswith("Helsinki-NLP")])
-        self.assertEqual(expected_num_models, actual_num_models)
+        model_ids = [x.modelId for x in model_list if x.modelId.startswith(ORG_NAME)]
+        bad_model_ids = [mid for mid in model_ids if "+" in model_ids]
+        self.assertListEqual([], bad_model_ids)
+        self.assertGreater(len(model_ids), 500)
 
 
 @require_torch
@@ -91,12 +97,12 @@ class MarianIntegrationTest(unittest.TestCase):
         self.assertListEqual(self.expected_text, generated_words)
 
     def translate_src_text(self, **tokenizer_kwargs):
-        model_inputs: dict = self.tokenizer.prepare_translation_batch(src_texts=self.src_text, **tokenizer_kwargs).to(
+        model_inputs = self.tokenizer.prepare_translation_batch(src_texts=self.src_text, **tokenizer_kwargs).to(
             torch_device
         )
-        self.assertEqual(self.model.device, model_inputs["input_ids"].device)
+        self.assertEqual(self.model.device, model_inputs.input_ids.device)
         generated_ids = self.model.generate(
-            model_inputs["input_ids"], attention_mask=model_inputs["attention_mask"], num_beams=2
+            model_inputs.input_ids, attention_mask=model_inputs.attention_mask, num_beams=2
         )
         generated_words = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
         return generated_words
@@ -106,10 +112,10 @@ class TestMarian_EN_DE_More(MarianIntegrationTest):
     @slow
     def test_forward(self):
         src, tgt = ["I am a small frog"], ["Ich bin ein kleiner Frosch."]
-        expected = [38, 121, 14, 697, 38848, 0]
+        expected_ids = [38, 121, 14, 697, 38848, 0]
 
         model_inputs: dict = self.tokenizer.prepare_translation_batch(src, tgt_texts=tgt).to(torch_device)
-        self.assertListEqual(expected, model_inputs["input_ids"][0].tolist())
+        self.assertListEqual(expected_ids, model_inputs.input_ids[0].tolist())
 
         desired_keys = {
             "input_ids",
@@ -125,20 +131,19 @@ class TestMarian_EN_DE_More(MarianIntegrationTest):
 
     def test_tokenizer_equivalence(self):
         batch = self.tokenizer.prepare_translation_batch(["I am a small frog"]).to(torch_device)
-        input_ids = batch["input_ids"][0]
         expected = [38, 121, 14, 697, 38848, 0]
-        self.assertListEqual(expected, input_ids.tolist())
+        self.assertListEqual(expected, batch.input_ids[0].tolist())
 
     def test_unk_support(self):
         t = self.tokenizer
-        ids = t.prepare_translation_batch(["||"]).to(torch_device)["input_ids"][0].tolist()
+        ids = t.prepare_translation_batch(["||"]).to(torch_device).input_ids[0].tolist()
         expected = [t.unk_token_id, t.unk_token_id, t.eos_token_id]
         self.assertEqual(expected, ids)
 
     def test_pad_not_split(self):
-        input_ids_w_pad = self.tokenizer.prepare_translation_batch(["I am a small frog <pad>"])["input_ids"][0]
+        input_ids_w_pad = self.tokenizer.prepare_translation_batch(["I am a small frog <pad>"]).input_ids[0].tolist()
         expected_w_pad = [38, 121, 14, 697, 38848, self.tokenizer.pad_token_id, 0]  # pad
-        self.assertListEqual(expected_w_pad, input_ids_w_pad.tolist())
+        self.assertListEqual(expected_w_pad, input_ids_w_pad)
 
     @slow
     def test_batch_generation_en_de(self):
@@ -187,9 +192,8 @@ class TestMarian_RU_FR(MarianIntegrationTest):
     src = "ru"
     tgt = "fr"
     src_text = ["Он показал мне рукопись своей новой пьесы."]
-    expected_text = ["Il me montre un manuscrit de sa nouvelle pièce."]
+    expected_text = ["Il m'a montré le manuscrit de sa nouvelle pièce."]
 
-    @slow
     def test_batch_generation_ru_fr(self):
         self._assert_generated_batch_equal_expected()
 
@@ -197,36 +201,59 @@ class TestMarian_RU_FR(MarianIntegrationTest):
 class TestMarian_MT_EN(MarianIntegrationTest):
     src = "mt"
     tgt = "en"
-    src_text = ["Il - Babiloniżi b'mod żbaljat ikkonkludew li l - Alla l - veru kien dgħajjef."]
-    expected_text = ["The Babylonians wrongly concluded that the true God was weak."]
+    src_text = ["Billi messu b'mod ġentili, Ġesù fejjaq raġel li kien milqut bil - marda kerha tal - ġdiem."]
+    expected_text = ["Touching gently, Jesus healed a man who was affected by the sad disease of leprosy."]
 
-    @unittest.skip("")  # Known Issue: This model generates a string of .... at the end of the translation.
     def test_batch_generation_mt_en(self):
         self._assert_generated_batch_equal_expected()
 
 
-class TestMarian_DE_Multi(MarianIntegrationTest):
-    src = "de"
-    tgt = "ch_group"
-    src_text = ["Er aber sprach: Das ist die Gottlosigkeit."]
+class TestMarian_en_ROMANCE(MarianIntegrationTest):
+    """Multilingual on target side."""
+
+    src = "en"
+    tgt = "ROMANCE"
+    src_text = [
+        ">>fr<< Don't spend so much time watching TV.",
+        ">>pt<< Your message has been sent.",
+        ">>es<< He's two years older than me.",
+    ]
+    expected_text = [
+        "Ne passez pas autant de temps à regarder la télé.",
+        "A sua mensagem foi enviada.",
+        "Es dos años más viejo que yo.",
+    ]
 
     @slow
-    def test_translation_de_multi_does_not_error(self):
-        self.translate_src_text()
-
-    @unittest.skip("")  # "Language codes are not yet supported."
-    def test_batch_generation_de_multi_tgt(self):
+    def test_batch_generation_en_ROMANCE_multi(self):
         self._assert_generated_batch_equal_expected()
 
-    @unittest.skip("")  # "Language codes are not yet supported."
-    def test_lang_code(self):
-        t = "Er aber sprach"
-        zh_code = self.code
-        tok_fn = self.tokenizer.prepare_translation_batch
-        pass_code = tok_fn(src_texts=[t], tgt_lang_code=zh_code)["input_ids"][0]
-        preprocess_with_code = tok_fn(src_texts=[zh_code + "  " + t])["input_ids"][0]
-        self.assertListEqual(pass_code.tolist(), preprocess_with_code.tolist())
-        for code in self.tokenizer.supported_language_codes:
-            self.assertIn(code, self.tokenizer.encoder)
-        pass_only_code = tok_fn(src_texts=[""], tgt_lang_code=zh_code)["input_ids"][0].tolist()
-        self.assertListEqual(pass_only_code, [self.tokenizer.encoder[zh_code], self.tokenizer.eos_token_id])
+    def test_tokenizer_handles_empty(self):
+        normalized = self.tokenizer.normalize("")
+        self.assertIsInstance(normalized, str)
+        with self.assertRaises(ValueError):
+            self.tokenizer.prepare_translation_batch([""])
+
+
+@require_torch
+class TestConversionUtils(unittest.TestCase):
+    def test_renaming_multilingual(self):
+        old_names = [
+            "opus-mt-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-fi",
+            "opus-mt-cmn+cn-fi",  # no group
+            "opus-mt-en-de",  # standard name
+            "opus-mt-en-de",  # standard name
+        ]
+        expected = ["opus-mt-ZH-fi", "opus-mt-cmn_cn-fi", "opus-mt-en-de", "opus-mt-en-de"]
+        self.assertListEqual(expected, [convert_opus_name_to_hf_name(x) for x in old_names])
+
+    def test_undoing_renaming(self):
+        hf_names = ["opus-mt-ZH-fi", "opus-mt-cmn_cn-fi", "opus-mt-en-de", "opus-mt-en-de"]
+        converted_opus_names = [convert_hf_name_to_opus_name(x) for x in hf_names]
+        expected_opus_names = [
+            "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-fi",
+            "cmn+cn-fi",
+            "en-de",  # standard name
+            "en-de",
+        ]
+        self.assertListEqual(expected_opus_names, converted_opus_names)

tests/test_trainer.py

@@ -98,7 +98,7 @@ class TrainerIntegrationTest(unittest.TestCase):
         training_args = TrainingArguments(output_dir="./examples", no_cuda=True)
         trainer = Trainer(model=model, args=training_args, eval_dataset=eval_dataset)
         result = trainer.evaluate()
-        self.assertLess(result["loss"], 0.2)
+        self.assertLess(result["eval_loss"], 0.2)
 
     def test_trainer_eval_lm(self):
         MODEL_ID = "distilroberta-base"