adding positional embeds masking to TFRoBERTa

closes #1960 Add saving and resuming functionality for remaining examples

.circleci/config.yml

@@ -1,9 +1,11 @@
 version: 2
 jobs:
-    build_py3_torch_and_tf:
+    run_tests_py3_torch_and_tf:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.5
+        environment:
+            OMP_NUM_THREADS: 1
         resource_class: xlarge
         parallelism: 1
         steps:
@@ -11,49 +13,67 @@ jobs:
             - run: sudo pip install torch
             - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
-            - run: sudo pip install pytest codecov pytest-cov
+            - run: sudo pip install pytest codecov pytest-cov pytest-xdist
             - run: sudo pip install tensorboardX scikit-learn
-            - run: python -m pytest -sv ./transformers/tests/ --cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./transformers/tests/ --cov
             - run: codecov
-    build_py3_torch:
+    run_tests_py3_torch:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.5
+        environment:
+            OMP_NUM_THREADS: 1
         resource_class: xlarge
         parallelism: 1
         steps:
             - checkout
             - run: sudo pip install torch
             - run: sudo pip install --progress-bar off .
-            - run: sudo pip install pytest codecov pytest-cov
+            - run: sudo pip install pytest codecov pytest-cov pytest-xdist
             - run: sudo pip install tensorboardX scikit-learn
-            - run: python -m pytest -sv ./transformers/tests/ --cov
-            - run: python -m pytest -sv ./examples/
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./transformers/tests/ --cov
             - run: codecov
-    build_py3_tf:
+    run_tests_py3_tf:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.5
+        environment:
+            OMP_NUM_THREADS: 1
         resource_class: xlarge
         parallelism: 1
         steps:
             - checkout
             - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
-            - run: sudo pip install pytest codecov pytest-cov
+            - run: sudo pip install pytest codecov pytest-cov pytest-xdist
             - run: sudo pip install tensorboardX scikit-learn
-            - run: python -m pytest -sv ./transformers/tests/ --cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./transformers/tests/ --cov
             - run: codecov
-    build_py3_custom_tokenizers:
+    run_tests_py3_custom_tokenizers:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.5
         steps:
             - checkout
             - run: sudo pip install --progress-bar off .
-            - run: sudo pip install pytest
+            - run: sudo pip install pytest pytest-xdist
             - run: sudo pip install mecab-python3
             - run: RUN_CUSTOM_TOKENIZERS=1 python -m pytest -sv ./transformers/tests/tokenization_bert_japanese_test.py
+    run_examples_py3_torch:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.5
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - run: sudo pip install torch
+            - run: sudo pip install --progress-bar off .
+            - run: sudo pip install pytest pytest-xdist
+            - run: sudo pip install tensorboardX scikit-learn
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./examples/
     deploy_doc:
         working_directory: ~/transformers
         docker:
@@ -66,7 +86,7 @@ jobs:
             - run: sudo pip install --progress-bar off -r docs/requirements.txt
             - run: sudo pip install --progress-bar off -r requirements.txt
             - run: ./.circleci/deploy.sh
-    repository_consistency:
+    check_repository_consistency:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.5
@@ -85,9 +105,10 @@ workflows:
     version: 2
     build_and_test:
         jobs:
-            - repository_consistency
-            - build_py3_custom_tokenizers
-            - build_py3_torch_and_tf
-            - build_py3_torch
-            - build_py3_tf
+            - check_repository_consistency
+            - run_examples_py3_torch
+            - run_tests_py3_custom_tokenizers
+            - run_tests_py3_torch_and_tf
+            - run_tests_py3_torch
+            - run_tests_py3_tf
             - deploy_doc: *workflow_filters

README.md

@@ -133,7 +133,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 
 ## Model architectures
 
-🤗 Transformers currently provides 10 NLU/NLG architectures:
+🤗 Transformers currently provides the following NLU/NLG architectures:
 
 1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.

docs/source/pretrained_models.rst

@@ -3,6 +3,7 @@ Pretrained models
 
 Here is the full list of the currently provided pretrained models together with a short presentation of each model.
 
+For a list that includes community-uploaded models, refer to `https://huggingface.co/models <https://huggingface.co/models>`__.
 
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | Architecture      | Shortcut name                                              | Details of the model                                                                                                                  |
@@ -154,6 +155,10 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
+|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``roberta-base-openai-detector``                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
 |                   |                                                            | | ``roberta-base`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                             |
 |                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
@@ -174,10 +179,6 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint.                                                               |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
-|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
-|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
-|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``distilbert-base-german-cased``                           | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
 |                   |                                                            | | The German DistilBERT model distilled from the German DBMDZ BERT model `bert-base-german-dbmdz-cased` checkpoint.                   |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |

examples/README.md

@@ -43,7 +43,7 @@ Quick benchmarks from the script (no other modifications):
 | Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
 | V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
 | V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
-| 1080 Ti | FP32 | 55s | - | 
+| 1080 Ti | FP32 | 55s | - |
 
 Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
 
@@ -357,9 +357,9 @@ eval_loss = 0.44457291918821606
 
 Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py).
 
-#### Fine-tuning on SQuAD
+#### Fine-tuning BERT on SQuAD1.0
 
-This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) 
+This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) 
 on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a 
 $SQUAD_DIR directory.
 
@@ -367,6 +367,12 @@ $SQUAD_DIR directory.
 * [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
 * [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
 
+And for SQuAD2.0, you need to download:
+
+- [train-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json)
+- [dev-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json)
+- [evaluate-v2.0.py](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/)
+
 ```bash
 export SQUAD_DIR=/path/to/SQUAD
 
@@ -396,7 +402,7 @@ exact_match = 81.22
 #### Distributed training
 
 
-Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:
+Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.0:
 
 ```bash
 python -m torch.distributed.launch --nproc_per_node=8 run_squad.py \
@@ -428,7 +434,9 @@ This fine-tuned model is available as a checkpoint under the reference
 
 #### Fine-tuning XLNet on SQuAD
 
-This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
+This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .
+
+##### Command for SQuAD1.0:
 
 ```bash
 export SQUAD_DIR=/path/to/SQUAD
@@ -451,7 +459,32 @@ python /data/home/hlu/transformers/examples/run_squad.py \
     --save_steps 5000
 ```
 
-Training with the previously defined hyper-parameters yields the following results:
+##### Command for SQuAD2.0:
+
+```bash
+export SQUAD_DIR=/path/to/SQUAD
+
+python run_squad.py \
+    --model_type xlnet \
+    --model_name_or_path xlnet-large-cased \
+    --do_train \
+    --do_eval \
+    --version_2_with_negative \
+    --train_file $SQUAD_DIR/train-v2.0.json \
+    --predict_file $SQUAD_DIR/dev-v2.0.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 4 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./wwm_cased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=2  \
+    --per_gpu_train_batch_size=2   \
+    --save_steps 5000
+```
+
+Larger batch size may improve the performance while costing more memory.
+
+##### Results for SQuAD1.0 with the previously defined hyper-parameters:
 
 ```python
 {
@@ -464,6 +497,24 @@ Training with the previously defined hyper-parameters yields the following resul
 }
 ```
 
+##### Results for SQuAD2.0 with the previously defined hyper-parameters:
+
+```python
+{
+"exact": 80.4177545691906,
+"f1": 84.07154997729623,
+"total": 11873,
+"HasAns_exact": 76.73751686909581,
+"HasAns_f1": 84.05558584352873,
+"HasAns_total": 5928,
+"NoAns_exact": 84.0874684608915,
+"NoAns_f1": 84.0874684608915,
+"NoAns_total": 5945
+}
+```
+
+
+
 ## Named Entity Recognition
 
 Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and

examples/contrib/run_openai_gpt.py

@@ -75,7 +75,7 @@ def pre_process_datasets(encoded_datasets, input_len, cap_length, start_token, d
         n_batch = len(dataset)
         input_ids = np.zeros((n_batch, 2, input_len), dtype=np.int64)
         mc_token_ids = np.zeros((n_batch, 2), dtype=np.int64)
-        lm_labels = np.full((n_batch, 2, input_len), fill_value=-1, dtype=np.int64)
+        lm_labels = np.full((n_batch, 2, input_len), fill_value=-100, dtype=np.int64)
         mc_labels = np.zeros((n_batch,), dtype=np.int64)
         for i, (story, cont1, cont2, mc_label), in enumerate(dataset):
             with_cont1 = [start_token] + story[:cap_length] + [delimiter_token] + cont1[:cap_length] + [clf_token]

examples/distillation/distiller.py

@@ -112,7 +112,7 @@ class Distiller:
         self.last_log = 0
 
         self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
-        self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+        self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
         if self.alpha_mse > 0.:
             self.mse_loss_fct = nn.MSELoss(reduction='sum')
         if self.alpha_cos > 0.:
@@ -186,7 +186,7 @@ class Distiller:
         -------
             token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
             attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
-            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -1 where there is nothing to predict.
+            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -100 where there is nothing to predict.
         """
         token_ids, lengths = batch
         token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -224,7 +224,7 @@ class Distiller:
         _token_ids = _token_ids_mask * (probs == 0).long() + _token_ids_real * (probs == 1).long() + _token_ids_rand * (probs == 2).long()
         token_ids = token_ids.masked_scatter(pred_mask, _token_ids)
 
-        mlm_labels[~pred_mask] = -1 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
+        mlm_labels[~pred_mask] = -100 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
 
         # sanity checks
         assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
@@ -246,7 +246,7 @@ class Distiller:
         -------
             token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
             attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
-            clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -1 where there is nothing to predict.
+            clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -100 where there is nothing to predict.
         """
         token_ids, lengths = batch
         token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -254,7 +254,7 @@ class Distiller:
 
         attn_mask = (torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None])
         clm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
-        clm_labels[~attn_mask] = -1 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
+        clm_labels[~attn_mask] = -100 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
 
         # sanity checks
         assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size

examples/run_generation.py

@@ -20,14 +20,10 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import argparse
 import logging
-from tqdm import trange
 
 import torch
-import torch.nn.functional as F
 import numpy as np
 
-from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
-
 from transformers import GPT2LMHeadModel, GPT2Tokenizer
 from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
 from transformers import XLNetLMHeadModel, XLNetTokenizer
@@ -36,22 +32,22 @@ from transformers import CTRLLMHeadModel, CTRLTokenizer
 from transformers import XLMWithLMHeadModel, XLMTokenizer
 
 
-logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                    datefmt = '%m/%d/%Y %H:%M:%S',
-                    level = logging.INFO)
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+    datefmt="%m/%d/%Y %H:%M:%S",
+    level=logging.INFO,
+)
 logger = logging.getLogger(__name__)
 
 MAX_LENGTH = int(10000)  # Hardcoded max length to avoid infinite loop
 
-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
-
 MODEL_CLASSES = {
-    'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
-    'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
-    'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
-    'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
-    'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
-    'xlm': (XLMWithLMHeadModel, XLMTokenizer),
+    "gpt2": (GPT2LMHeadModel, GPT2Tokenizer),
+    "ctrl": (CTRLLMHeadModel, CTRLTokenizer),
+    "openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
+    "xlnet": (XLNetLMHeadModel, XLNetTokenizer),
+    "transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer),
+    "xlm": (XLMWithLMHeadModel, XLMTokenizer),
 }
 
 # Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
@@ -75,81 +71,79 @@ def set_seed(args):
     if args.n_gpu > 0:
         torch.cuda.manual_seed_all(args.seed)
 
-
-def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
-    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
-        Args:
-            logits: logits distribution shape (batch size x vocabulary size)
-            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
-            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
-                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
-        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
-    """
-    top_k = min(top_k, logits.size(-1))  # Safety check
-    if top_k > 0:
-        # Remove all tokens with a probability less than the last token of the top-k
-        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
-        logits[indices_to_remove] = filter_value
-
-    if top_p > 0.0:
-        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
-        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
-
-        # Remove tokens with cumulative probability above the threshold
-        sorted_indices_to_remove = cumulative_probs > top_p
-        # Shift the indices to the right to keep also the first token above the threshold
-        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
-        sorted_indices_to_remove[..., 0] = 0
-
-        # scatter sorted tensors to original indexing
-        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
-        logits[indices_to_remove] = filter_value
-    return logits
+#
+# Functions to prepare models' input
+#
 
 
-def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
-                    is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
-    context = torch.tensor(context, dtype=torch.long, device=device)
-    context = context.unsqueeze(0).repeat(num_samples, 1)
-    generated = context
-    with torch.no_grad():
-        for _ in trange(length):
+def prepare_ctrl_input(args, _, tokenizer, prompt_text):
+    if args.temperature > 0.7:
+        logger.info(
+            "CTRL typically works better with lower temperatures (and lower top_k)."
+        )
 
-            inputs = {'input_ids': generated}
-            if is_xlnet: 
-                # XLNet is a direct (predict same token, not next token) and bi-directional model by default
-                # => need one additional dummy token in the input (will be masked), attention mask and target mapping (see model docstring)
-                input_ids = torch.cat((generated, torch.zeros((1, 1), dtype=torch.long, device=device)), dim=1)
-                perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float, device=device)
-                perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
-                target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float, device=device)
-                target_mapping[0, 0, -1] = 1.0  # predict last token
-                inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)
+    if not any(encoded_prompt[0] == x for x in tokenizer.control_codes.values()):
+        logger.info(
+            "WARNING! You are not starting your generation from a control code so you won't get good results"
+        )
+    return prompt_text
 
-            if is_xlm_mlm and xlm_mask_token:
-                # XLM MLM models are direct models (predict same token, not next token)
-                # => need one additional dummy token in the input (will be masked and guessed)
-                input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
-                inputs = {'input_ids': input_ids}
 
-            if xlm_lang is not None:
-                inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
+def prepare_xlm_input(args, model, tokenizer, prompt_text):
+    # kwargs = {"language": None, "mask_token_id": None}
 
-            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
-            next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
+    # Set the language
+    use_lang_emb = hasattr(model.config, "use_lang_emb") and model.config.use_lang_emb
+    if hasattr(model.config, "lang2id") and use_lang_emb:
+        available_languages = model.config.lang2id.keys()
+        if args.xlm_language in available_languages:
+            language = args.xlm_language
+        else:
+            language = None
+            while language not in available_languages:
+                language = input(
+                    "Using XLM. Select language in "
+                    + str(list(available_languages))
+                    + " >>> "
+                )
+        # kwargs["language"] = tokenizer.lang2id[language]
 
-            # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
-            for i in range(num_samples):
-                for _ in set(generated[i].tolist()):
-                    next_token_logits[i, _] /= repetition_penalty
-                
-            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
-            if temperature == 0: # greedy sampling:
-                next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
-            else:
-                next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
-            generated = torch.cat((generated, next_token), dim=1)
-    return generated
+    # TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
+    # XLM masked-language modeling (MLM) models need masked token
+    # is_xlm_mlm = "mlm" in args.model_name_or_path
+    # if is_xlm_mlm:
+    #     kwargs["mask_token_id"] = tokenizer.mask_token_id
+
+    return prompt_text
+
+
+def prepare_xlnet_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text, {}
+
+
+def prepare_transfoxl_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text, {}
+
+
+PREPROCESSING_FUNCTIONS = {
+    "ctrl": prepare_ctrl_input,
+    "xlm": prepare_xlm_input,
+    "xlnet": prepare_xlnet_input,
+    "transfo-xl": prepare_transfoxl_input,
+}
+
+
+def adjust_length_to_model(length, max_sequence_length):
+    if length < 0 and max_sequence_length > 0:
+        length = max_sequence_length
+    elif 0 < max_sequence_length < length:
+        length = max_sequence_length  # No generation bigger than model size
+    elif length < 0:
+        length = MAX_LENGTH  # avoid infinite loop
+    return length
 
 
 def main():
@@ -157,108 +151,76 @@ def main():
     parser.add_argument("--model_type", default=None, type=str, required=True,
                         help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
     parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
-                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+
     parser.add_argument("--prompt", type=str, default="")
-    parser.add_argument("--padding_text", type=str, default="")
-    parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
     parser.add_argument("--length", type=int, default=20)
-    parser.add_argument("--num_samples", type=int, default=1)
-    parser.add_argument("--temperature", type=float, default=1.0,
-                        help="temperature of 0 implies greedy sampling")
-    parser.add_argument("--repetition_penalty", type=float, default=1.0,
-                        help="primarily useful for CTRL model; in that case, use 1.2")
-    parser.add_argument("--top_k", type=int, default=0)
-    parser.add_argument("--top_p", type=float, default=0.9)
-    parser.add_argument("--no_cuda", action='store_true',
-                        help="Avoid using CUDA when available")
-    parser.add_argument('--seed', type=int, default=42,
-                        help="random seed for initialization")
-    parser.add_argument('--stop_token', type=str, default=None,
-                        help="Token at which text generation is stopped")
+    parser.add_argument("--stop_token", type=str, default=None, help="Token at which text generation is stopped")
+
+    parser.add_argument("--temperature", type=float, default=1.0, help="temperature of 1.0 has no effect, lower tend toward greedy sampling")
+    parser.add_argument("--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2")
+    parser.add_argument("--k", type=int, default=0)
+    parser.add_argument("--p", type=float, default=0.9)
+
+    parser.add_argument("--padding_text", type=str, default="", help="Padding text for Transfo-XL and XLNet.")
+    parser.add_argument("--xlm_language", type=str, default="", help="Optional language when used with the XLM model.")
+
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
     args = parser.parse_args()
 
-    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+    args.device = torch.device(
+        "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu"
+    )
     args.n_gpu = torch.cuda.device_count()
 
     set_seed(args)
 
-    args.model_type = args.model_type.lower()
-    model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    # Initialize the model and tokenizer
+    try:
+        args.model_type = args.model_type.lower()
+        model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    except KeyError:
+        raise KeyError(
+            "the model {} you specified is not supported. You are welcome to add it and open a PR :)"
+        )
+
     tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
     model = model_class.from_pretrained(args.model_name_or_path)
     model.to(args.device)
-    model.eval()
-
-    if args.length < 0 and model.config.max_position_embeddings > 0:
-        args.length = model.config.max_position_embeddings
-    elif 0 < model.config.max_position_embeddings < args.length:
-        args.length = model.config.max_position_embeddings  # No generation bigger than model size 
-    elif args.length < 0:
-        args.length = MAX_LENGTH  # avoid infinite loop
 
+    args.length = adjust_length_to_model(
+        args.length, max_sequence_length=model.config.max_position_embeddings
+    )
     logger.info(args)
-    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7:
-            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
 
-    while True:
-        xlm_lang = None
-        # XLM Language usage detailed in the issues #1414
-        if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
-                and model.config.use_lang_emb:
-            if args.xlm_lang:
-                language = args.xlm_lang
-            else:
-                language = None
-                while language not in tokenizer.lang2id.keys():
-                    language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
-            xlm_lang = tokenizer.lang2id[language]
+    prompt_text = args.prompt if args.prompt else input("Model prompt >>> ")
 
-        # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
-        is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
-        if is_xlm_mlm:
-            xlm_mask_token = tokenizer.mask_token_id
-        else:
-            xlm_mask_token = None
+    # Different models need different input formatting and/or extra arguments
+    requires_preprocessing = args.model_type in PREPROCESSING_FUNCTIONS.keys()
+    if requires_preprocessing:
+        prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
+        prompt_text = prepare_input(args, model, tokenizer, prompt_text)
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors='pt')
 
-        raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
-        if args.model_type in ["transfo-xl", "xlnet"]:
-            # Models with memory likes to have a long prompt for short inputs.
-            raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
-        context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
-        if args.model_type == "ctrl":
-            if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
-                logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
-        out = sample_sequence(
-            model=model,
-            context=context_tokens,
-            num_samples=args.num_samples,
-            length=args.length,
-            temperature=args.temperature,
-            top_k=args.top_k,
-            top_p=args.top_p,
-            repetition_penalty=args.repetition_penalty,
-            is_xlnet=bool(args.model_type == "xlnet"),
-            is_xlm_mlm=is_xlm_mlm,
-            xlm_mask_token=xlm_mask_token,
-            xlm_lang=xlm_lang,
-            device=args.device,
-        )
-        out = out[:, len(context_tokens):].tolist()
-        for o in out:
-            text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
-            if args.stop_token:
-                index =  text.find(args.stop_token)
-                if index == -1:
-                    index = None
-                text = text[:index]
+    output_sequences = model.generate(
+        input_ids=encoded_prompt,
+        max_length=args.length,
+        temperature=args.temperature,
+        top_k=args.k,
+        top_p=args.p,
+        repetition_penalty=args.repetition_penalty,
+    )
 
-            print(text)
+    # Batch size == 1. to add more examples please use num_return_sequences > 1 
+    generated_sequence = output_sequences[0].tolist()
+    text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
+    text = text[: t.find(args.stop_token) if args.stop_token else None]
+
+    print(text)
 
-        if args.prompt:
-            break
     return text
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

examples/run_glue.py

@@ -112,6 +112,13 @@ def train(args, train_dataset, model, tokenizer):
 
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -140,13 +147,33 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // 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("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {'input_ids':      batch[0],
@@ -205,9 +232,15 @@ def train(args, train_dataset, model, tokenizer):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break

examples/run_lm_finetuning.py

@@ -150,7 +150,7 @@ def mask_tokens(inputs, tokenizer, args):
     special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
     probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
     masked_indices = torch.bernoulli(probability_matrix).bool()
-    labels[~masked_indices] = -1  # We only compute loss on masked tokens
+    labels[~masked_indices] = -100  # We only compute loss on masked tokens
 
     # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
     indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices

examples/run_ner.py

@@ -88,6 +88,13 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
     ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -117,13 +124,33 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // 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("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {"input_ids": batch[0],
@@ -175,9 +202,15 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, "training_args.bin"))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break

examples/run_squad.py

@@ -27,7 +27,8 @@ import glob
 import timeit
 import numpy as np
 import torch
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)
+from torch.utils.data import (
+    DataLoader, RandomSampler, SequentialSampler, TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
 
 try:
@@ -39,6 +40,7 @@ from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,
+                                  RobertaForQuestionAnswering, RobertaTokenizer, RobertaConfig,
                                   XLMConfig, XLMForQuestionAnswering,
                                   XLMTokenizer, XLNetConfig,
                                   XLNetForQuestionAnswering,
@@ -53,16 +55,18 @@ from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_e
 logger = logging.getLogger(__name__)
 
 ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
-                  for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
+                  for conf in (BertConfig, RobertaConfig, XLNetConfig, XLMConfig)), ())
 
 MODEL_CLASSES = {
     'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
+    'roberta': (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
     'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
     'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
     'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
     'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
 }
 
+
 def set_seed(args):
     random.seed(args.seed)
     np.random.seed(args.seed)
@@ -70,40 +74,60 @@ def set_seed(args):
     if args.n_gpu > 0:
         torch.cuda.manual_seed_all(args.seed)
 
+
 def to_list(tensor):
     return tensor.detach().cpu().tolist()
 
+
 def train(args, train_dataset, model, tokenizer):
     """ Train the model """
     if args.local_rank in [-1, 0]:
         tb_writer = SummaryWriter()
 
     args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+    train_sampler = RandomSampler(
+        train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(
+        train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
 
     if args.max_steps > 0:
         t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+        args.num_train_epochs = args.max_steps // (
+            len(train_dataloader) // args.gradient_accumulation_steps) + 1
     else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+        t_total = len(
+            train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
 
     # Prepare optimizer and schedule (linear warmup and decay)
     no_decay = ['bias', 'LayerNorm.weight']
     optimizer_grouped_parameters = [
-        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
-        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+        {'params': [p for n, p in model.named_parameters() if not any(
+            nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+        {'params': [p for n, p in model.named_parameters() if any(
+            nd in n for nd in no_decay)], 'weight_decay': 0.0}
     ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    optimizer = AdamW(optimizer_grouped_parameters,
+                      lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(
+            os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(
+            os.path.join(args.model_name_or_path, 'scheduler.pt')))
 
     if args.fp16:
         try:
             from apex import amp
         except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+            raise ImportError(
+                "Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
+        model, optimizer = amp.initialize(
+            model, optimizer, opt_level=args.fp16_opt_level)
 
     # multi-gpu training (should be after apex fp16 initialization)
     if args.n_gpu > 1:
@@ -119,42 +143,72 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("***** Running training *****")
     logger.info("  Num examples = %d", len(train_dataset))
     logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Instantaneous batch size per GPU = %d",
+                args.per_gpu_train_batch_size)
     logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
-                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+                args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d",
+                args.gradient_accumulation_steps)
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 1
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) //
+                                         args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (
+            len(train_dataloader) // 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("  Will skip the first %d steps in the first epoch",
+                    steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
-    
+    train_iterator = trange(epochs_trained, int(
+        args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    # Added here for reproductibility (even between python 2 and 3)
+    set_seed(args)
+
     for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration",
+                              disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
 
             inputs = {
                 'input_ids':       batch[0],
                 'attention_mask':  batch[1],
+                'token_type_ids': None if args.model_type in ['xlm', 'roberta', 'distilbert'] else batch[2],
                 'start_positions': batch[3],
-                'end_positions':   batch[4]
+                'end_positions':   batch[4],
             }
 
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
-
             if args.model_type in ['xlnet', 'xlm']:
-                inputs.update({'cls_index': batch[5], 'p_mask': batch[6]})
-
+                inputs.update({'cls_index': batch[5],
+                               'p_mask':       batch[6]})
+                if args.version_2_with_negative:
+                    inputs.update({'is_impossible': batch[7]})
             outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+            # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]
 
             if args.n_gpu > 1:
-                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
+                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
             if args.gradient_accumulation_steps > 1:
                 loss = loss / args.gradient_accumulation_steps
 
@@ -167,9 +221,11 @@ def train(args, train_dataset, model, tokenizer):
             tr_loss += loss.item()
             if (step + 1) % args.gradient_accumulation_steps == 0:
                 if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                    torch.nn.utils.clip_grad_norm_(
+                        amp.master_params(optimizer), args.max_grad_norm)
                 else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+                    torch.nn.utils.clip_grad_norm_(
+                        model.parameters(), args.max_grad_norm)
 
                 optimizer.step()
                 scheduler.step()  # Update learning rate schedule
@@ -178,24 +234,41 @@ def train(args, train_dataset, model, tokenizer):
 
                 # Log metrics
                 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                    # Only evaluate when single GPU otherwise metrics may not average well
+                    if args.local_rank == -1 and args.evaluate_during_training:
                         results = evaluate(args, model, tokenizer)
                         for key, value in results.items():
-                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
-                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
-                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
+                            tb_writer.add_scalar(
+                                'eval_{}'.format(key), value, global_step)
+                    tb_writer.add_scalar(
+                        'lr', scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar(
+                        'loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
                     logging_loss = tr_loss
 
                 # Save model checkpoint
                 if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
+                    output_dir = os.path.join(
+                        args.output_dir, 'checkpoint-{}'.format(global_step))
                     if not os.path.exists(output_dir):
                         os.makedirs(output_dir)
-                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+                    # Take care of distributed/parallel training
+                    model_to_save = model.module if hasattr(
+                        model, 'module') else model
                     model_to_save.save_pretrained(output_dir)
-                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
+                    tokenizer.save_pretrained(output_dir)
+
+                    torch.save(args, os.path.join(
+                        output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(
+                        output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(
+                        output_dir, 'scheduler.pt'))
+                    logger.info(
+                        "Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break
@@ -210,7 +283,8 @@ def train(args, train_dataset, model, tokenizer):
 
 
 def evaluate(args, model, tokenizer, prefix=""):
-    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+    dataset, examples, features = load_and_cache_examples(
+        args, tokenizer, evaluate=True, output_examples=True)
 
     if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
         os.makedirs(args.output_dir)
@@ -219,7 +293,8 @@ def evaluate(args, model, tokenizer, prefix=""):
 
     # Note that DistributedSampler samples randomly
     eval_sampler = SequentialSampler(dataset)
-    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+    eval_dataloader = DataLoader(
+        dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
     # multi-gpu evaluate
     if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
@@ -240,14 +315,11 @@ def evaluate(args, model, tokenizer, prefix=""):
         with torch.no_grad():
             inputs = {
                 'input_ids':      batch[0],
-                'attention_mask': batch[1]
+                'attention_mask': batch[1],
+                'token_type_ids': None if args.model_type in ['xlm', 'roberta', 'distilbert'] else batch[2],
             }
-            
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
-
             example_indices = batch[3]
-            
+
             # XLNet and XLM use more arguments for their predictions
             if args.model_type in ['xlnet', 'xlm']:
                 inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
@@ -270,9 +342,9 @@ def evaluate(args, model, tokenizer, prefix=""):
                 cls_logits = output[4]
 
                 result = SquadResult(
-                    unique_id, start_logits, end_logits, 
-                    start_top_index=start_top_index, 
-                    end_top_index=end_top_index, 
+                    unique_id, start_logits, end_logits,
+                    start_top_index=start_top_index,
+                    end_top_index=end_top_index,
                     cls_logits=cls_logits
                 )
 
@@ -285,40 +357,48 @@ def evaluate(args, model, tokenizer, prefix=""):
             all_results.append(result)
 
     evalTime = timeit.default_timer() - start_time
-    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
+    logger.info("  Evaluation done in total %f secs (%f sec per example)",
+                evalTime, evalTime / len(dataset))
 
     # Compute predictions
-    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
-    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+    output_prediction_file = os.path.join(
+        args.output_dir, "predictions_{}.json".format(prefix))
+    output_nbest_file = os.path.join(
+        args.output_dir, "nbest_predictions_{}.json".format(prefix))
 
     if args.version_2_with_negative:
-        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
+        output_null_log_odds_file = os.path.join(
+            args.output_dir, "null_odds_{}.json".format(prefix))
     else:
         output_null_log_odds_file = None
 
     # XLNet and XLM use a more complex post-processing procedure
     if args.model_type in ['xlnet', 'xlm']:
-        start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
-        end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
+        start_n_top = model.config.start_n_top if hasattr(
+            model, "config") else model.module.config.start_n_top
+        end_n_top = model.config.end_n_top if hasattr(
+            model, "config") else model.module.config.end_n_top
 
         predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
-                        args.max_answer_length, output_prediction_file,
-                        output_nbest_file, output_null_log_odds_file,
-                        start_n_top, end_n_top,
-                        args.version_2_with_negative, tokenizer, args.verbose_logging)
+                                                    args.max_answer_length, output_prediction_file,
+                                                    output_nbest_file, output_null_log_odds_file,
+                                                    start_n_top, end_n_top,
+                                                    args.version_2_with_negative, tokenizer, args.verbose_logging)
     else:
         predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
                         args.max_answer_length, args.do_lower_case, output_prediction_file,
                         output_nbest_file, output_null_log_odds_file, args.verbose_logging,
-                        args.version_2_with_negative, args.null_score_diff_threshold)
+                        args.version_2_with_negative, args.null_score_diff_threshold, tokenizer)
 
     # Compute the F1 and exact scores.
     results = squad_evaluate(examples, predictions)
     return results
 
+
 def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
     if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
 
     # Load data features from cache or dataset file
     input_dir = args.data_dir if args.data_dir else "."
@@ -330,7 +410,8 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
 
     # Init features and dataset from cache if it exists
     if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
-        logger.info("Loading features from cached file %s", cached_features_file)
+        logger.info("Loading features from cached file %s",
+                    cached_features_file)
         features_and_dataset = torch.load(cached_features_file)
         features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
     else:
@@ -340,37 +421,43 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
             try:
                 import tensorflow_datasets as tfds
             except ImportError:
-                raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
+                raise ImportError(
+                    "If not data_dir is specified, tensorflow_datasets needs to be installed.")
 
             if args.version_2_with_negative:
-                logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")
+                logger.warn(
+                    "tensorflow_datasets does not handle version 2 of SQuAD.")
 
             tfds_examples = tfds.load("squad")
-            examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
+            examples = SquadV1Processor().get_examples_from_dataset(
+                tfds_examples, evaluate=evaluate)
         else:
             processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
-
             if evaluate:
                 examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
             else:
                 examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
 
-        features, dataset = squad_convert_examples_to_features( 
+        features, dataset = squad_convert_examples_to_features(
             examples=examples,
             tokenizer=tokenizer,
             max_seq_length=args.max_seq_length,
             doc_stride=args.doc_stride,
             max_query_length=args.max_query_length,
             is_training=not evaluate,
-            return_dataset='pt'
+            return_dataset='pt',
+            threads=args.threads,
         )
 
         if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save({"features": features, "dataset": dataset}, cached_features_file)
+            logger.info("Saving features into cached file %s",
+                        cached_features_file)
+            torch.save({"features": features, "dataset": dataset},
+                       cached_features_file)
 
     if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
 
     if output_examples:
         return dataset, examples, features
@@ -380,7 +467,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
 def main():
     parser = argparse.ArgumentParser()
 
-    ## Required parameters
+    # Required parameters
     parser.add_argument("--model_type", default=None, type=str, required=True,
                         help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
     parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
@@ -388,7 +475,7 @@ def main():
     parser.add_argument("--output_dir", default=None, type=str, required=True,
                         help="The output directory where the model checkpoints and predictions will be written.")
 
-    ## Other parameters
+    # Other parameters
     parser.add_argument("--data_dir", default=None, type=str,
                         help="The input data dir. Should contain the .json files for the task." +
                              "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
@@ -480,22 +567,27 @@ def main():
                              "See details at https://nvidia.github.io/apex/amp.html")
     parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
     parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+
+    parser.add_argument('--threads', type=int, default=1, help='multiple threads for converting example to features')
     args = parser.parse_args()
 
     if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
-        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
 
     # Setup distant debugging if needed
     if args.server_ip and args.server_port:
         # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
         import ptvsd
         print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.enable_attach(
+            address=(args.server_ip, args.server_port), redirect_output=True)
         ptvsd.wait_for_attach()
 
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        device = torch.device(
+            "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
         args.n_gpu = torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
@@ -505,18 +597,19 @@ def main():
     args.device = device
 
     # Setup logging
-    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                        datefmt = '%m/%d/%Y %H:%M:%S',
-                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                        datefmt='%m/%d/%Y %H:%M:%S',
+                        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
     logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+                   args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
 
     # Set seed
     set_seed(args)
 
     # Load pretrained model and tokenizer
     if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
 
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
@@ -526,12 +619,14 @@ def main():
                                                 do_lower_case=args.do_lower_case,
                                                 cache_dir=args.cache_dir if args.cache_dir else None)
     model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        from_tf=bool(
+                                            '.ckpt' in args.model_name_or_path),
                                         config=config,
                                         cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
 
     model.to(args.device)
 
@@ -545,14 +640,16 @@ def main():
             import apex
             apex.amp.register_half_function(torch, 'einsum')
         except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+            raise ImportError(
+                "Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
 
     # Training
     if args.do_train:
-        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        train_dataset = load_and_cache_examples(
+            args, tokenizer, evaluate=False, output_examples=False)
         global_step, tr_loss = train(args, train_dataset, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
+        logger.info(" global_step = %s, average loss = %s",
+                    global_step, tr_loss)
 
     # Save the trained model and the tokenizer
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
@@ -563,7 +660,8 @@ def main():
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`
-        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+        # Take care of distributed/parallel training
+        model_to_save = model.module if hasattr(model, 'module') else model
         model_to_save.save_pretrained(args.output_dir)
         tokenizer.save_pretrained(args.output_dir)
 
@@ -571,15 +669,15 @@ def main():
         torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir, force_download=True)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model = model_class.from_pretrained(
+            args.output_dir, force_download=True)
+        tokenizer = tokenizer_class.from_pretrained(
+            args.output_dir, do_lower_case=args.do_lower_case)
         model.to(args.device)
 
-
     # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
     results = {}
     if args.do_eval and args.local_rank in [-1, 0]:
-
         if args.do_train:
             logger.info("Loading checkpoints saved during training for evaluation")
             checkpoints = [args.output_dir]
@@ -594,14 +692,17 @@ def main():
 
         for checkpoint in checkpoints:
             # Reload the model
-            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            model = model_class.from_pretrained(checkpoint, force_download=True)
+            global_step = checkpoint.split(
+                '-')[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(
+                checkpoint, force_download=True)
             model.to(args.device)
 
             # Evaluate
             result = evaluate(args, model, tokenizer, prefix=global_step)
 
-            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
+            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v)
+                          for k, v in result.items())
             results.update(result)
 
     logger.info("Results: {}".format(results))

examples/run_xnli.py

@@ -92,6 +92,13 @@ def train(args, train_dataset, model, tokenizer):
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -120,13 +127,32 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // 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("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {'input_ids':      batch[0],
@@ -177,9 +203,15 @@ def train(args, train_dataset, model, tokenizer):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break

examples/utils_ner.py

@@ -94,7 +94,7 @@ def convert_examples_to_features(examples,
                                  pad_on_left=False,
                                  pad_token=0,
                                  pad_token_segment_id=0,
-                                 pad_token_label_id=-1,
+                                 pad_token_label_id=-100,
                                  sequence_a_segment_id=0,
                                  mask_padding_with_zero=True):
     """ Loads a data file into a list of `InputBatch`s

setup.py

@@ -59,6 +59,7 @@ setup(
                                     "tests.*", "tests"]),
     install_requires=['numpy',
                       'boto3',
+                      'filelock',
                       'requests',
                       'tqdm',
                       'regex != 2019.12.17',

templates/adding_a_new_model/modeling_xxx.py

@@ -364,7 +364,7 @@ class XxxForMaskedLM(XxxPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -415,7 +415,7 @@ class XxxForMaskedLM(XxxPreTrainedModel):
 
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 

templates/adding_a_new_model/tests/modeling_tf_xxx_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import XxxConfig, is_tf_available
 
@@ -245,10 +244,8 @@ class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in ['xxx-base-uncased']:
-            model = TFXxxModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFXxxModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

templates/adding_a_new_model/tests/modeling_xxx_test.py

@@ -17,13 +17,12 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (XxxConfig, XxxModel, XxxForMaskedLM,
@@ -249,10 +248,8 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(XXX_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XxxModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = XxxModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/__init__.py

@@ -106,7 +106,7 @@ if is_torch_available():
                             XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_roberta import (RobertaForMaskedLM, RobertaModel,
                                 RobertaForSequenceClassification, RobertaForMultipleChoice,
-                                RobertaForTokenClassification,
+                                RobertaForTokenClassification, RobertaForQuestionAnswering,
                                 ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_distilbert import (DistilBertPreTrainedModel, DistilBertForMaskedLM, DistilBertModel,
                                 DistilBertForSequenceClassification, DistilBertForQuestionAnswering,

transformers/configuration_utils.py

@@ -56,8 +56,24 @@ class PretrainedConfig(object):
         self.torchscript = kwargs.pop('torchscript', False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop('use_bfloat16', False)
         self.pruned_heads = kwargs.pop('pruned_heads', {})
+
+        # Is decoder is used in encoder-decoder models to differentiate encoder from decoder
         self.is_decoder = kwargs.pop('is_decoder', False)
 
+        # Parameters for sequence generation
+        self.max_length = kwargs.pop('max_length', 20)
+        self.do_sample = kwargs.pop('do_sample', False)
+        self.num_beams = kwargs.pop('num_beams', 1)
+        self.temperature = kwargs.pop('temperature', 1.0)
+        self.top_k = kwargs.pop('top_k', 50)
+        self.top_p = kwargs.pop('top_p', 1.0)
+        self.repetition_penalty = kwargs.pop('repetition_penalty', 1.0)
+        self.bos_token_id = kwargs.pop('bos_token_id', 0)
+        self.pad_token_id = kwargs.pop('pad_token_id', 0)
+        self.eos_token_ids = kwargs.pop('eos_token_ids', 0)
+        self.length_penalty = kwargs.pop('length_penalty', 1.)
+        self.num_return_sequences = kwargs.pop('num_return_sequences', 1)
+
         # Fine-tuning task arguments
         self.finetuning_task = kwargs.pop('finetuning_task', None)
         self.num_labels = kwargs.pop('num_labels', 2)

transformers/configuration_xlm.py

@@ -110,6 +110,8 @@ class XLMConfig(PretrainedConfig):
                  summary_first_dropout=0.1,
                  start_n_top=5,
                  end_n_top=5,
+                 mask_token_id=0,
+                 lang_id=0,
                  **kwargs):
         """Constructs XLMConfig.
         """
@@ -143,6 +145,8 @@ class XLMConfig(PretrainedConfig):
         self.summary_first_dropout = summary_first_dropout
         self.start_n_top = start_n_top
         self.end_n_top = end_n_top
+        self.mask_token_id = mask_token_id
+        self.lang_id = lang_id
 
         if "n_words" in kwargs:
             self.n_words = kwargs["n_words"]

transformers/data/metrics/squad_metrics.py

@@ -377,7 +377,8 @@ def compute_predictions_logits(
     output_null_log_odds_file,
     verbose_logging,
     version_2_with_negative,
-    null_score_diff_threshold
+    null_score_diff_threshold,
+    tokenizer,
 ):
     """Write final predictions to the json file and log-odds of null if needed."""
     logger.info("Writing predictions to: %s" % (output_prediction_file))
@@ -474,11 +475,14 @@ def compute_predictions_logits(
                 orig_doc_start = feature.token_to_orig_map[pred.start_index]
                 orig_doc_end = feature.token_to_orig_map[pred.end_index]
                 orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
-                tok_text = " ".join(tok_tokens)
 
-                # De-tokenize WordPieces that have been split off.
-                tok_text = tok_text.replace(" ##", "")
-                tok_text = tok_text.replace("##", "")
+                tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
+
+                # tok_text = " ".join(tok_tokens)
+                #
+                # # De-tokenize WordPieces that have been split off.
+                # tok_text = tok_text.replace(" ##", "")
+                # tok_text = tok_text.replace("##", "")
 
                 # Clean whitespace
                 tok_text = tok_text.strip()

transformers/data/processors/squad.py

@@ -4,6 +4,9 @@ import logging
 import os
 import json
 import numpy as np
+from multiprocessing import Pool
+from multiprocessing import cpu_count
+from functools import partial
 
 from ...tokenization_bert import BasicTokenizer, whitespace_tokenize
 from .utils import DataProcessor, InputExample, InputFeatures
@@ -79,10 +82,168 @@ def _is_whitespace(c):
         return True
     return False
 
+def squad_convert_example_to_features(example, max_seq_length,
+                                       doc_stride, max_query_length, is_training):
+    features = []
+    if is_training and not example.is_impossible:
+        # Get start and end position
+        start_position = example.start_position
+        end_position = example.end_position
 
-def squad_convert_examples_to_features(
-    examples, tokenizer, max_seq_length, doc_stride, max_query_length, is_training, return_dataset=False
-):
+        # If the answer cannot be found in the text, then skip this example.
+        actual_text = " ".join(example.doc_tokens[start_position:(end_position + 1)])
+        cleaned_answer_text = " ".join(whitespace_tokenize(example.answer_text))
+        if actual_text.find(cleaned_answer_text) == -1:
+            logger.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text)
+            return []
+
+    tok_to_orig_index = []
+    orig_to_tok_index = []
+    all_doc_tokens = []
+    for (i, token) in enumerate(example.doc_tokens):
+        orig_to_tok_index.append(len(all_doc_tokens))
+        sub_tokens = tokenizer.tokenize(token)
+        for sub_token in sub_tokens:
+            tok_to_orig_index.append(i)
+            all_doc_tokens.append(sub_token)
+
+    if is_training and not example.is_impossible:
+        tok_start_position = orig_to_tok_index[example.start_position]
+        if example.end_position < len(example.doc_tokens) - 1:
+            tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
+        else:
+            tok_end_position = len(all_doc_tokens) - 1
+
+        (tok_start_position, tok_end_position) = _improve_answer_span(
+            all_doc_tokens, tok_start_position, tok_end_position, tokenizer, example.answer_text
+        )
+
+    spans = []
+
+    truncated_query = tokenizer.encode(example.question_text, add_special_tokens=False, max_length=max_query_length)
+    sequence_added_tokens = tokenizer.max_len - tokenizer.max_len_single_sentence + 1 \
+        if 'roberta' in str(type(tokenizer)) else tokenizer.max_len - tokenizer.max_len_single_sentence
+    sequence_pair_added_tokens = tokenizer.max_len - tokenizer.max_len_sentences_pair
+
+    span_doc_tokens = all_doc_tokens
+    while len(spans) * doc_stride < len(all_doc_tokens):
+
+        encoded_dict = tokenizer.encode_plus(
+            truncated_query if tokenizer.padding_side == "right" else span_doc_tokens,
+            span_doc_tokens if tokenizer.padding_side == "right" else truncated_query,
+            max_length=max_seq_length,
+            return_overflowing_tokens=True,
+            pad_to_max_length=True,
+            stride=max_seq_length - doc_stride - len(truncated_query) - sequence_pair_added_tokens,
+            truncation_strategy='only_second' if tokenizer.padding_side == "right" else 'only_first'
+        )
+
+        paragraph_len = min(len(all_doc_tokens) - len(spans) * doc_stride,
+                            max_seq_length - len(truncated_query) - sequence_pair_added_tokens)
+
+        if tokenizer.pad_token_id in encoded_dict['input_ids']:
+            non_padded_ids = encoded_dict['input_ids'][:encoded_dict['input_ids'].index(tokenizer.pad_token_id)]
+        else:
+            non_padded_ids = encoded_dict['input_ids']
+
+        tokens = tokenizer.convert_ids_to_tokens(non_padded_ids)
+
+        token_to_orig_map = {}
+        for i in range(paragraph_len):
+            index = len(truncated_query) + sequence_added_tokens + i if tokenizer.padding_side == "right" else i
+            token_to_orig_map[index] = tok_to_orig_index[len(spans) * doc_stride + i]
+
+        encoded_dict["paragraph_len"] = paragraph_len
+        encoded_dict["tokens"] = tokens
+        encoded_dict["token_to_orig_map"] = token_to_orig_map
+        encoded_dict["truncated_query_with_special_tokens_length"] = len(truncated_query) + sequence_added_tokens
+        encoded_dict["token_is_max_context"] = {}
+        encoded_dict["start"] = len(spans) * doc_stride
+        encoded_dict["length"] = paragraph_len
+
+        spans.append(encoded_dict)
+
+        if "overflowing_tokens" not in encoded_dict:
+            break
+        span_doc_tokens = encoded_dict["overflowing_tokens"]
+
+    for doc_span_index in range(len(spans)):
+        for j in range(spans[doc_span_index]["paragraph_len"]):
+            is_max_context = _new_check_is_max_context(spans, doc_span_index, doc_span_index * doc_stride + j)
+            index = j if tokenizer.padding_side == "left" else spans[doc_span_index][
+                                                                   "truncated_query_with_special_tokens_length"] + j
+            spans[doc_span_index]["token_is_max_context"][index] = is_max_context
+
+    for span in spans:
+        # Identify the position of the CLS token
+        cls_index = span['input_ids'].index(tokenizer.cls_token_id)
+
+        # p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
+        # Original TF implem also keep the classification token (set to 0) (not sure why...)
+        p_mask = np.array(span['token_type_ids'])
+
+        p_mask = np.minimum(p_mask, 1)
+
+        if tokenizer.padding_side == "right":
+            # Limit positive values to one
+            p_mask = 1 - p_mask
+
+        p_mask[np.where(np.array(span["input_ids"]) == tokenizer.sep_token_id)[0]] = 1
+
+        # Set the CLS index to '0'
+        p_mask[cls_index] = 0
+
+        span_is_impossible = example.is_impossible
+        start_position = 0
+        end_position = 0
+        if is_training and not span_is_impossible:
+            # For training, if our document chunk does not contain an annotation
+            # we throw it out, since there is nothing to predict.
+            doc_start = span["start"]
+            doc_end = span["start"] + span["length"] - 1
+            out_of_span = False
+
+            if not (tok_start_position >= doc_start and tok_end_position <= doc_end):
+                out_of_span = True
+
+            if out_of_span:
+                start_position = cls_index
+                end_position = cls_index
+                span_is_impossible = True
+            else:
+                if tokenizer.padding_side == "left":
+                    doc_offset = 0
+                else:
+                    doc_offset = len(truncated_query) + sequence_added_tokens
+
+                start_position = tok_start_position - doc_start + doc_offset
+                end_position = tok_end_position - doc_start + doc_offset
+
+        features.append(SquadFeatures(
+            span['input_ids'],
+            span['attention_mask'],
+            span['token_type_ids'],
+            cls_index,
+            p_mask.tolist(),
+            example_index=0, # Can not set unique_id and example_index here. They will be set after multiple processing.
+            unique_id=0,
+            paragraph_len=span['paragraph_len'],
+            token_is_max_context=span["token_is_max_context"],
+            tokens=span["tokens"],
+            token_to_orig_map=span["token_to_orig_map"],
+
+            start_position=start_position,
+            end_position=end_position
+        ))
+    return features
+
+def squad_convert_example_to_features_init(tokenizer_for_convert):
+    global tokenizer
+    tokenizer = tokenizer_for_convert
+
+def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
+                                       doc_stride, max_query_length, is_training, 
+                                       return_dataset=False, threads=1):
     """
     Converts a list of examples into a list of features that can be directly given as input to a model.
     It is model-dependant and takes advantage of many of the tokenizer's features to create the model's inputs.
@@ -97,6 +258,8 @@ def squad_convert_examples_to_features(
         return_dataset: Default False. Either 'pt' or 'tf'.
             if 'pt': returns a torch.data.TensorDataset,
             if 'tf': returns a tf.data.Dataset
+        threads: multiple processing threadsa-smi
+
 
     Returns:
         list of :class:`~transformers.data.processors.squad.SquadFeatures`
@@ -116,172 +279,28 @@ def squad_convert_examples_to_features(
         )
     """
 
-    # Defining helper methods
-    unique_id = 1000000000
-
+    # Defining helper methods    
     features = []
-    for (example_index, example) in enumerate(tqdm(examples, desc="Converting examples to features")):
-        if is_training and not example.is_impossible:
-            # Get start and end position
-            start_position = example.start_position
-            end_position = example.end_position
-
-            # If the answer cannot be found in the text, then skip this example.
-            actual_text = " ".join(example.doc_tokens[start_position : (end_position + 1)])
-            cleaned_answer_text = " ".join(whitespace_tokenize(example.answer_text))
-            if actual_text.find(cleaned_answer_text) == -1:
-                logger.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text)
-                continue
-
-        tok_to_orig_index = []
-        orig_to_tok_index = []
-        all_doc_tokens = []
-        for (i, token) in enumerate(example.doc_tokens):
-            orig_to_tok_index.append(len(all_doc_tokens))
-            sub_tokens = tokenizer.tokenize(token)
-            for sub_token in sub_tokens:
-                tok_to_orig_index.append(i)
-                all_doc_tokens.append(sub_token)
-
-        if is_training and not example.is_impossible:
-            tok_start_position = orig_to_tok_index[example.start_position]
-            if example.end_position < len(example.doc_tokens) - 1:
-                tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
-            else:
-                tok_end_position = len(all_doc_tokens) - 1
-
-            (tok_start_position, tok_end_position) = _improve_answer_span(
-                all_doc_tokens, tok_start_position, tok_end_position, tokenizer, example.answer_text
-            )
-
-        spans = []
-
-        truncated_query = tokenizer.encode(
-            example.question_text, add_special_tokens=False, max_length=max_query_length
-        )
-        sequence_added_tokens = tokenizer.max_len - tokenizer.max_len_single_sentence
-        sequence_pair_added_tokens = tokenizer.max_len - tokenizer.max_len_sentences_pair
-
-        span_doc_tokens = all_doc_tokens
-        while len(spans) * doc_stride < len(all_doc_tokens):
-
-            encoded_dict = tokenizer.encode_plus(
-                truncated_query if tokenizer.padding_side == "right" else span_doc_tokens,
-                span_doc_tokens if tokenizer.padding_side == "right" else truncated_query,
-                max_length=max_seq_length,
-                return_overflowing_tokens=True,
-                pad_to_max_length=True,
-                stride=max_seq_length - doc_stride - len(truncated_query) - sequence_pair_added_tokens,
-                truncation_strategy="only_second" if tokenizer.padding_side == "right" else "only_first",
-            )
-
-            paragraph_len = min(
-                len(all_doc_tokens) - len(spans) * doc_stride,
-                max_seq_length - len(truncated_query) - sequence_pair_added_tokens,
-            )
-
-            if tokenizer.pad_token_id in encoded_dict["input_ids"]:
-                non_padded_ids = encoded_dict["input_ids"][: encoded_dict["input_ids"].index(tokenizer.pad_token_id)]
-            else:
-                non_padded_ids = encoded_dict["input_ids"]
-
-            tokens = tokenizer.convert_ids_to_tokens(non_padded_ids)
-
-            token_to_orig_map = {}
-            for i in range(paragraph_len):
-                index = len(truncated_query) + sequence_added_tokens + i if tokenizer.padding_side == "right" else i
-                token_to_orig_map[index] = tok_to_orig_index[len(spans) * doc_stride + i]
-
-            encoded_dict["paragraph_len"] = paragraph_len
-            encoded_dict["tokens"] = tokens
-            encoded_dict["token_to_orig_map"] = token_to_orig_map
-            encoded_dict["truncated_query_with_special_tokens_length"] = len(truncated_query) + sequence_added_tokens
-            encoded_dict["token_is_max_context"] = {}
-            encoded_dict["start"] = len(spans) * doc_stride
-            encoded_dict["length"] = paragraph_len
-
-            spans.append(encoded_dict)
-
-            if "overflowing_tokens" not in encoded_dict:
-                break
-            span_doc_tokens = encoded_dict["overflowing_tokens"]
-
-        for doc_span_index in range(len(spans)):
-            for j in range(spans[doc_span_index]["paragraph_len"]):
-                is_max_context = _new_check_is_max_context(spans, doc_span_index, doc_span_index * doc_stride + j)
-                index = (
-                    j
-                    if tokenizer.padding_side == "left"
-                    else spans[doc_span_index]["truncated_query_with_special_tokens_length"] + j
-                )
-                spans[doc_span_index]["token_is_max_context"][index] = is_max_context
-
-        for span in spans:
-            # Identify the position of the CLS token
-            cls_index = span["input_ids"].index(tokenizer.cls_token_id)
-
-            # p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
-            # Original TF implem also keep the classification token (set to 0) (not sure why...)
-            p_mask = np.array(span["token_type_ids"])
-
-            p_mask = np.minimum(p_mask, 1)
-
-            if tokenizer.padding_side == "right":
-                # Limit positive values to one
-                p_mask = 1 - p_mask
-
-            p_mask[np.where(np.array(span["input_ids"]) == tokenizer.sep_token_id)[0]] = 1
-
-            # Set the CLS index to '0'
-            p_mask[cls_index] = 0
-
-            span_is_impossible = example.is_impossible
-            start_position = 0
-            end_position = 0
-            if is_training and not span_is_impossible:
-                # For training, if our document chunk does not contain an annotation
-                # we throw it out, since there is nothing to predict.
-                doc_start = span["start"]
-                doc_end = span["start"] + span["length"] - 1
-                out_of_span = False
-
-                if not (tok_start_position >= doc_start and tok_end_position <= doc_end):
-                    out_of_span = True
-
-                if out_of_span:
-                    start_position = cls_index
-                    end_position = cls_index
-                    span_is_impossible = True
-                else:
-                    if tokenizer.padding_side == "left":
-                        doc_offset = 0
-                    else:
-                        doc_offset = len(truncated_query) + sequence_added_tokens
-
-                    start_position = tok_start_position - doc_start + doc_offset
-                    end_position = tok_end_position - doc_start + doc_offset
-
-            features.append(
-                SquadFeatures(
-                    span["input_ids"],
-                    span["attention_mask"],
-                    span["token_type_ids"],
-                    cls_index,
-                    p_mask.tolist(),
-                    example_index=example_index,
-                    unique_id=unique_id,
-                    paragraph_len=span["paragraph_len"],
-                    token_is_max_context=span["token_is_max_context"],
-                    tokens=span["tokens"],
-                    token_to_orig_map=span["token_to_orig_map"],
-                    start_position=start_position,
-                    end_position=end_position,
-                )
-            )
-
+    threads = min(threads, cpu_count())
+    with Pool(threads, initializer=squad_convert_example_to_features_init, initargs=(tokenizer,)) as p:
+        annotate_ = partial(squad_convert_example_to_features, max_seq_length=max_seq_length,
+                                       doc_stride=doc_stride, max_query_length=max_query_length, is_training=is_training)
+        features = list(tqdm(p.imap(annotate_, examples, chunksize=32), total=len(examples), desc='convert squad examples to features'))
+    new_features = []
+    unique_id = 1000000000
+    example_index = 0
+    for example_features in tqdm(features, total=len(features), desc='add example index and unique id'):
+        if not example_features:
+            continue
+        for example_feature in example_features:
+            example_feature.example_index = example_index
+            example_feature.unique_id = unique_id
+            new_features.append(example_feature)
             unique_id += 1
-
-    if return_dataset == "pt":
+        example_index += 1
+    features = new_features
+    del new_features
+    if return_dataset == 'pt':
         if not is_torch_available():
             raise ImportError("Pytorch must be installed to return a pytorch dataset.")
 

transformers/file_utils.py

@@ -10,10 +10,9 @@ import json
 import logging
 import os
 import six
-import shutil
 import tempfile
 import fnmatch
-from functools import wraps
+from functools import partial, wraps
 from hashlib import sha256
 from io import open
 
@@ -25,6 +24,8 @@ from tqdm.auto import tqdm
 from contextlib import contextmanager
 from . import __version__
 
+from filelock import FileLock
+
 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
 try:
@@ -334,59 +335,60 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag
     # If we don't have a connection (etag is None) and can't identify the file
     # try to get the last downloaded one
     if not os.path.exists(cache_path) and etag is None:
-        matching_files = fnmatch.filter(os.listdir(cache_dir), filename + '.*')
-        matching_files = list(filter(lambda s: not s.endswith('.json'), matching_files))
+        matching_files = [
+            file
+            for file in fnmatch.filter(os.listdir(cache_dir), filename + '.*')
+            if not file.endswith('.json') and not file.endswith('.lock')
+        ]
         if matching_files:
             cache_path = os.path.join(cache_dir, matching_files[-1])
 
-    if resume_download:
-        incomplete_path = cache_path + '.incomplete'
-        @contextmanager
-        def _resumable_file_manager():
-            with open(incomplete_path,'a+b') as f:
-                yield f
-            os.remove(incomplete_path)
-        temp_file_manager = _resumable_file_manager
-        if os.path.exists(incomplete_path):
-            resume_size = os.stat(incomplete_path).st_size
-        else:
-            resume_size = 0
-    else:
-        temp_file_manager = tempfile.NamedTemporaryFile
-        resume_size = 0
+    # Prevent parallel downloads of the same file with a lock.
+    lock_path = cache_path + '.lock'
+    with FileLock(lock_path):
 
-    if etag is not None and (not os.path.exists(cache_path) or force_download):
-        # Download to temporary file, then copy to cache dir once finished.
-        # Otherwise you get corrupt cache entries if the download gets interrupted.
-        with temp_file_manager() as temp_file:
-            logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
-
-            # GET file object
-            if url.startswith("s3://"):
-                if resume_download:
-                    logger.warn('Warning: resumable downloads are not implemented for "s3://" urls')
-                s3_get(url, temp_file, proxies=proxies)
+        if resume_download:
+            incomplete_path = cache_path + '.incomplete'
+            @contextmanager
+            def _resumable_file_manager():
+                with open(incomplete_path,'a+b') as f:
+                    yield f
+            temp_file_manager = _resumable_file_manager
+            if os.path.exists(incomplete_path):
+                resume_size = os.stat(incomplete_path).st_size
             else:
-                http_get(url, temp_file, proxies=proxies, resume_size=resume_size, user_agent=user_agent)
+                resume_size = 0
+        else:
+            temp_file_manager = partial(tempfile.NamedTemporaryFile, dir=cache_dir, delete=False)
+            resume_size = 0
 
-            # we are copying the file before closing it, so flush to avoid truncation
-            temp_file.flush()
-            # shutil.copyfileobj() starts at the current position, so go to the start
-            temp_file.seek(0)
+        if etag is not None and (not os.path.exists(cache_path) or force_download):
+            # Download to temporary file, then copy to cache dir once finished.
+            # Otherwise you get corrupt cache entries if the download gets interrupted.
+            with temp_file_manager() as temp_file:
+                logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
 
-            logger.info("copying %s to cache at %s", temp_file.name, cache_path)
-            with open(cache_path, 'wb') as cache_file:
-                shutil.copyfileobj(temp_file, cache_file)
+                # GET file object
+                if url.startswith("s3://"):
+                    if resume_download:
+                        logger.warn('Warning: resumable downloads are not implemented for "s3://" urls')
+                    s3_get(url, temp_file, proxies=proxies)
+                else:
+                    http_get(url, temp_file, proxies=proxies, resume_size=resume_size, user_agent=user_agent)
 
-            logger.info("creating metadata file for %s", cache_path)
-            meta = {'url': url, 'etag': etag}
-            meta_path = cache_path + '.json'
-            with open(meta_path, 'w') as meta_file:
-                output_string = json.dumps(meta)
-                if sys.version_info[0] == 2 and isinstance(output_string, str):
-                    output_string = unicode(output_string, 'utf-8')  # The beauty of python 2
-                meta_file.write(output_string)
+                # we are copying the file before closing it, so flush to avoid truncation
+                temp_file.flush()
 
-            logger.info("removing temp file %s", temp_file.name)
+                logger.info("storing %s in cache at %s", url, cache_path)
+                os.rename(temp_file.name, cache_path)
+
+                logger.info("creating metadata file for %s", cache_path)
+                meta = {'url': url, 'etag': etag}
+                meta_path = cache_path + '.json'
+                with open(meta_path, 'w') as meta_file:
+                    output_string = json.dumps(meta)
+                    if sys.version_info[0] == 2 and isinstance(output_string, str):
+                        output_string = unicode(output_string, 'utf-8')  # The beauty of python 2
+                    meta_file.write(output_string)
 
     return cache_path

transformers/modeling_albert.py

@@ -572,7 +572,7 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -624,7 +624,7 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
 
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 

transformers/modeling_bert.py

@@ -754,7 +754,7 @@ class BertForPreTraining(BertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
         **next_sentence_label**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see ``input_ids`` docstring)
@@ -813,7 +813,7 @@ class BertForPreTraining(BertPreTrainedModel):
         outputs = (prediction_scores, seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
 
         if masked_lm_labels is not None and next_sentence_label is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
             total_loss = masked_lm_loss + next_sentence_loss
@@ -830,12 +830,12 @@ class BertForMaskedLM(BertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
         **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the left-to-right language modeling loss (next word prediction).
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -897,7 +897,7 @@ class BertForMaskedLM(BertPreTrainedModel):
         # 2. If `lm_labels` is provided we are in a causal scenario where we
         #    try to predict the next token for each input in the decoder.
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)  # -1 index = padding token
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 
@@ -905,7 +905,7 @@ class BertForMaskedLM(BertPreTrainedModel):
             # we are doing next-token prediction; shift prediction scores and input ids by one
             prediction_scores = prediction_scores[:, :-1, :].contiguous()
             lm_labels = lm_labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             ltr_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), lm_labels.view(-1))
             outputs = (ltr_lm_loss,) + outputs
 
@@ -969,7 +969,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
         if next_sentence_label is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
             outputs = (next_sentence_loss,) + outputs
 

transformers/modeling_camembert.py

@@ -156,7 +156,7 @@ class CamembertForMaskedLM(RobertaForMaskedLM):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:

transformers/modeling_ctrl.py

@@ -268,7 +268,7 @@ class CTRLModel(CTRLPreTrainedModel):
 
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLModel.from_pretrained('ctrl')
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -429,7 +429,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -458,7 +458,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLLMHeadModel.from_pretrained('ctrl')
 
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -494,7 +494,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs

transformers/modeling_distilbert.py

@@ -415,7 +415,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertModel.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -491,7 +491,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -511,7 +511,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -528,7 +528,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
 
         self.init_weights()
 
-        self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+        self.mlm_loss_fct = nn.CrossEntropyLoss()
 
     def get_output_embeddings(self):
         return self.vocab_projector
@@ -581,7 +581,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -656,7 +656,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForQuestionAnswering.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/modeling_encoder_decoder.py

@@ -18,9 +18,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 import os
+import warnings
 
 import torch
 from torch import nn
+from tqdm import trange
 
 from .modeling_auto import AutoModel, AutoModelWithLMHead
 
@@ -119,8 +121,7 @@ class PreTrainedEncoderDecoder(nn.Module):
         kwargs_common = {
             argument: value
             for argument, value in kwargs.items()
-            if not argument.startswith("encoder_")
-            and not argument.startswith("decoder_")
+            if not argument.startswith("encoder_") and not argument.startswith("decoder_")
         }
         kwargs_decoder = kwargs_common.copy()
         kwargs_encoder = kwargs_common.copy()
@@ -166,7 +167,39 @@ class PreTrainedEncoderDecoder(nn.Module):
 
         We save the encoder' and decoder's parameters in two separate directories.
         """
+
+        # If the root output directory does not exist, create it 
+        if not os.path.exists(save_directory):
+            os.mkdir(save_directory)
+
+        # Check whether the output directory is empty or not
+        sub_directories = [directory for directory in os.listdir(save_directory)
+            if os.path.isdir(os.path.join(save_directory, directory))]
+
+        if len(sub_directories) > 0:
+            if "encoder" in sub_directories and "decoder" in sub_directories:
+                print("WARNING: there is an older version of encoder-decoder saved in" +\
+                    " the output directory. The default behaviour is to overwrite them.")
+
+            # Empty the output directory
+            for directory_to_remove in sub_directories:
+                # Remove all files into the subdirectory
+                files_to_remove = os.listdir(os.path.join(save_directory, directory_to_remove))
+                for file_to_remove in files_to_remove:
+                    os.remove(os.path.join(save_directory, directory_to_remove, file_to_remove))
+                # Remove the subdirectory itself
+                os.rmdir(os.path.join(save_directory, directory_to_remove))
+
+            assert(len(os.listdir(save_directory)) == 0) # sanity check
+
+        # Create the "encoder" directory inside the output directory and save the encoder into it
+        if not os.path.exists(os.path.join(save_directory, "encoder")):
+            os.mkdir(os.path.join(save_directory, "encoder"))
         self.encoder.save_pretrained(os.path.join(save_directory, "encoder"))
+
+        # Create the "encoder" directory inside the output directory and save the decoder into it
+        if not os.path.exists(os.path.join(save_directory, "decoder")):
+            os.mkdir(os.path.join(save_directory, "decoder"))
         self.decoder.save_pretrained(os.path.join(save_directory, "decoder"))
 
     def forward(self, encoder_input_ids, decoder_input_ids, **kwargs):
@@ -188,32 +221,7 @@ class PreTrainedEncoderDecoder(nn.Module):
                 Indices of decoder input sequence tokens in the vocabulary.
             kwargs: (`optional`) Remaining dictionary of keyword arguments.
         """
-        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
-        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
-        # that apply to the model as whole.
-        # We let the specific kwargs override the common ones in case of conflict.
-        kwargs_common = {
-            argument: value
-            for argument, value in kwargs.items()
-            if not argument.startswith("encoder_")
-            and not argument.startswith("decoder_")
-        }
-        kwargs_decoder = kwargs_common.copy()
-        kwargs_encoder = kwargs_common.copy()
-        kwargs_encoder.update(
-            {
-                argument[len("encoder_") :]: value
-                for argument, value in kwargs.items()
-                if argument.startswith("encoder_")
-            }
-        )
-        kwargs_decoder.update(
-            {
-                argument[len("decoder_") :]: value
-                for argument, value in kwargs.items()
-                if argument.startswith("decoder_")
-            }
-        )
+        kwargs_encoder, kwargs_decoder = self.prepare_model_kwargs(**kwargs)
 
         # Encode if needed (training, first prediction pass)
         encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
@@ -223,15 +231,47 @@ class PreTrainedEncoderDecoder(nn.Module):
         else:
             encoder_outputs = ()
 
-        # Decode
         kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
-        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get(
-            "attention_mask", None
-        )
-        decoder_outputs = self.decoder(decoder_input_ids, **kwargs_decoder)
+        decoder_outputs = self.decoder(decoder_input_ids, encoder_hidden_states, **kwargs_decoder)
 
         return decoder_outputs + encoder_outputs
 
+    @staticmethod
+    def prepare_model_kwargs(**kwargs):
+        """ Prepare the encoder and decoder's keyword arguments.
+
+        Keyword arguments come in 3 flavors:
+        - encoder-specific (prefixed by `encoder_`)
+        - decoder-specific (prefixed by `decoder_`)
+        - those that apply to the model as whole.
+
+        We let the specific kwargs override the common ones in case of
+        conflict.
+        """
+        kwargs_common = {
+            argument: value
+            for argument, value in kwargs.items()
+            if not argument.startswith("encoder_") and not argument.startswith("decoder_")
+        }
+        decoder_kwargs = kwargs_common.copy()
+        encoder_kwargs = kwargs_common.copy()
+        encoder_kwargs.update(
+            {
+                argument[len("encoder_") :]: value
+                for argument, value in kwargs.items()
+                if argument.startswith("encoder_")
+            }
+        )
+        decoder_kwargs.update(
+            {
+                argument[len("decoder_") :]: value
+                for argument, value in kwargs.items()
+                if argument.startswith("decoder_")
+            }
+        )
+        decoder_kwargs["encoder_attention_mask"] = encoder_kwargs.get("attention_mask", None)
+        return encoder_kwargs, decoder_kwargs
+
 
 class Model2Model(PreTrainedEncoderDecoder):
     r"""

transformers/modeling_gpt2.py

@@ -345,7 +345,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2Model.from_pretrained('gpt2')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -494,7 +494,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -523,7 +523,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2LMHeadModel.from_pretrained('gpt2')
 
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -557,7 +557,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs
@@ -579,7 +579,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
         **mc_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
             Labels for computing the multiple choice classification loss.
@@ -668,7 +668,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         if lm_labels is not None:
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = lm_labels[..., 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs

transformers/modeling_openai.py

@@ -349,7 +349,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -471,7 +471,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -491,7 +491,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -523,7 +523,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs
@@ -545,7 +545,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
         **mc_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
             Labels for computing the multiple choice classification loss.
@@ -622,7 +622,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
         if lm_labels is not None:
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = lm_labels[..., 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs

transformers/modeling_roberta.py

@@ -51,24 +51,44 @@ class RobertaEmbeddings(BertEmbeddings):
                                                 padding_idx=self.padding_idx)
 
     def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
-        if input_ids is not None:
-            input_shape = input_ids.size()
-        else:
-            input_shape = inputs_embeds.size()[:-1]
-
-        seq_length = input_shape[1]
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
         if position_ids is None:
-            # Position numbers begin at padding_idx+1. Padding symbols are ignored.
-            # cf. fairseq's `utils.make_positions`
-            position_ids = torch.arange(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=torch.long, device=device)
-            position_ids = position_ids.unsqueeze(0).expand(input_shape)
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = self.create_position_ids_from_input_ids(input_ids).to(input_ids.device)
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+
         return super(RobertaEmbeddings, self).forward(input_ids,
                                                       token_type_ids=token_type_ids,
                                                       position_ids=position_ids,
                                                       inputs_embeds=inputs_embeds)
 
+    def create_position_ids_from_input_ids(self, x):
+        """ Replace non-padding symbols with their position numbers. Position numbers begin at
+        padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
+        `utils.make_positions`.
+
+        :param torch.Tensor x:
+        :return torch.Tensor:
+        """
+        mask = x.ne(self.padding_idx).long()
+        incremental_indicies = torch.cumsum(mask, dim=1) * mask
+        return incremental_indicies + self.padding_idx
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """ We are provided embeddings directly. We cannot infer which are padded so just generate
+        sequential position ids.
+
+        :param torch.Tensor inputs_embeds:
+        :return torch.Tensor:
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(self.padding_idx+1, sequence_length+self.padding_idx+1, dtype=torch.long,
+                                    device=inputs_embeds.device)
+        return position_ids.unsqueeze(0).expand(input_shape)
+
 
 ROBERTA_START_DOCSTRING = r"""    The RoBERTa model was proposed in
     `RoBERTa: A Robustly Optimized BERT Pretraining Approach`_
@@ -168,7 +188,7 @@ class RobertaModel(BertModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaModel.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -196,7 +216,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -216,7 +236,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaForMaskedLM.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -250,7 +270,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
 
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 
@@ -307,7 +327,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaForSequenceClassification.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -555,3 +575,89 @@ class RobertaClassificationHead(nn.Module):
         x = self.dropout(x)
         x = self.out_proj(x)
         return x
+
+
+@add_start_docstrings("""Roberta Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class RobertaForQuestionAnswering(BertPreTrainedModel):
+    r"""
+        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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::
+        tokenizer = RobertaTokenizer.from_pretrained('roberta-large')
+        model = RobertaForQuestionAnswering.from_pretrained('roberta-large')
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        input_ids = tokenizer.encode(question, text)
+        start_scores, end_scores = model(torch.tensor([input_ids]))
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
+        answer = ' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1])
+    """
+    config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super(RobertaForQuestionAnswering, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = RobertaModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                start_positions=None, end_positions=None):
+
+        outputs = self.roberta(input_ids,
+                               attention_mask=attention_mask,
+                               token_type_ids=token_type_ids,
+                               position_ids=position_ids,
+                               head_mask=head_mask)
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

transformers/modeling_tf_bert.py

@@ -677,7 +677,7 @@ class TFBertModel(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertModel.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -716,7 +716,7 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForPreTraining.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, seq_relationship_scores = outputs[:2]
 
@@ -765,7 +765,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForMaskedLM.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores = outputs[0]
 
@@ -812,7 +812,7 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForNextSentencePrediction.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         seq_relationship_scores = outputs[0]
 
@@ -857,7 +857,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 
@@ -994,7 +994,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForTokenClassification.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         scores = outputs[0]
 
@@ -1047,7 +1047,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForQuestionAnswering.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         start_scores, end_scores = outputs[:2]
 

transformers/modeling_tf_ctrl.py

@@ -418,7 +418,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
 
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = TFCTRLModel.from_pretrained('ctrl')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -481,7 +481,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = TFCTRLLMHeadModel.from_pretrained('ctrl')
 
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 

transformers/modeling_tf_gpt2.py

@@ -454,7 +454,7 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
 
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = TFGPT2Model.from_pretrained('gpt2')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -495,7 +495,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = TFGPT2LMHeadModel.from_pretrained('gpt2')
 
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 

transformers/modeling_tf_openai.py

@@ -431,7 +431,7 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = TFOpenAIGPTModel.from_pretrained('openai-gpt')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -467,7 +467,7 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = TFOpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 

transformers/modeling_tf_roberta.py

@@ -20,7 +20,6 @@ from __future__ import (absolute_import, division, print_function,
 
 import logging
 
-import numpy as np
 import tensorflow as tf
 
 from .configuration_roberta import RobertaConfig
@@ -46,17 +45,40 @@ class TFRobertaEmbeddings(TFBertEmbeddings):
         super(TFRobertaEmbeddings, self).__init__(config, **kwargs)
         self.padding_idx = 1
 
+    def create_position_ids_from_input_ids(self, x):
+        """ Replace non-padding symbols with their position numbers. Position numbers begin at
+        padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
+        `utils.make_positions`.
+        :param torch.Tensor x:
+        :return torch.Tensor:
+        """
+        mask = tf.cast(tf.math.not_equal(x, self.padding_idx), dtype=tf.int32)
+        incremental_indicies = tf.math.cumsum(mask, axis=1) * mask
+        return incremental_indicies + self.padding_idx
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """ We are provided embeddings directly. We cannot infer which are padded so just generate
+        sequential position ids.
+        :param torch.Tensor inputs_embeds:
+        :return torch.Tensor:
+        """
+        seq_length = shape_list(inputs_embeds)[1]
+
+        position_ids = tf.range(self.padding_idx + 1,
+                                seq_length + self.padding_idx + 1,
+                                dtype=tf.int32)[tf.newaxis, :]
+        return position_ids
+
     def _embedding(self, inputs, training=False):
         """Applies embedding based on inputs tensor."""
         input_ids, position_ids, token_type_ids, inputs_embeds = inputs
 
-        if input_ids is not None:
-            seq_length = shape_list(input_ids)[1]
-        else:
-            seq_length = shape_list(inputs_embeds)[1]
-
         if position_ids is None:
-            position_ids = tf.range(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=tf.int32)[tf.newaxis, :]
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = self.create_position_ids_from_input_ids(input_ids)
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
 
         return super(TFRobertaEmbeddings, self)._embedding([input_ids, position_ids, token_type_ids, inputs_embeds], training=training)
 
@@ -199,7 +221,7 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = TFRobertaModel.from_pretrained('roberta-base')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -250,12 +272,6 @@ class TFRobertaLMHead(tf.keras.layers.Layer):
     ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
 class TFRobertaForMaskedLM(TFRobertaPreTrainedModel):
     r"""
-        **masked_lm_labels**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
-            Labels for computing the masked language modeling loss.
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
-            in ``[0, ..., config.vocab_size]``
-
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``tf.Tensor`` of shape ``(1,)``:
             Masked language modeling loss.
@@ -276,7 +292,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = TFRobertaForMaskedLM.from_pretrained('roberta-base')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         prediction_scores = outputs[0]
 
@@ -347,7 +363,7 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel):
 
         tokenizer = RoertaTokenizer.from_pretrained('roberta-base')
         model = TFRobertaForSequenceClassification.from_pretrained('roberta-base')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         labels = tf.constant([1])[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]

transformers/modeling_tf_transfo_xl.py

@@ -673,7 +673,7 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
 
         tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
         model = TFTransfoXLModel.from_pretrained('transfo-xl-wt103')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states, mems = outputs[:2]
 
@@ -715,7 +715,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
 
         tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
         model = TFTransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, mems = outputs[:2]
 

transformers/modeling_tf_xlm.py

@@ -576,7 +576,7 @@ class TFXLMModel(TFXLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = TFXLMModel.from_pretrained('xlm-mlm-en-2048')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -649,7 +649,7 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = TFXLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -695,7 +695,7 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = TFXLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         labels = tf.constant([1])[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
@@ -743,7 +743,7 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = TFXLMForQuestionAnsweringSimple.from_pretrained('xlm-mlm-en-2048')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         start_scores, end_scores = outputs[:2]
 

transformers/modeling_tf_xlnet.py

@@ -552,7 +552,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) " \
             "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
         if input_mask is None and attention_mask is not None:
-            input_mask = 1.0 - attention_mask
+            input_mask = 1.0 - tf.cast(attention_mask, dtype=dtype_float)
         if input_mask is not None and perm_mask is not None:
             data_mask = input_mask[None] + perm_mask
         elif input_mask is not None and perm_mask is None:
@@ -811,7 +811,7 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
 
         tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
         model = TFXLNetModel.from_pretrained('xlnet-large-cased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -855,7 +855,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
         model = TFXLNetLMHeadModel.from_pretrained('xlnet-large-cased')
 
         # We show how to setup inputs to predict a next token using a bi-directional context.
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is very <mask>"))[None, :]  # We will predict the masked token
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is very <mask>", add_special_tokens=True))[None, :]  # We will predict the masked token
         perm_mask = tf.zeros((1, input_ids.shape[1], input_ids.shape[1]))
         perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
         target_mapping = tf.zeros((1, 1, input_ids.shape[1]))  # Shape [1, 1, seq_length] => let's predict one token
@@ -911,7 +911,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
 
         tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
         model = TFXLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 
@@ -1022,7 +1022,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
 
         tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
         model = TFXLNetForQuestionAnsweringSimple.from_pretrained('xlnet-base-cased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         start_scores, end_scores = outputs[:2]
 
@@ -1086,7 +1086,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
 
 #         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
 #         model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
-#         input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+#         input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
 #         start_positions = tf.constant([1])
 #         end_positions = tf.constant([3])
 #         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/modeling_transfo_xl.py

@@ -36,7 +36,7 @@ from torch.nn.parameter import Parameter
 
 from .modeling_utils import PreTrainedModel, Conv1D, prune_conv1d_layer, SequenceSummary
 from .configuration_transfo_xl import TransfoXLConfig
-from .modeling_transfo_xl_utilities import ProjectedAdaptiveLogSoftmax, sample_logits
+from .modeling_transfo_xl_utilities import ProjectedAdaptiveLogSoftmax, sample_logits, LogUniformSampler
 from .file_utils import add_start_docstrings
 
 logger = logging.getLogger(__name__)
@@ -582,7 +582,7 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
 
         tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
         model = TransfoXLModel.from_pretrained('transfo-xl-wt103')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states, mems = outputs[:2]
 
@@ -796,17 +796,17 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
     TRANSFO_XL_START_DOCSTRING, TRANSFO_XL_INPUTS_DOCSTRING)
 class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
     r"""
-        **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
             Language modeling loss.
-        **prediction_scores**: ``None`` if ``lm_labels`` is provided else ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+        **prediction_scores**: ``None`` if ``labels`` is provided else ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
             We don't output them when the loss is computed to speedup adaptive softmax decoding.
         **mems**:
@@ -825,7 +825,7 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
 
         tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
         model = TransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, mems = outputs[:2]
 
@@ -908,3 +908,11 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
                 outputs = [softmax_output, None] + outputs
 
         return outputs  # (loss), logits or None if labels is not None (speed up adaptive softmax), new_mems, (all hidden states), (all attentions)
+
+    def get_output_embeddings(self):
+        """ Double-check if you are using adaptive softmax.
+        """
+        if self.sample_softmax > 0:
+            return self.out_layer
+        else:
+            return self.crit.out_layers[-1]

transformers/modeling_utils.py

@@ -1,5 +1,5 @@
 # coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
 # Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -496,6 +496,403 @@ class PreTrainedModel(nn.Module):
 
         return model
 
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        return {"input_ids": input_ids}
+
+    @torch.no_grad()
+    def generate(self, input_ids=None, max_length=None, do_sample=None, num_beams=None,
+                 temperature=None, top_k=None, top_p=None, repetition_penalty=None,
+                 bos_token_id=None, pad_token_id=None, eos_token_ids=None,
+                 length_penalty=None, num_return_sequences=None):
+        """ Sequence generator for models with a LM head.
+
+        The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
+        and beam-search.
+
+        Adapted in part from Facebook's XLM beam search code: https://github.com/facebookresearch/XLM
+
+        Params:
+            **input_ids**: (`optional`) `torch.LongTensor` of shape (1, sequence_length)
+                The sequence used as a prompt for the generation. If `None` the method initializes
+                it as an empty `torch.LongTensor` of shape (1,)
+            **max_length**: (`optional`) int
+                The max length of the sequence to be generated.  Between 1 and infinity. Default to 20.
+            **do_sample**: (`optional`) bool
+                If set to `False` we use greedy decoding; otherwise sampling. Default to greedy sampling.
+            **num_beams**: (`optional`) int
+                Number of beams for beam search. 1 means no beam serach. Default to 1.
+            **temperature**: (`optional`) float
+                The value used to module the next token probabilities.
+            **top_k**: (`optional`) int
+                The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
+            **top_p**: (`optional`) float
+                The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.
+            **repetition_penalty**: (`optional`) float
+                The parameter for repetition penalty. Between 1.0 and + infinity. 1.0 means no penalty. Default to 1.
+            **bos_token_id**: (`optional`) int
+                Beginning of sentence token if no prompt is provided. Default to 0.
+            **eos_token_ids**: (`optional`) int or list of int
+                End of sequence token or list of tokens to stop the generation. Default to 0.
+            **length_penalty**: (`optional`) int
+                Exponential penalty to the length. Default to 0.
+            **length_penalty**: (`optional`) float
+                Exponential penalty to the length. Default to 1.
+            **num_return_sequences**: (`optional`) int
+                The number of independantly computed returned sequences for each element in the batch. Default to 1.
+        """
+
+        # We cannot generate if the model does not have a LM head
+        if self.get_output_embeddings() is None:
+            raise AttributeError("You tried to generate sequences with a model that does not have a LM Head."
+                                 "Please use another model class (e.g. `OpenAIGPTLMHeadModel`)")
+
+        max_length = max_length if max_length is not None else self.config.max_length
+        do_sample = do_sample if do_sample is not None else self.config.do_sample
+        num_beams = num_beams if num_beams is not None else self.config.num_beams
+        temperature = temperature if temperature is not None else self.config.temperature
+        top_k = top_k if top_k is not None else self.config.top_k
+        top_p = top_p if top_p is not None else self.config.top_p
+        repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
+        bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
+        pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
+        eos_token_ids = eos_token_ids if eos_token_ids is not None else self.config.eos_token_ids
+        length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
+        num_return_sequences = num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]  # overriden by the input batch_size
+        else:
+            batch_size = 1
+        if isinstance(eos_token_ids, int):
+            eos_token_ids = [eos_token_ids]
+
+        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictely positive integer."
+        assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
+        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictely positive integer."
+        # assert temperature >= 0, "`temperature` should be positive."
+        assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
+        assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
+        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
+        assert isinstance(bos_token_id, int) and bos_token_id >= 0, "`bos_token_id` should be a positive integer."
+        assert isinstance(pad_token_id, int) and pad_token_id >= 0, "`pad_token_id` should be a positive integer."
+        assert isinstance(eos_token_ids, (list, tuple)) and (e >= 0 for e in eos_token_ids), \
+                   "`eos_token_ids` should be a positive integer or a list/tuple of positive integers."
+        assert length_penalty > 0, "`length_penalty` should be strictely positive."
+        assert isinstance(num_return_sequences, int) and num_return_sequences > 0, "`num_return_sequences` should be a strictely positive integer."
+
+        if input_ids is None:
+            input_ids = torch.full((batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device)
+        else:
+            assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
+
+        # current position and vocab size
+        cur_len = input_ids.shape[1]
+        vocab_size = self.config.vocab_size
+
+        if num_return_sequences != 1:
+            # Expand input to num return sequences
+            input_ids = input_ids.unsqueeze(1).expand(batch_size, num_return_sequences, cur_len)
+            input_ids = input_ids.contiguous().view(batch_size * num_return_sequences, cur_len)   # (batch_size * num_return_sequences, cur_len)
+            effective_batch_size = batch_size * num_return_sequences
+        else:
+            effective_batch_size = batch_size
+
+        if num_beams > 1:
+            output = self._generate_beam_search(input_ids, cur_len, max_length, do_sample,
+                                                temperature, top_k, top_p, repetition_penalty,
+                                                pad_token_id, eos_token_ids, effective_batch_size,
+                                                length_penalty, num_beams, vocab_size)
+        else:
+            output = self._generate_no_beam_search(input_ids, cur_len, max_length, do_sample,
+                                             temperature, top_k, top_p, repetition_penalty,
+                                             pad_token_id, eos_token_ids, effective_batch_size)
+
+        if num_return_sequences != 1:
+            output = output.view(batch_size, num_return_sequences, -1)
+        return output
+
+    def _generate_no_beam_search(self, input_ids, cur_len, max_length, do_sample,
+                                 temperature, top_k, top_p, repetition_penalty,
+                                 pad_token_id, eos_token_ids, batch_size):
+        """ Generate sequences for each example without beam search (num_beams == 1).
+            All returned sequence are generated independantly.
+        """
+        # current position / max lengths / length of generated sentences / unfinished sentences
+        unfinished_sents = input_ids.new(batch_size).fill_(1)
+
+        # TODO: add cached compute states
+        pasts = None
+
+        while cur_len < max_length:
+            model_inputs = self.prepare_inputs_for_generation(input_ids, pasts=pasts)
+            outputs = self(**model_inputs)
+            next_token_logits = outputs[0][:, -1, :]
+
+            # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
+            if repetition_penalty != 1.0:
+                for i in range(batch_size):
+                    for previous_tokens in set(input_ids[i].tolist()):
+                        next_token_logits[i, previous_tokens] /= repetition_penalty
+
+            if do_sample:
+                # Temperature (higher temperature => more likely to sample low probability tokens)
+                if temperature > 0 and temperature != 1.0:
+                    next_token_logits = next_token_logits / temperature
+                # Top-p/top-k filtering
+                next_token_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
+                # Sample
+                next_token = torch.multinomial(F.softmax(next_token_logits, dim=-1), num_samples=1).squeeze(1)
+            else:
+                # Greedy decoding
+                next_token = torch.argmax(next_token_logits, dim=-1)
+
+            # update generations and finished sentences
+            tokens_to_add = next_token * unfinished_sents + pad_token_id * (1 - unfinished_sents)
+            input_ids = torch.cat([input_ids, tokens_to_add.unsqueeze(-1)], dim=-1)
+            for eos_token_id in eos_token_ids:
+                unfinished_sents.mul_(tokens_to_add.ne(eos_token_id).long())
+            cur_len = cur_len + 1
+
+            # stop when there is a </s> in each sentence, or if we exceed the maximul length
+            if unfinished_sents.max() == 0:
+                break
+
+        # add eos_token_ids to unfinished sentences
+        if cur_len == max_length:
+            input_ids[:, -1].masked_fill_(unfinished_sents.to(dtype=torch.bool), eos_token_ids[0])
+
+        return input_ids
+
+    def _generate_beam_search(self, input_ids, cur_len, max_length, do_sample,
+                              temperature, top_k, top_p, repetition_penalty,
+                              pad_token_id, eos_token_ids, batch_size,
+                              length_penalty, num_beams, vocab_size):
+        """ Generate sequences for each example with beam search.
+        """
+        # Expand input to num beams
+        input_ids = input_ids.unsqueeze(1).expand(batch_size, num_beams, cur_len)
+        input_ids = input_ids.contiguous().view(batch_size * num_beams, cur_len)   # (batch_size * num_beams, cur_len)
+
+        # generated hypotheses
+        generated_hyps = [BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=False) for _ in range(batch_size)]
+
+        # scores for each sentence in the beam
+        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
+        beam_scores[:, 1:] = -1e9
+        beam_scores = beam_scores.view(-1)                                      # shape (batch_size * num_beams,)
+
+        # cache compute states
+        pasts = None  # self.prepare_pasts()
+
+        # done sentences
+        done = [False for _ in range(batch_size)]
+
+        while cur_len < max_length:
+            model_inputs = self.prepare_inputs_for_generation(input_ids, pasts=pasts)
+            scores = self(**model_inputs)[0]                                    # (batch_size * num_beams, cur_len, vocab_size)
+            scores = scores[:, -1, :]                                           # (batch_size * num_beams, vocab_size)
+
+            # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
+            if repetition_penalty != 1.0:
+                for i in range(batch_size * num_beams):
+                    for previous_tokens in set(input_ids[i].tolist()):
+                        scores[i, previous_tokens] /= repetition_penalty
+
+            if do_sample:
+                # Temperature (higher temperature => more likely to sample low probability tokens)
+                if temperature > 0 and temperature != 1.0:
+                    scores = scores / temperature
+                # Top-p/top-k filtering
+                scores = top_k_top_p_filtering(scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2)  # (batch_size * num_beams, vocab_size)
+                # Sample 2 next words for each beam (so we have some spare tokens and match output of greedy beam search)
+                next_words = torch.multinomial(F.softmax(scores, dim=-1), num_samples=2)    # (batch_size * num_beams, 2)
+                # Compute next scores
+                _scores = F.log_softmax(scores, dim=-1)                                     # (batch_size * num_beams, vocab_size)
+                _scores = torch.gather(_scores, -1, next_words)                             # (batch_size * num_beams, 2)
+                next_scores = _scores + beam_scores[:, None].expand_as(_scores)             # (batch_size * num_beams, 2)
+                # Match shape of greedy beam search
+                next_words = next_words.view(batch_size, 2 * num_beams)                     # (batch_size, 2 * num_beams)
+                next_scores = next_scores.view(batch_size, 2 * num_beams)                   # (batch_size, 2 * num_beams)
+            else:
+                # do greedy beam search
+                scores = F.log_softmax(scores, dim=-1)                          # (batch_size * num_beams, vocab_size)
+                assert scores.size() == (batch_size * num_beams, vocab_size)
+                # Add the log prob of the new beams to the log prob of the beginning of the sequence (sum of logs == log of the product)
+                _scores = scores + beam_scores[:, None].expand_as(scores)       # (batch_size * num_beams, vocab_size)
+                # re-organize to group the beam together (we are keeping top hypothesis accross beams)
+                _scores = _scores.view(batch_size, num_beams * vocab_size)      # (batch_size, num_beams * vocab_size)
+                next_scores, next_words = torch.topk(_scores, 2*num_beams, dim=1, largest=True, sorted=True)
+
+            assert next_scores.size() == next_words.size() == (batch_size, 2 * num_beams)
+
+            # next batch beam content
+            # list of (batch_size * num_beams) tuple(next hypothesis score, next word, current position in the batch)
+            next_batch_beam = []
+
+            # for each sentence
+            for batch_ex in range(batch_size):
+
+                # if we are done with this sentence
+                done[batch_ex] = done[batch_ex] or generated_hyps[batch_ex].is_done(next_scores[batch_ex].max().item())
+                if done[batch_ex]:
+                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
+                    continue
+
+                # next sentence beam content
+                next_sent_beam = []
+
+                # next words for this sentence
+                for idx, score in zip(next_words[batch_ex], next_scores[batch_ex]):
+
+                    # get beam and word IDs
+                    beam_id = idx // vocab_size
+                    word_id = idx % vocab_size
+
+                    # end of sentence, or next word
+                    if word_id.item() in eos_token_ids or cur_len + 1 == max_length:
+                        generated_hyps[batch_ex].add(input_ids[batch_ex * num_beams + beam_id, :cur_len].clone(), score.item())
+                    else:
+                        next_sent_beam.append((score, word_id, batch_ex * num_beams + beam_id))
+
+                    # the beam for next step is full
+                    if len(next_sent_beam) == num_beams:
+                        break
+
+                # update next beam content
+                assert len(next_sent_beam) == 0 if cur_len + 1 == max_length else num_beams
+                if len(next_sent_beam) == 0:
+                    next_sent_beam = [(0, pad_token_id, 0)] * num_beams  # pad the batch
+                next_batch_beam.extend(next_sent_beam)
+                assert len(next_batch_beam) == num_beams * (batch_ex + 1)
+
+            # sanity check / prepare next batch
+            assert len(next_batch_beam) == batch_size * num_beams
+            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
+            beam_words = input_ids.new([x[1] for x in next_batch_beam])
+            beam_idx = input_ids.new([x[2] for x in next_batch_beam])
+
+            # re-order batch and internal states
+            input_ids = input_ids[beam_idx, :]
+            input_ids = torch.cat([input_ids, beam_words.unsqueeze(1)], dim=-1)
+            # TODO: Activate cache
+            # for k in cache.keys():
+            #     if k != 'slen':
+            #         cache[k] = (cache[k][0][beam_idx], cache[k][1][beam_idx])
+
+            # update current length
+            cur_len = cur_len + 1
+
+            # stop when we are done with each sentence
+            if all(done):
+                break
+
+        # visualize hypotheses
+        # print([len(x) for x in generated_hyps], cur_len)
+        # globals().update( locals() );
+        # !import code; code.interact(local=vars())
+        # for ii in range(batch_size):
+        #     for ss, ww in sorted(generated_hyps[ii].hyp, key=lambda x: x[0], reverse=True):
+        #         print("%.3f " % ss + " ".join(self.dico[x] for x in ww.tolist()))
+        #     print("")
+
+        # select the best hypotheses
+        tgt_len = input_ids.new(batch_size)
+        best = []
+
+        for i, hypotheses in enumerate(generated_hyps):
+            best_hyp = max(hypotheses.hyp, key=lambda x: x[0])[1]
+            tgt_len[i] = len(best_hyp) + 1  # +1 for the <EOS> symbol
+            best.append(best_hyp)
+
+        # generate target batch
+        decoded = input_ids.new(batch_size, tgt_len.max().item()).fill_(pad_token_id)
+        for i, hypo in enumerate(best):
+            decoded[i, :tgt_len[i] - 1] = hypo
+            decoded[i, tgt_len[i] - 1] = eos_token_ids[0]
+
+        return decoded
+
+
+def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float('Inf'), min_tokens_to_keep=1):
+    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+        Args:
+            logits: logits distribution shape (batch size, vocabulary size)
+            if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+            if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+            Make sure we keep at least min_tokens_to_keep per batch example in the output
+        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    if top_k > 0:
+        top_k = min(max(top_k, min_tokens_to_keep), logits.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > top_p
+        if min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., :min_tokens_to_keep] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
+        logits[indices_to_remove] = filter_value
+    return logits
+
+
+class BeamHypotheses(object):
+
+    def __init__(self, n_hyp, max_length, length_penalty, early_stopping):
+        """
+        Initialize n-best list of hypotheses.
+        """
+        self.max_length = max_length - 1  # ignoring bos_token
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        self.n_hyp = n_hyp
+        self.hyp = []
+        self.worst_score = 1e9
+
+    def __len__(self):
+        """
+        Number of hypotheses in the list.
+        """
+        return len(self.hyp)
+
+    def add(self, hyp, sum_logprobs):
+        """
+        Add a new hypothesis to the list.
+        """
+        score = sum_logprobs / len(hyp) ** self.length_penalty
+        if len(self) < self.n_hyp or score > self.worst_score:
+            self.hyp.append((score, hyp))
+            if len(self) > self.n_hyp:
+                sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.hyp)])
+                del self.hyp[sorted_scores[0][1]]
+                self.worst_score = sorted_scores[1][0]
+            else:
+                self.worst_score = min(score, self.worst_score)
+
+    def is_done(self, best_sum_logprobs):
+        """
+        If there are enough hypotheses and that none of the hypotheses being generated
+        can become better than the worst one in the heap, then we are done with this sentence.
+        """
+        if len(self) < self.n_hyp:
+            return False
+        elif self.early_stopping:
+            return True
+        else:
+            return self.worst_score >= best_sum_logprobs / self.max_length ** self.length_penalty
+
 
 class Conv1D(nn.Module):
     def __init__(self, nf, nx):

transformers/modeling_xlm.py

@@ -346,7 +346,7 @@ class XLMModel(XLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = XLMModel.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -614,7 +614,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -634,7 +634,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = XLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -649,6 +649,18 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
     def get_output_embeddings(self):
         return self.pred_layer.proj
 
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        mask_token_id = self.config.mask_token_id
+        lang_id = self.config.lang_id
+
+        mask_token = torch.full((1, 1), mask_token_id, dtype=torch.long, device=input_ids.device)
+        input_ids = torch.cat([input_ids, mask_token], dim=1)
+        if lang_id is not None:
+            langs = torch.full_like(input_ids, lang_id)
+        else:
+            langs = None
+        return {"input_ids": input_ids, "langs": langs}
+
     def forward(self, input_ids=None, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
                 lengths=None, cache=None, head_mask=None, inputs_embeds=None, labels=None):
         transformer_outputs = self.transformer(input_ids,
@@ -696,7 +708,7 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = XLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -780,7 +792,7 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = XLMForQuestionAnsweringSimple.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
@@ -876,7 +888,7 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = XLMForQuestionAnswering.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/modeling_xlnet.py

@@ -589,7 +589,7 @@ class XLNetModel(XLNetPreTrainedModel):
 
         tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
         model = XLNetModel.from_pretrained('xlnet-large-cased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -898,7 +898,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -925,7 +925,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
         model = XLNetLMHeadModel.from_pretrained('xlnet-large-cased')
         # We show how to setup inputs to predict a next token using a bi-directional context.
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>")).unsqueeze(0)  # We will predict the masked token
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>", add_special_tokens=True)).unsqueeze(0)  # We will predict the masked token
         perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float)
         perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
         target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float)  # Shape [1, 1, seq_length] => let's predict one token
@@ -947,6 +947,30 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_loss
 
+    def prepare_inputs_for_generation(self, input_ids, **model_kwargs):
+        # Add dummy token at the end (no attention on this one)
+        dummy_token = torch.zeros((1, 1), dtype=torch.long, device=input_ids.device)
+        input_ids = torch.cat([input_ids, dummy_token], dim=1)
+
+        # Build permutation mask so that previous tokens don't see last token
+        perm_mask = torch.zeros(
+            (input_ids.shape[0], input_ids.shape[1], input_ids.shape[1]),
+            dtype=torch.float, device=input_ids.device
+        )
+        perm_mask[:, :, -1] = 1.0
+
+        # We'll only predict the last token
+        target_mapping = torch.zeros(
+            (input_ids.shape[0], 1, input_ids.shape[1]),
+            dtype=torch.float, device=input_ids.device
+        )
+        target_mapping[0, 0, -1] = 1.0
+
+        return {"input_ids": input_ids,
+                "perm_mask": perm_mask,
+                "target_mapping": target_mapping
+               }
+
     def forward(self, input_ids=None, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
                 token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None, labels=None):
         transformer_outputs = self.transformer(input_ids,
@@ -965,7 +989,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
 
         if labels is not None:
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, logits.size(-1)),
                             labels.view(-1))
             outputs = (loss,) + outputs
@@ -1007,7 +1031,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
 
         tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
         model = XLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -1294,7 +1318,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
 
         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
         model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
@@ -1409,7 +1433,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
 
         tokenizer =  XLNetTokenizer.from_pretrained('xlnet-large-cased')
         model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/tests/modeling_albert_test.py

@@ -17,13 +17,12 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (AlbertConfig, AlbertModel, AlbertForMaskedLM,
@@ -230,10 +229,8 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = AlbertModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = AlbertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_bert_test.py

@@ -17,13 +17,12 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor, floats_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (BertConfig, BertModel, BertForMaskedLM,
@@ -360,10 +359,8 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = BertModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = BertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_common_test.py

@@ -18,7 +18,7 @@ from __future__ import print_function
 
 import copy
 import sys
-import os
+import os.path
 import shutil
 import tempfile
 import json
@@ -30,7 +30,7 @@ import logging
 
 from transformers import is_torch_available
 
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 if is_torch_available():
     import torch
@@ -218,21 +218,22 @@ class CommonTestCases:
                 inputs = inputs_dict['input_ids']  # Let's keep only input_ids
 
                 try:
-                    torch.jit.trace(model, inputs)
+                    traced_gpt2 = torch.jit.trace(model, inputs)
                 except RuntimeError:
                     self.fail("Couldn't trace module.")
 
-                try:
-                    traced_gpt2 = torch.jit.trace(model, inputs)
-                    torch.jit.save(traced_gpt2, "traced_model.pt")
-                except RuntimeError:
-                    self.fail("Couldn't save module.")
+                with TemporaryDirectory() as tmp_dir_name:
+                    pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
 
-                try:
-                    loaded_model = torch.jit.load("traced_model.pt")
-                    os.remove("traced_model.pt")
-                except ValueError:
-                    self.fail("Couldn't load module.")
+                    try:
+                        torch.jit.save(traced_gpt2, pt_file_name)
+                    except Exception:
+                        self.fail("Couldn't save module.")
+
+                    try:
+                        loaded_model = torch.jit.load(pt_file_name)
+                    except Exception:
+                        self.fail("Couldn't load module.")
 
                 model.to(torch_device)
                 model.eval()
@@ -352,12 +353,11 @@ class CommonTestCases:
                 heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
                                 -1: [0]}
                 model.prune_heads(heads_to_prune)
-                directory = "pruned_model"
-                if not os.path.exists(directory):
-                    os.makedirs(directory)
-                model.save_pretrained(directory)
-                model = model_class.from_pretrained(directory)
-                model.to(torch_device)
+
+                with TemporaryDirectory() as temp_dir_name:
+                    model.save_pretrained(temp_dir_name)
+                    model = model_class.from_pretrained(temp_dir_name)
+                    model.to(torch_device)
 
                 with torch.no_grad():
                     outputs = model(**inputs_dict)
@@ -366,7 +366,6 @@ class CommonTestCases:
                 self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
                 self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
 
-                shutil.rmtree(directory)
 
         def test_head_pruning_save_load_from_config_init(self):
             if not self.test_pruning:
@@ -426,14 +425,10 @@ class CommonTestCases:
                 self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
                 self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
 
-                directory = "pruned_model"
-
-                if not os.path.exists(directory):
-                    os.makedirs(directory)
-                model.save_pretrained(directory)
-                model = model_class.from_pretrained(directory)
-                model.to(torch_device)
-                shutil.rmtree(directory)
+                with TemporaryDirectory() as temp_dir_name:
+                    model.save_pretrained(temp_dir_name)
+                    model = model_class.from_pretrained(temp_dir_name)
+                    model.to(torch_device)
 
                 with torch.no_grad():
                     outputs = model(**inputs_dict)
@@ -758,10 +753,8 @@ class CommonTestCases:
                 [[], []])
 
         def create_and_check_model_from_pretrained(self):
-            cache_dir = "/tmp/transformers_test/"
             for model_name in list(self.base_model_class.pretrained_model_archive_map.keys())[:1]:
-                model = self.base_model_class.from_pretrained(model_name, cache_dir=cache_dir)
-                shutil.rmtree(cache_dir)
+                model = self.base_model_class.from_pretrained(model_name, cache_dir=CACHE_DIR)
                 self.parent.assertIsNotNone(model)
 
         def prepare_config_and_inputs_for_common(self):

transformers/tests/modeling_ctrl_test.py

@@ -16,7 +16,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import pdb
 
 from transformers import is_torch_available
@@ -27,7 +26,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -205,10 +204,8 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(CTRL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = CTRLModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = CTRLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_distilbert_test.py

@@ -27,7 +27,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -235,10 +235,8 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
 
     # @slow
     # def test_model_from_pretrained(self):
-    #     cache_dir = "/tmp/transformers_test/"
     #     for model_name in list(DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-    #         model = DistilBertModel.from_pretrained(model_name, cache_dir=cache_dir)
-    #         shutil.rmtree(cache_dir)
+    #         model = DistilBertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
     #         self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_gpt2_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
@@ -27,7 +26,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -239,10 +238,8 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = GPT2Model.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = GPT2Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_openai_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
@@ -27,7 +26,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -207,10 +206,8 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = OpenAIGPTModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = OpenAIGPTModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_roberta_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
@@ -25,11 +24,12 @@ if is_torch_available():
     import torch
     from transformers import (RobertaConfig, RobertaModel, RobertaForMaskedLM,
                               RobertaForSequenceClassification, RobertaForTokenClassification)
+    from transformers.modeling_roberta import RobertaEmbeddings
     from transformers.modeling_roberta import ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -199,12 +199,61 @@ class RobertaModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = RobertaModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = RobertaModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
+    def test_create_position_ids_respects_padding_index(self):
+        """ Ensure that the default position ids only assign a sequential . This is a regression
+        test for https://github.com/huggingface/transformers/issues/1761
+
+        The position ids should be masked with the embedding object's padding index. Therefore, the
+        first available non-padding position index is RobertaEmbeddings.padding_idx + 1
+        """
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        model = RobertaEmbeddings(config=config)
+
+        input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
+        expected_positions = torch.as_tensor([[
+            0 + model.padding_idx + 1,
+            1 + model.padding_idx + 1,
+            2 + model.padding_idx + 1,
+            model.padding_idx
+        ]])
+
+        position_ids = model.create_position_ids_from_input_ids(input_ids)
+        self.assertEqual(
+            position_ids.shape,
+            expected_positions.shape
+        )
+        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
+
+    def test_create_position_ids_from_inputs_embeds(self):
+        """ Ensure that the default position ids only assign a sequential . This is a regression
+        test for https://github.com/huggingface/transformers/issues/1761
+
+        The position ids should be masked with the embedding object's padding index. Therefore, the
+        first available non-padding position index is RobertaEmbeddings.padding_idx + 1
+        """
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        embeddings = RobertaEmbeddings(config=config)
+
+        inputs_embeds = torch.Tensor(2, 4, 30)
+        expected_single_positions = [
+            0 + embeddings.padding_idx + 1,
+            1 + embeddings.padding_idx + 1,
+            2 + embeddings.padding_idx + 1,
+            3 + embeddings.padding_idx + 1,
+        ]
+        expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
+        position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
+        self.assertEqual(
+            position_ids.shape,
+            expected_positions.shape
+        )
+        self.assertTrue(
+            torch.all(torch.eq(position_ids, expected_positions))
+        )
 
 
 class RobertaModelIntegrationTest(unittest.TestCase):

transformers/tests/modeling_t5_test.py

@@ -17,13 +17,12 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor, floats_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (T5Config, T5Model, T5WithLMHeadModel)
@@ -175,10 +174,8 @@ class T5ModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(T5_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = T5Model.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = T5Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_albert_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import AlbertConfig, is_tf_available
 
@@ -217,12 +216,8 @@ class TFAlbertModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
-        # for model_name in list(TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-        for model_name in ['albert-base-uncased']:
-            model = TFAlbertModel.from_pretrained(
-                model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+        for model_name in list(TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = TFAlbertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_tf_auto_test.py

@@ -46,11 +46,11 @@ class TFAutoModelTest(unittest.TestCase):
         logging.basicConfig(level=logging.INFO)
         # for model_name in list(TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
         for model_name in ['bert-base-uncased']:
-            config = AutoConfig.from_pretrained(model_name, force_download=True)
+            config = AutoConfig.from_pretrained(model_name)
             self.assertIsNotNone(config)
             self.assertIsInstance(config, BertConfig)
 
-            model = TFAutoModel.from_pretrained(model_name, force_download=True)
+            model = TFAutoModel.from_pretrained(model_name)
             self.assertIsNotNone(model)
             self.assertIsInstance(model, TFBertModel)
 
@@ -59,11 +59,11 @@ class TFAutoModelTest(unittest.TestCase):
         logging.basicConfig(level=logging.INFO)
         # for model_name in list(TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
         for model_name in ['bert-base-uncased']:
-            config = AutoConfig.from_pretrained(model_name, force_download=True)
+            config = AutoConfig.from_pretrained(model_name)
             self.assertIsNotNone(config)
             self.assertIsInstance(config, BertConfig)
 
-            model = TFAutoModelWithLMHead.from_pretrained(model_name, force_download=True)
+            model = TFAutoModelWithLMHead.from_pretrained(model_name)
             self.assertIsNotNone(model)
             self.assertIsInstance(model, TFBertForMaskedLM)
 
@@ -72,11 +72,11 @@ class TFAutoModelTest(unittest.TestCase):
         logging.basicConfig(level=logging.INFO)
         # for model_name in list(TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
         for model_name in ['bert-base-uncased']:
-            config = AutoConfig.from_pretrained(model_name, force_download=True)
+            config = AutoConfig.from_pretrained(model_name)
             self.assertIsNotNone(config)
             self.assertIsInstance(config, BertConfig)
 
-            model = TFAutoModelForSequenceClassification.from_pretrained(model_name, force_download=True)
+            model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
             self.assertIsNotNone(model)
             self.assertIsInstance(model, TFBertForSequenceClassification)
 
@@ -85,17 +85,17 @@ class TFAutoModelTest(unittest.TestCase):
         logging.basicConfig(level=logging.INFO)
         # for model_name in list(TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
         for model_name in ['bert-base-uncased']:
-            config = AutoConfig.from_pretrained(model_name, force_download=True)
+            config = AutoConfig.from_pretrained(model_name)
             self.assertIsNotNone(config)
             self.assertIsInstance(config, BertConfig)
 
-            model = TFAutoModelForQuestionAnswering.from_pretrained(model_name, force_download=True)
+            model = TFAutoModelForQuestionAnswering.from_pretrained(model_name)
             self.assertIsNotNone(model)
             self.assertIsInstance(model, TFBertForQuestionAnswering)
 
     def test_from_pretrained_identifier(self):
         logging.basicConfig(level=logging.INFO)
-        model = TFAutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER, force_download=True)
+        model = TFAutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER)
         self.assertIsInstance(model, TFBertForMaskedLM)
 
 

transformers/tests/modeling_tf_bert_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import BertConfig, is_tf_available
 
@@ -310,11 +309,9 @@ class TFBertModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         # for model_name in list(TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
         for model_name in ['bert-base-uncased']:
-            model = TFBertModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFBertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_ctrl_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import CTRLConfig, is_tf_available
 
@@ -189,10 +188,8 @@ class TFCTRLModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFCTRLModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFCTRLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_distilbert_test.py

@@ -20,7 +20,7 @@ import unittest
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import DistilBertConfig, is_tf_available
 
@@ -211,10 +211,8 @@ class TFDistilBertModelTest(TFCommonTestCases.TFCommonModelTester):
 
     # @slow
     # def test_model_from_pretrained(self):
-    #     cache_dir = "/tmp/transformers_test/"
     #     for model_name in list(DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-    #         model = DistilBertModel.from_pretrained(model_name, cache_dir=cache_dir)
-    #         shutil.rmtree(cache_dir)
+    #         model = DistilBertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
     #         self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_gpt2_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import GPT2Config, is_tf_available
 
@@ -220,10 +219,8 @@ class TFGPT2ModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFGPT2Model.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFGPT2Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_openai_gpt_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import OpenAIGPTConfig, is_tf_available
 
@@ -219,10 +218,8 @@ class TFOpenAIGPTModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFOpenAIGPTModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFOpenAIGPTModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_roberta_test.py

@@ -17,11 +17,10 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import RobertaConfig, is_tf_available
 
@@ -192,10 +191,8 @@ class TFRobertaModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFRobertaModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFRobertaModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_tf_t5_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import T5Config, is_tf_available
 
@@ -162,10 +161,8 @@ class TFT5ModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in ['t5-small']:
-            model = TFT5Model.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFT5Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

transformers/tests/modeling_tf_transfo_xl_test.py

@@ -18,11 +18,10 @@ from __future__ import print_function
 
 import unittest
 import random
-import shutil
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import TransfoXLConfig, is_tf_available
 
@@ -205,10 +204,8 @@ class TFTransfoXLModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFTransfoXLModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFTransfoXLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_tf_xlm_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_tf_available
 
@@ -31,7 +30,7 @@ if is_tf_available():
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 
 @require_tf
@@ -252,10 +251,8 @@ class TFXLMModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XLMModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFXLMModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_tf_xlnet_test.py

@@ -20,7 +20,6 @@ import os
 import unittest
 import json
 import random
-import shutil
 
 from transformers import XLNetConfig, is_tf_available
 
@@ -35,7 +34,7 @@ if is_tf_available():
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_tf, slow
+from .utils import CACHE_DIR, require_tf, slow
 
 
 @require_tf
@@ -319,10 +318,8 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFXLNetModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFXLNetModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_transfo_xl_test.py

@@ -18,7 +18,6 @@ from __future__ import print_function
 
 import unittest
 import random
-import shutil
 
 from transformers import is_torch_available
 
@@ -29,7 +28,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -208,10 +207,8 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TransfoXLModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TransfoXLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_xlm_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
 
 from transformers import is_torch_available
 
@@ -28,7 +27,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -318,10 +317,8 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(XLM_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XLMModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = XLMModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_xlnet_test.py

@@ -20,7 +20,6 @@ import os
 import unittest
 import json
 import random
-import shutil
 
 from transformers import is_torch_available
 
@@ -33,7 +32,7 @@ if is_torch_available():
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
-from .utils import require_torch, slow, torch_device
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 @require_torch
@@ -385,10 +384,8 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
 
     @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(XLNET_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XLNetModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = XLNetModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/utils.py

@@ -1,11 +1,14 @@
 import os
 import unittest
+import tempfile
 
 from distutils.util import strtobool
 
 from transformers.file_utils import _tf_available, _torch_available
 
 
+CACHE_DIR = os.path.join(tempfile.gettempdir(), "transformers_test")
+
 SMALL_MODEL_IDENTIFIER = "julien-c/bert-xsmall-dummy"
 
 

transformers/tokenization_utils.py

@@ -231,6 +231,7 @@ class PreTrainedTokenizer(object):
         
         # Added tokens
         self.added_tokens_encoder = {}
+        self.unique_added_tokens_encoder = set()
         self.added_tokens_decoder = {}
 
         # inputs and kwargs for saving and re-loading (see ``from_pretrained`` and ``save_pretrained``)
@@ -554,6 +555,7 @@ class PreTrainedTokenizer(object):
         added_tok_encoder = dict((tok, len(self) + i) for i, tok in enumerate(to_add_tokens))
         added_tok_decoder = {v:k for k, v in added_tok_encoder.items()}
         self.added_tokens_encoder.update(added_tok_encoder)
+        self.unique_added_tokens_encoder = set(self.added_tokens_encoder.keys()).union(set(self.all_special_tokens))
         self.added_tokens_decoder.update(added_tok_decoder)
 
         return len(to_add_tokens)
@@ -631,6 +633,7 @@ class PreTrainedTokenizer(object):
 
         return added_tokens
 
+
     def tokenize(self, text, **kwargs):
         """ Converts a string in a sequence of tokens (string), using the tokenizer.
             Split in words for word-based vocabulary or sub-words for sub-word-based
@@ -685,18 +688,17 @@ class PreTrainedTokenizer(object):
             for tok in tok_list:
                 tokenized_text = []
                 for sub_text in text_list:
-                    if sub_text not in self.added_tokens_encoder \
-                            and sub_text not in all_special_tokens:
+                    if sub_text not in self.unique_added_tokens_encoder:
                         tokenized_text += split_on_token(tok, sub_text)
                     else:
                         tokenized_text += [sub_text]
                 text_list = tokenized_text
 
-            return list(itertools.chain.from_iterable((self._tokenize(token, **kwargs) if token not \
-                    in self.added_tokens_encoder and token not in all_special_tokens \
+            return list(itertools.chain.from_iterable((self._tokenize(token, **kwargs) \
+                    if token not in self.unique_added_tokens_encoder
                     else [token] for token in tokenized_text)))
 
-        added_tokens = list(self.added_tokens_encoder.keys()) + all_special_tokens
+        added_tokens = self.unique_added_tokens_encoder
         tokenized_text = split_on_tokens(added_tokens, text)
         return tokenized_text
 
@@ -773,7 +775,7 @@ class PreTrainedTokenizer(object):
                 padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
                 The tokenizer padding sides are handled by the following strings:
                 - 'left': pads on the left of the sequences
-                - 'right': pads on the right of the sequences   
+                - 'right': pads on the right of the sequences
                 Defaults to False: no padding.
             return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
                 or PyTorch torch.Tensor instead of a list of python integers.