Merge remote-tracking branch 'refs/remotes/huggingface/master'

transformers/configuration_gpt2.py

@@ -29,6 +29,7 @@ logger = logging.getLogger(__name__)
 GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.json",
                                       "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-config.json",
                                       "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-config.json",
+                                      "gpt2-xl": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-xl-config.json",
                                       "distilgpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-config.json",}
 
 class GPT2Config(PretrainedConfig):

transformers/configuration_roberta.py

@@ -29,6 +29,8 @@ ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-config.json",
     'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-config.json",
     'distilroberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/distilroberta-base-config.json",
+    'roberta-base-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-openai-detector-config.json",
+    'roberta-large-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-openai-detector-config.json",
 }
 
 

transformers/configuration_utils.py

@@ -24,7 +24,7 @@ import logging
 import os
 from io import open
 
-from .file_utils import cached_path, CONFIG_NAME
+from .file_utils import CONFIG_NAME, cached_path, is_remote_url, hf_bucket_url
 
 logger = logging.getLogger(__name__)
 
@@ -57,6 +57,7 @@ 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', {})
+        self.is_decoder = kwargs.pop('is_decoder', False)
 
     def save_pretrained(self, save_directory):
         """ Save a configuration object to the directory `save_directory`, so that it
@@ -78,6 +79,7 @@ class PretrainedConfig(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model configuration that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing a configuration file saved using the :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
                 - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.
 
@@ -93,6 +95,9 @@ class PretrainedConfig(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
             proxies: (`optional`) dict, default None:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
@@ -119,6 +124,7 @@ class PretrainedConfig(object):
         """
         cache_dir = kwargs.pop('cache_dir', None)
         force_download = kwargs.pop('force_download', False)
+        resume_download = kwargs.pop('resume_download', False)
         proxies = kwargs.pop('proxies', None)
         return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
 
@@ -126,11 +132,14 @@ class PretrainedConfig(object):
             config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
         elif os.path.isdir(pretrained_model_name_or_path):
             config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
-        else:
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
             config_file = pretrained_model_name_or_path
+        else:
+            config_file = hf_bucket_url(pretrained_model_name_or_path, postfix=CONFIG_NAME)
         # redirect to the cache, if necessary
         try:
-            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download,
+                                               proxies=proxies, resume_download=resume_download)
         except EnvironmentError:
             if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
                 msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
@@ -181,7 +190,7 @@ class PretrainedConfig(object):
 
     @classmethod
     def from_json_file(cls, json_file):
-        """Constructs a `BertConfig` from a json file of parameters."""
+        """Constructs a `Config` from a json file of parameters."""
         with open(json_file, "r", encoding='utf-8') as reader:
             text = reader.read()
         return cls.from_dict(json.loads(text))

transformers/convert_albert_original_tf_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ALBERT checkpoint."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import torch
+
+from transformers import AlbertConfig, AlbertForMaskedLM, load_tf_weights_in_albert
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, albert_config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = AlbertConfig.from_json_file(albert_config_file)
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    model = AlbertForMaskedLM(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_albert(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--tf_checkpoint_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--albert_config_file",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "The config json file corresponding to the pre-trained ALBERT model. \n"
+                            "This specifies the model architecture.")
+    parser.add_argument("--pytorch_dump_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
+                                     args.albert_config_file,
+                                     args.pytorch_dump_path)
+ 
\ No newline at end of file

transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -33,7 +33,8 @@ from transformers import (load_pytorch_checkpoint_in_tf2_model,
                                   OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
+                                  CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
 
 if is_torch_available():
     import torch
@@ -46,7 +47,8 @@ if is_torch_available():
                                       OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
-                                      CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
+                                      CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+                                      AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 else:
     (BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -56,7 +58,8 @@ else:
     OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
     RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
-    CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP) = (
+    CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+    AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP) = (
         None, None, None, None,
         None, None,
         None, None,
@@ -65,6 +68,7 @@ else:
         None, None,
         None, None, None,
         None, None, None,
+        None, None,
         None, None)
 
 
@@ -85,7 +89,8 @@ MODEL_CLASSES = {
     'roberta-large-mnli': (RobertaConfig, TFRobertaForSequenceClassification, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'ctrl': (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
+    'ctrl': (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
 }
 
 def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file, tf_dump_path, compare_with_pt_model=False, use_cached_models=True):
@@ -114,10 +119,11 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
         tf_inputs = tf.constant(inputs_list)
         tfo = tf_model(tf_inputs, training=False)  # build the network
 
-        pt_model = pt_model_class.from_pretrained(None,
+        state_dict = torch.load(pytorch_checkpoint_path, map_location='cpu')
+        pt_model = pt_model_class.from_pretrained(pretrained_model_name_or_path=None,
                                                   config=config,
-                                                  state_dict=torch.load(pytorch_checkpoint_path,
-                                                                        map_location='cpu'))
+                                                  state_dict=state_dict)
+
         pt_inputs = torch.tensor(inputs_list)
         with torch.no_grad():
             pto = pt_model(pt_inputs)
@@ -134,7 +140,7 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
 
 
 def convert_all_pt_checkpoints_to_tf(args_model_type, tf_dump_path, model_shortcut_names_or_path=None, config_shortcut_names_or_path=None,
-                                     compare_with_pt_model=False, use_cached_models=False, only_convert_finetuned_models=False):
+                                     compare_with_pt_model=False, use_cached_models=False, remove_cached_files=False, only_convert_finetuned_models=False):
     assert os.path.isdir(args.tf_dump_path), "--tf_dump_path should be a directory"
 
     if args_model_type is None:
@@ -182,13 +188,15 @@ def convert_all_pt_checkpoints_to_tf(args_model_type, tf_dump_path, model_shortc
 
             if os.path.isfile(model_shortcut_name):
                 model_shortcut_name = 'converted_model'
+
             convert_pt_checkpoint_to_tf(model_type=model_type,
                                         pytorch_checkpoint_path=model_file,
                                         config_file=config_file,
                                         tf_dump_path=os.path.join(tf_dump_path, model_shortcut_name + '-tf_model.h5'),
                                         compare_with_pt_model=compare_with_pt_model)
-            os.remove(config_file)
-            os.remove(model_file)
+            if remove_cached_files:
+                os.remove(config_file)
+                os.remove(model_file)
 
 
 if __name__ == "__main__":
@@ -221,6 +229,9 @@ if __name__ == "__main__":
     parser.add_argument("--use_cached_models",
                         action='store_true',
                         help = "Use cached models if possible instead of updating to latest checkpoint versions.")
+    parser.add_argument("--remove_cached_files",
+                        action='store_true',
+                        help = "Remove pytorch models after conversion (save memory when converting in batches).")
     parser.add_argument("--only_convert_finetuned_models",
                         action='store_true',
                         help = "Only convert finetuned models.")
@@ -240,4 +251,5 @@ if __name__ == "__main__":
                                         config_shortcut_names_or_path=[args.config_file] if args.config_file is not None else None,
                                         compare_with_pt_model=args.compare_with_pt_model,
                                         use_cached_models=args.use_cached_models,
+                                        remove_cached_files=args.remove_cached_files,
                                         only_convert_finetuned_models=args.only_convert_finetuned_models)

transformers/data/__init__.py

-from .processors import InputExample, InputFeatures, DataProcessor
+from .processors import InputExample, InputFeatures, DataProcessor, SquadFeatures
 from .processors import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
+from .processors import squad_convert_examples_to_features, SquadExample, SquadV1Processor, SquadV2Processor
+from .processors import xnli_output_modes, xnli_processors, xnli_tasks_num_labels
 
 from .metrics import is_sklearn_available
 if is_sklearn_available():
-    from .metrics import glue_compute_metrics
+    from .metrics import glue_compute_metrics, xnli_compute_metrics

transformers/data/metrics/__init__.py

@@ -81,3 +81,11 @@ if _has_sklearn:
             return {"acc": simple_accuracy(preds, labels)}
         else:
             raise KeyError(task_name)
+
+
+    def xnli_compute_metrics(task_name, preds, labels):
+        assert len(preds) == len(labels)
+        if task_name == "xnli":
+            return {"acc": simple_accuracy(preds, labels)}
+        else:
+            raise KeyError(task_name)

transformers/data/metrics/squad_metrics.py

+""" Very heavily inspired by the official evaluation script for SQuAD version 2.0 which was
+modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
+
+In addition to basic functionality, we also compute additional statistics and
+plot precision-recall curves if an additional na_prob.json file is provided.
+This file is expected to map question ID's to the model's predicted probability
+that a question is unanswerable.
+"""
+
+
+import json
+import logging
+import math
+import collections
+from io import open
+from tqdm import tqdm
+import string
+import re
+
+from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
+
+logger = logging.getLogger(__name__)
+
+
+def normalize_answer(s):
+    """Lower text and remove punctuation, articles and extra whitespace."""
+    def remove_articles(text):
+        regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
+        return re.sub(regex, ' ', text)
+
+    def white_space_fix(text):
+        return ' '.join(text.split())
+
+    def remove_punc(text):
+        exclude = set(string.punctuation)
+        return ''.join(ch for ch in text if ch not in exclude)
+
+    def lower(text):
+        return text.lower()
+    return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+
+def get_tokens(s):
+    if not s:
+        return []
+    return normalize_answer(s).split()
+
+
+def compute_exact(a_gold, a_pred):
+    return int(normalize_answer(a_gold) == normalize_answer(a_pred))
+
+
+def compute_f1(a_gold, a_pred):
+    gold_toks = get_tokens(a_gold)
+    pred_toks = get_tokens(a_pred)
+    common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
+    num_same = sum(common.values())
+    if len(gold_toks) == 0 or len(pred_toks) == 0:
+        # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
+        return int(gold_toks == pred_toks)
+    if num_same == 0:
+        return 0
+    precision = 1.0 * num_same / len(pred_toks)
+    recall = 1.0 * num_same / len(gold_toks)
+    f1 = (2 * precision * recall) / (precision + recall)
+    return f1
+
+
+def get_raw_scores(examples, preds):
+    """
+    Computes the exact and f1 scores from the examples and the model predictions
+    """
+    exact_scores = {}
+    f1_scores = {}
+
+    for example in examples:
+        qas_id = example.qas_id
+        gold_answers = [answer['text'] for answer in example.answers if normalize_answer(answer['text'])]
+
+        if not gold_answers:
+            # For unanswerable questions, only correct answer is empty string
+            gold_answers = ['']
+
+        if qas_id not in preds:
+            print('Missing prediction for %s' % qas_id)
+            continue
+
+        prediction = preds[qas_id]
+        exact_scores[qas_id] = max(compute_exact(a, prediction) for a in gold_answers)
+        f1_scores[qas_id] = max(compute_f1(a, prediction) for a in gold_answers)
+
+    return exact_scores, f1_scores
+
+
+def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
+    new_scores = {}
+    for qid, s in scores.items():
+        pred_na = na_probs[qid] > na_prob_thresh
+        if pred_na:
+            new_scores[qid] = float(not qid_to_has_ans[qid])
+        else:
+            new_scores[qid] = s
+    return new_scores
+
+
+def make_eval_dict(exact_scores, f1_scores, qid_list=None):
+    if not qid_list:
+        total = len(exact_scores)
+        return collections.OrderedDict([
+            ('exact', 100.0 * sum(exact_scores.values()) / total),
+            ('f1', 100.0 * sum(f1_scores.values()) / total),
+            ('total', total),
+        ])
+    else:
+        total = len(qid_list)
+        return collections.OrderedDict([
+            ('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
+            ('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
+            ('total', total),
+        ])
+
+
+def merge_eval(main_eval, new_eval, prefix):
+    for k in new_eval:
+        main_eval['%s_%s' % (prefix, k)] = new_eval[k]
+
+
+def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
+    num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+    cur_score = num_no_ans
+    best_score = cur_score
+    best_thresh = 0.0
+    qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+    for i, qid in enumerate(qid_list):
+        if qid not in scores:
+            continue
+        if qid_to_has_ans[qid]:
+            diff = scores[qid]
+        else:
+            if preds[qid]:
+                diff = -1
+            else:
+                diff = 0
+        cur_score += diff
+        if cur_score > best_score:
+            best_score = cur_score
+            best_thresh = na_probs[qid]
+
+    has_ans_score, has_ans_cnt = 0, 0
+    for qid in qid_list:
+        if not qid_to_has_ans[qid]:
+            continue
+        has_ans_cnt += 1
+
+        if qid not in scores:
+            continue
+        has_ans_score += scores[qid]
+
+    return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
+
+
+def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+    best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(
+        preds, exact_raw, na_probs, qid_to_has_ans)
+    best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(
+        preds, f1_raw, na_probs, qid_to_has_ans)
+    main_eval['best_exact'] = best_exact
+    main_eval['best_exact_thresh'] = exact_thresh
+    main_eval['best_f1'] = best_f1
+    main_eval['best_f1_thresh'] = f1_thresh
+    main_eval['has_ans_exact'] = has_ans_exact
+    main_eval['has_ans_f1'] = has_ans_f1
+
+
+def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
+    num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+    cur_score = num_no_ans
+    best_score = cur_score
+    best_thresh = 0.0
+    qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+    for _, qid in enumerate(qid_list):
+        if qid not in scores:
+            continue
+        if qid_to_has_ans[qid]:
+            diff = scores[qid]
+        else:
+            if preds[qid]:
+                diff = -1
+            else:
+                diff = 0
+        cur_score += diff
+        if cur_score > best_score:
+            best_score = cur_score
+            best_thresh = na_probs[qid]
+    return 100.0 * best_score / len(scores), best_thresh
+
+
+def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+    best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
+    best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
+
+    main_eval['best_exact'] = best_exact
+    main_eval['best_exact_thresh'] = exact_thresh
+    main_eval['best_f1'] = best_f1
+    main_eval['best_f1_thresh'] = f1_thresh
+
+
+def squad_evaluate(examples, preds, no_answer_probs=None, no_answer_probability_threshold=1.0):
+    qas_id_to_has_answer = {example.qas_id: bool(example.answers) for example in examples}
+    has_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if has_answer]
+    no_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if not has_answer]
+
+    if no_answer_probs is None:
+        no_answer_probs = {k: 0.0 for k in preds}
+
+    exact, f1 = get_raw_scores(examples, preds)
+
+    exact_threshold = apply_no_ans_threshold(exact, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
+    f1_threshold = apply_no_ans_threshold(f1, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
+
+    evaluation = make_eval_dict(exact_threshold, f1_threshold)
+
+    if has_answer_qids:
+        has_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=has_answer_qids)
+        merge_eval(evaluation, has_ans_eval, 'HasAns')
+
+    if no_answer_qids:
+        no_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=no_answer_qids)
+        merge_eval(evaluation, no_ans_eval, 'NoAns')
+
+    if no_answer_probs:
+        find_all_best_thresh(evaluation, preds, exact, f1, no_answer_probs, qas_id_to_has_answer)
+
+    return evaluation
+
+
+def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False):
+    """Project the tokenized prediction back to the original text."""
+
+    # When we created the data, we kept track of the alignment between original
+    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
+    # now `orig_text` contains the span of our original text corresponding to the
+    # span that we predicted.
+    #
+    # However, `orig_text` may contain extra characters that we don't want in
+    # our prediction.
+    #
+    # For example, let's say:
+    #   pred_text = steve smith
+    #   orig_text = Steve Smith's
+    #
+    # We don't want to return `orig_text` because it contains the extra "'s".
+    #
+    # We don't want to return `pred_text` because it's already been normalized
+    # (the SQuAD eval script also does punctuation stripping/lower casing but
+    # our tokenizer does additional normalization like stripping accent
+    # characters).
+    #
+    # What we really want to return is "Steve Smith".
+    #
+    # Therefore, we have to apply a semi-complicated alignment heuristic between
+    # `pred_text` and `orig_text` to get a character-to-character alignment. This
+    # can fail in certain cases in which case we just return `orig_text`.
+
+    def _strip_spaces(text):
+        ns_chars = []
+        ns_to_s_map = collections.OrderedDict()
+        for (i, c) in enumerate(text):
+            if c == " ":
+                continue
+            ns_to_s_map[len(ns_chars)] = i
+            ns_chars.append(c)
+        ns_text = "".join(ns_chars)
+        return (ns_text, ns_to_s_map)
+
+    # We first tokenize `orig_text`, strip whitespace from the result
+    # and `pred_text`, and check if they are the same length. If they are
+    # NOT the same length, the heuristic has failed. If they are the same
+    # length, we assume the characters are one-to-one aligned.
+    tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+
+    tok_text = " ".join(tokenizer.tokenize(orig_text))
+
+    start_position = tok_text.find(pred_text)
+    if start_position == -1:
+        if verbose_logging:
+            logger.info(
+                "Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
+        return orig_text
+    end_position = start_position + len(pred_text) - 1
+
+    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
+    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
+
+    if len(orig_ns_text) != len(tok_ns_text):
+        if verbose_logging:
+            logger.info("Length not equal after stripping spaces: '%s' vs '%s'",
+                        orig_ns_text, tok_ns_text)
+        return orig_text
+
+    # We then project the characters in `pred_text` back to `orig_text` using
+    # the character-to-character alignment.
+    tok_s_to_ns_map = {}
+    for (i, tok_index) in tok_ns_to_s_map.items():
+        tok_s_to_ns_map[tok_index] = i
+
+    orig_start_position = None
+    if start_position in tok_s_to_ns_map:
+        ns_start_position = tok_s_to_ns_map[start_position]
+        if ns_start_position in orig_ns_to_s_map:
+            orig_start_position = orig_ns_to_s_map[ns_start_position]
+
+    if orig_start_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map start position")
+        return orig_text
+
+    orig_end_position = None
+    if end_position in tok_s_to_ns_map:
+        ns_end_position = tok_s_to_ns_map[end_position]
+        if ns_end_position in orig_ns_to_s_map:
+            orig_end_position = orig_ns_to_s_map[ns_end_position]
+
+    if orig_end_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map end position")
+        return orig_text
+
+    output_text = orig_text[orig_start_position:(orig_end_position + 1)]
+    return output_text
+
+
+def _get_best_indexes(logits, n_best_size):
+    """Get the n-best logits from a list."""
+    index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
+
+    best_indexes = []
+    for i in range(len(index_and_score)):
+        if i >= n_best_size:
+            break
+        best_indexes.append(index_and_score[i][0])
+    return best_indexes
+
+
+def _compute_softmax(scores):
+    """Compute softmax probability over raw logits."""
+    if not scores:
+        return []
+
+    max_score = None
+    for score in scores:
+        if max_score is None or score > max_score:
+            max_score = score
+
+    exp_scores = []
+    total_sum = 0.0
+    for score in scores:
+        x = math.exp(score - max_score)
+        exp_scores.append(x)
+        total_sum += x
+
+    probs = []
+    for score in exp_scores:
+        probs.append(score / total_sum)
+    return probs
+
+
+def compute_predictions_logits(
+    all_examples,
+    all_features,
+    all_results,
+    n_best_size,
+    max_answer_length,
+    do_lower_case,
+    output_prediction_file,
+    output_nbest_file,
+    output_null_log_odds_file,
+    verbose_logging,
+    version_2_with_negative,
+    null_score_diff_threshold
+):
+    """Write final predictions to the json file and log-odds of null if needed."""
+    logger.info("Writing predictions to: %s" % (output_prediction_file))
+    logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+        min_null_feature_index = 0  # the paragraph slice with min null score
+        null_start_logit = 0  # the start logit at the slice with min null score
+        null_end_logit = 0  # the end logit at the slice with min null score
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
+            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
+            # if we could have irrelevant answers, get the min score of irrelevant
+            if version_2_with_negative:
+                feature_null_score = result.start_logits[0] + result.end_logits[0]
+                if feature_null_score < score_null:
+                    score_null = feature_null_score
+                    min_null_feature_index = feature_index
+                    null_start_logit = result.start_logits[0]
+                    null_end_logit = result.end_logits[0]
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= len(feature.tokens):
+                        continue
+                    if end_index >= len(feature.tokens):
+                        continue
+                    if start_index not in feature.token_to_orig_map:
+                        continue
+                    if end_index not in feature.token_to_orig_map:
+                        continue
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_logit=result.start_logits[start_index],
+                            end_logit=result.end_logits[end_index]))
+        if version_2_with_negative:
+            prelim_predictions.append(
+                _PrelimPrediction(
+                    feature_index=min_null_feature_index,
+                    start_index=0,
+                    end_index=0,
+                    start_logit=null_start_logit,
+                    end_logit=null_end_logit))
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_logit + x.end_logit),
+            reverse=True)
+
+        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+            "NbestPrediction", ["text", "start_logit", "end_logit"])
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+            if pred.start_index > 0:  # this is a non-null prediction
+                tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+                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("##", "")
+
+                # Clean whitespace
+                tok_text = tok_text.strip()
+                tok_text = " ".join(tok_text.split())
+                orig_text = " ".join(orig_tokens)
+
+                final_text = get_final_text(tok_text, orig_text, do_lower_case, verbose_logging)
+                if final_text in seen_predictions:
+                    continue
+
+                seen_predictions[final_text] = True
+            else:
+                final_text = ""
+                seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_logit=pred.start_logit,
+                    end_logit=pred.end_logit))
+        # if we didn't include the empty option in the n-best, include it
+        if version_2_with_negative:
+            if "" not in seen_predictions:
+                nbest.append(
+                    _NbestPrediction(
+                        text="",
+                        start_logit=null_start_logit,
+                        end_logit=null_end_logit))
+
+            # In very rare edge cases we could only have single null prediction.
+            # So we just create a nonce prediction in this case to avoid failure.
+            if len(nbest) == 1:
+                nbest.insert(0,
+                             _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        assert len(nbest) >= 1
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_logit + entry.end_logit)
+            if not best_non_null_entry:
+                if entry.text:
+                    best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_logit"] = entry.start_logit
+            output["end_logit"] = entry.end_logit
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+
+        if not version_2_with_negative:
+            all_predictions[example.qas_id] = nbest_json[0]["text"]
+        else:
+            # predict "" iff the null score - the score of best non-null > threshold
+            score_diff = score_null - best_non_null_entry.start_logit - (
+                best_non_null_entry.end_logit)
+            scores_diff_json[example.qas_id] = score_diff
+            if score_diff > null_score_diff_threshold:
+                all_predictions[example.qas_id] = ""
+            else:
+                all_predictions[example.qas_id] = best_non_null_entry.text
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions
+
+
+def compute_predictions_log_probs(
+    all_examples,
+    all_features,
+    all_results,
+    n_best_size,
+    max_answer_length,
+    output_prediction_file,
+    output_nbest_file,
+    output_null_log_odds_file,
+    start_n_top,
+    end_n_top,
+    version_2_with_negative,
+    tokenizer,
+    verbose_logging
+):
+    """ XLNet write prediction logic (more complex than Bert's).
+        Write final predictions to the json file and log-odds of null if needed.
+
+        Requires utils_squad_evaluate.py
+    """
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index",
+         "start_log_prob", "end_log_prob"])
+
+    _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "NbestPrediction", ["text", "start_log_prob", "end_log_prob"])
+
+    logger.info("Writing predictions to: %s", output_prediction_file)
+    # logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+
+            cur_null_score = result.cls_logits
+
+            # if we could have irrelevant answers, get the min score of irrelevant
+            score_null = min(score_null, cur_null_score)
+
+            for i in range(start_n_top):
+                for j in range(end_n_top):
+                    start_log_prob = result.start_logits[i]
+                    start_index = result.start_top_index[i]
+
+                    j_index = i * end_n_top + j
+
+                    end_log_prob = result.end_logits[j_index]
+                    end_index = result.end_top_index[j_index]
+
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= feature.paragraph_len - 1:
+                        continue
+                    if end_index >= feature.paragraph_len - 1:
+                        continue
+
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_log_prob=start_log_prob,
+                            end_log_prob=end_log_prob))
+
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_log_prob + x.end_log_prob),
+            reverse=True)
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+
+            # XLNet un-tokenizer
+            # Let's keep it simple for now and see if we need all this later.
+            #
+            # tok_start_to_orig_index = feature.tok_start_to_orig_index
+            # tok_end_to_orig_index = feature.tok_end_to_orig_index
+            # start_orig_pos = tok_start_to_orig_index[pred.start_index]
+            # end_orig_pos = tok_end_to_orig_index[pred.end_index]
+            # paragraph_text = example.paragraph_text
+            # final_text = paragraph_text[start_orig_pos: end_orig_pos + 1].strip()
+
+            # Previously used Bert untokenizer
+            tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+            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 = tokenizer.convert_tokens_to_string(tok_tokens)
+
+            # Clean whitespace
+            tok_text = tok_text.strip()
+            tok_text = " ".join(tok_text.split())
+            orig_text = " ".join(orig_tokens)
+
+            if hasattr(tokenizer, "do_lower_case"):
+                do_lower_case = tokenizer.do_lower_case
+            else:
+                do_lower_case = tokenizer.do_lowercase_and_remove_accent
+
+            final_text = get_final_text(tok_text, orig_text, do_lower_case,
+                                        verbose_logging)
+
+            if final_text in seen_predictions:
+                continue
+
+            seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_log_prob=pred.start_log_prob,
+                    end_log_prob=pred.end_log_prob))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="", start_log_prob=-1e6,
+                end_log_prob=-1e6))
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_log_prob + entry.end_log_prob)
+            if not best_non_null_entry:
+                best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_log_prob"] = entry.start_log_prob
+            output["end_log_prob"] = entry.end_log_prob
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+        assert best_non_null_entry is not None
+
+        score_diff = score_null
+        scores_diff_json[example.qas_id] = score_diff
+        # note(zhiliny): always predict best_non_null_entry
+        # and the evaluation script will search for the best threshold
+        all_predictions[example.qas_id] = best_non_null_entry.text
+
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions

transformers/data/processors/__init__.py

 from .utils import InputExample, InputFeatures, DataProcessor
 from .glue import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
-
+from .squad import squad_convert_examples_to_features, SquadFeatures, SquadExample, SquadV1Processor, SquadV2Processor
+from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels
\ No newline at end of file

transformers/data/processors/glue.py

@@ -80,6 +80,7 @@ def glue_convert_examples_to_features(examples, tokenizer,
             logger.info("Writing example %d" % (ex_index))
         if is_tf_dataset:
             example = processor.get_example_from_tensor_dict(example)
+            example = processor.tfds_map(example)
 
         inputs = tokenizer.encode_plus(
             example.text_a,

transformers/data/processors/squad.py

+from tqdm import tqdm
+import collections
+import logging
+import os
+import json
+import numpy as np
+
+from ...tokenization_bert import BasicTokenizer, whitespace_tokenize
+from .utils import DataProcessor, InputExample, InputFeatures
+from ...file_utils import is_tf_available, is_torch_available
+
+if is_torch_available():
+    import torch
+    from torch.utils.data import TensorDataset
+
+if is_tf_available():
+    import tensorflow as tf
+
+logger = logging.getLogger(__name__)
+
+def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
+                        orig_answer_text):
+    """Returns tokenized answer spans that better match the annotated answer."""
+    tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
+
+    for new_start in range(input_start, input_end + 1):
+        for new_end in range(input_end, new_start - 1, -1):
+            text_span = " ".join(doc_tokens[new_start:(new_end + 1)])
+            if text_span == tok_answer_text:
+                return (new_start, new_end)
+
+    return (input_start, input_end)
+
+def _check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+    best_score = None
+    best_span_index = None
+    for (span_index, doc_span) in enumerate(doc_spans):
+        end = doc_span.start + doc_span.length - 1
+        if position < doc_span.start:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span.start
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + 0.01 * doc_span.length
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+def _new_check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+    # if len(doc_spans) == 1:
+        # return True
+    best_score = None
+    best_span_index = None
+    for (span_index, doc_span) in enumerate(doc_spans):
+        end = doc_span["start"] + doc_span["length"] - 1
+        if position < doc_span["start"]:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span["start"]
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + 0.01 * doc_span["length"]
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+def _is_whitespace(c):
+    if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
+        return True
+    return False
+
+def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
+                                       doc_stride, max_query_length, is_training, 
+                                       return_dataset=False):
+    """
+    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.
+
+    Args:
+        examples: list of :class:`~transformers.data.processors.squad.SquadExample`
+        tokenizer: an instance of a child of :class:`~transformers.PreTrainedTokenizer`
+        max_seq_length: The maximum sequence length of the inputs.
+        doc_stride: The stride used when the context is too large and is split across several features.
+        max_query_length: The maximum length of the query.
+        is_training: whether to create features for model evaluation or model training.
+        return_dataset: Default False. Either 'pt' or 'tf'.
+            if 'pt': returns a torch.data.TensorDataset,
+            if 'tf': returns a tf.data.Dataset
+
+    Returns:
+        list of :class:`~transformers.data.processors.squad.SquadFeatures`
+
+    Example::
+
+        processor = SquadV2Processor()
+        examples = processor.get_dev_examples(data_dir)
+
+        features = 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,
+        )
+    """
+
+    # Defining helper methods    
+    unique_id = 1000000000
+
+    features = []
+    for (example_index, example) in enumerate(tqdm(examples)):
+        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
+            ))
+
+            unique_id += 1
+
+    if return_dataset == 'pt':
+        if not is_torch_available():
+            raise ImportError("Pytorch must be installed to return a pytorch dataset.")
+
+        # Convert to Tensors and build dataset
+        all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+        all_input_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
+        all_segment_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
+        all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
+        all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
+
+        if not is_training:
+            all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
+            dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                    all_example_index, all_cls_index, all_p_mask)
+        else:
+            all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
+            all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
+            dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                    all_start_positions, all_end_positions,
+                                    all_cls_index, all_p_mask)
+
+        return features, dataset
+        
+
+    return features
+
+
+class SquadProcessor(DataProcessor):
+    """
+    Processor for the SQuAD data set.
+    Overriden by SquadV1Processor and SquadV2Processor, used by the version 1.1 and version 2.0 of SQuAD, respectively.
+    """
+    train_file = None
+    dev_file = None
+
+    def _get_example_from_tensor_dict(self, tensor_dict, evaluate=False):
+        if not evaluate:
+            answer = tensor_dict['answers']['text'][0].numpy().decode('utf-8')
+            answer_start = tensor_dict['answers']['answer_start'][0].numpy()
+            answers = []
+        else:
+            answers = [{
+                "answer_start": start.numpy(), 
+                "text": text.numpy().decode('utf-8')
+            } for start, text in zip(tensor_dict['answers']["answer_start"], tensor_dict['answers']["text"])]
+
+            answer = None
+            answer_start = None
+
+        return SquadExample(
+            qas_id=tensor_dict['id'].numpy().decode("utf-8"),
+            question_text=tensor_dict['question'].numpy().decode('utf-8'),
+            context_text=tensor_dict['context'].numpy().decode('utf-8'),
+            answer_text=answer,
+            start_position_character=answer_start,
+            title=tensor_dict['title'].numpy().decode('utf-8'),
+            answers=answers
+        )
+
+    def get_examples_from_dataset(self, dataset, evaluate=False):
+        """
+        Creates a list of :class:`~transformers.data.processors.squad.SquadExample` using a TFDS dataset.
+
+        Args:
+            dataset: The tfds dataset loaded from `tensorflow_datasets.load("squad")`
+            evaluate: boolean specifying if in evaluation mode or in training mode
+
+        Returns:
+            List of SquadExample
+
+        Examples::
+
+            import tensorflow_datasets as tfds
+            dataset = tfds.load("squad")
+
+            training_examples = get_examples_from_dataset(dataset, evaluate=False)
+            evaluation_examples = get_examples_from_dataset(dataset, evaluate=True)
+        """
+
+        if evaluate:
+            dataset = dataset["validation"]
+        else:
+            dataset = dataset["train"]
+
+        examples = []
+        for tensor_dict in tqdm(dataset):
+            examples.append(self._get_example_from_tensor_dict(tensor_dict, evaluate=evaluate)) 
+
+        return examples
+
+    def get_train_examples(self, data_dir, filename=None):
+        """
+        Returns the training examples from the data directory.
+
+        Args:
+            data_dir: Directory containing the data files used for training and evaluating.
+            filename: None by default, specify this if the training file has a different name than the original one
+                which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
+
+        """
+        if data_dir is None:
+            data_dir = ""
+
+        if self.train_file is None:
+            raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
+
+        with open(os.path.join(data_dir, self.train_file if filename is None else filename), "r", encoding='utf-8') as reader:
+            input_data = json.load(reader)["data"]
+        return self._create_examples(input_data, "train")
+
+    def get_dev_examples(self, data_dir, filename=None):
+        """
+        Returns the evaluation example from the data directory.
+
+        Args:
+            data_dir: Directory containing the data files used for training and evaluating.
+            filename: None by default, specify this if the evaluation file has a different name than the original one
+                which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
+        """
+        if data_dir is None:
+            data_dir = ""
+
+        if self.dev_file is None:
+            raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
+        
+        with open(os.path.join(data_dir, self.dev_file if filename is None else filename), "r", encoding='utf-8') as reader:
+            input_data = json.load(reader)["data"]
+        return self._create_examples(input_data, "dev")
+
+    def _create_examples(self, input_data, set_type):
+        is_training = set_type == "train"
+        examples = []
+        for entry in tqdm(input_data):
+            title = entry['title']
+            for paragraph in entry["paragraphs"]:
+                context_text = paragraph["context"]
+                for qa in paragraph["qas"]:
+                    qas_id = qa["id"]
+                    question_text = qa["question"]
+                    start_position_character = None
+                    answer_text = None
+                    answers = []
+                    
+                    if "is_impossible" in qa:
+                        is_impossible = qa["is_impossible"]
+                    else:
+                        is_impossible = False
+
+                    if not is_impossible:
+                        if is_training:
+                            answer = qa["answers"][0]
+                            answer_text = answer['text']
+                            start_position_character = answer['answer_start']
+                        else:
+                            answers = qa["answers"]
+
+                    example = SquadExample(
+                        qas_id=qas_id,
+                        question_text=question_text,
+                        context_text=context_text,
+                        answer_text=answer_text,
+                        start_position_character=start_position_character,
+                        title=title,
+                        is_impossible=is_impossible,
+                        answers=answers
+                    )
+
+                    examples.append(example)
+        return examples
+
+class SquadV1Processor(SquadProcessor):
+    train_file = "train-v1.1.json"
+    dev_file = "dev-v1.1.json"
+
+
+class SquadV2Processor(SquadProcessor):
+    train_file = "train-v2.0.json"
+    dev_file = "dev-v2.0.json"
+    
+
+class SquadExample(object):
+    """
+    A single training/test example for the Squad dataset, as loaded from disk.
+
+    Args:
+        qas_id: The example's unique identifier
+        question_text: The question string
+        context_text: The context string
+        answer_text: The answer string
+        start_position_character: The character position of the start of the answer
+        title: The title of the example
+        answers: None by default, this is used during evaluation. Holds answers as well as their start positions.
+        is_impossible: False by default, set to True if the example has no possible answer.
+    """
+
+    def __init__(self,
+                 qas_id,
+                 question_text,
+                 context_text,
+                 answer_text,
+                 start_position_character,
+                 title,
+                 answers=[],
+                 is_impossible=False):
+        self.qas_id = qas_id
+        self.question_text = question_text
+        self.context_text = context_text
+        self.answer_text = answer_text
+        self.title = title
+        self.is_impossible = is_impossible 
+        self.answers = answers
+
+        self.start_position, self.end_position = 0, 0
+
+        doc_tokens = []
+        char_to_word_offset = []
+        prev_is_whitespace = True
+
+        # Split on whitespace so that different tokens may be attributed to their original position.
+        for c in self.context_text:
+            if _is_whitespace(c):
+                prev_is_whitespace = True
+            else:
+                if prev_is_whitespace:
+                    doc_tokens.append(c)
+                else:
+                    doc_tokens[-1] += c
+                prev_is_whitespace = False
+            char_to_word_offset.append(len(doc_tokens) - 1)
+
+        self.doc_tokens = doc_tokens
+        self.char_to_word_offset = char_to_word_offset
+
+        # Start end end positions only has a value during evaluation.
+        if start_position_character is not None and not is_impossible:
+            self.start_position = char_to_word_offset[start_position_character]
+            self.end_position = char_to_word_offset[start_position_character + len(answer_text) - 1]
+
+
+class SquadFeatures(object):
+    """
+    Single squad example features to be fed to a model.
+    Those features are model-specific and can be crafted from :class:`~transformers.data.processors.squad.SquadExample`
+    using the :method:`~transformers.data.processors.squad.squad_convert_examples_to_features` method.
+
+    Args:
+        input_ids: Indices of input sequence tokens in the vocabulary.
+        attention_mask: Mask to avoid performing attention on padding token indices.
+        token_type_ids: Segment token indices to indicate first and second portions of the inputs.
+        cls_index: the index of the CLS token.
+        p_mask: Mask identifying tokens that can be answers vs. tokens that cannot.
+            Mask with 1 for tokens than cannot be in the answer and 0 for token that can be in an answer
+        example_index: the index of the example
+        unique_id: The unique Feature identifier
+        paragraph_len: The length of the context
+        token_is_max_context: List of booleans identifying which tokens have their maximum context in this feature object.
+            If a token does not have their maximum context in this feature object, it means that another feature object
+            has more information related to that token and should be prioritized over this feature for that token.
+        tokens: list of tokens corresponding to the input ids
+        token_to_orig_map: mapping between the tokens and the original text, needed in order to identify the answer.
+        start_position: start of the answer token index 
+        end_position: end of the answer token index 
+    """
+
+    def __init__(self,
+                 input_ids,
+                 attention_mask,
+                 token_type_ids,
+                 cls_index,
+                 p_mask,
+                 
+                 example_index,
+                 unique_id,
+                 paragraph_len,
+                 token_is_max_context,
+                 tokens,
+                 token_to_orig_map,
+
+                 start_position,
+                 end_position
+        ):
+        self.input_ids = input_ids 
+        self.attention_mask = attention_mask
+        self.token_type_ids = token_type_ids
+        self.cls_index = cls_index
+        self.p_mask = p_mask
+
+        self.example_index = example_index
+        self.unique_id = unique_id
+        self.paragraph_len = paragraph_len
+        self.token_is_max_context = token_is_max_context
+        self.tokens = tokens
+        self.token_to_orig_map = token_to_orig_map
+
+        self.start_position = start_position
+        self.end_position = end_position
+
+
+class SquadResult(object):
+    """
+    Constructs a SquadResult which can be used to evaluate a model's output on the SQuAD dataset.
+
+    Args:
+        unique_id: The unique identifier corresponding to that example.
+        start_logits: The logits corresponding to the start of the answer
+        end_logits: The logits corresponding to the end of the answer
+    """
+    def __init__(self, unique_id, start_logits, end_logits, start_top_index=None, end_top_index=None, cls_logits=None):
+        self.start_logits = start_logits
+        self.end_logits = end_logits
+        self.unique_id = unique_id
+        
+        if start_top_index:
+            self.start_top_index = start_top_index
+            self.end_top_index = end_top_index
+            self.cls_logits = cls_logits
\ No newline at end of file

transformers/data/processors/utils.py

@@ -107,6 +107,13 @@ class DataProcessor(object):
         """Gets the list of labels for this data set."""
         raise NotImplementedError()
 
+    def tfds_map(self, example):
+        """Some tensorflow_datasets datasets are not formatted the same way the GLUE datasets are. 
+        This method converts examples to the correct format."""
+        if len(self.get_labels()) > 1:
+            example.label = self.get_labels()[int(example.label)]
+        return example
+
     @classmethod
     def _read_tsv(cls, input_file, quotechar=None):
         """Reads a tab separated value file."""

transformers/data/processors/xnli.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" XNLI utils (dataset loading and evaluation) """
+
+from __future__ import absolute_import, division, print_function
+
+import logging
+import os
+
+from .utils import DataProcessor, InputExample
+
+logger = logging.getLogger(__name__)
+
+class XnliProcessor(DataProcessor):
+    """Processor for the XNLI dataset.
+    Adapted from https://github.com/google-research/bert/blob/f39e881b169b9d53bea03d2d341b31707a6c052b/run_classifier.py#L207"""
+
+    def __init__(self, language, train_language = None):
+        self.language = language
+        self.train_language = train_language
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        lg = self.language if self.train_language is None else self.train_language
+        lines = self._read_tsv(os.path.join(data_dir, "XNLI-MT-1.0/multinli/multinli.train.{}.tsv".format(lg)))
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % ('train', i)
+            text_a = line[0]
+            text_b = line[1]
+            label = "contradiction" if line[2] == "contradictory" else line[2]
+            assert isinstance(text_a, str) and isinstance(text_b, str) and isinstance(label, str)
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        lines = self._read_tsv(os.path.join(data_dir, "XNLI-1.0/xnli.test.tsv"))
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            language = line[0]
+            if language != self.language:
+                continue
+            guid = "%s-%s" % ('test', i)
+            text_a = line[6]
+            text_b = line[7]
+            label = line[1]
+            assert isinstance(text_a, str) and isinstance(text_b, str) and isinstance(label, str)
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+    def get_labels(self):
+        """See base class."""
+        return ["contradiction", "entailment", "neutral"]
+
+xnli_processors = {
+    "xnli": XnliProcessor,
+}
+
+xnli_output_modes = {
+    "xnli": "classification",
+}
+
+xnli_tasks_num_labels = {
+    "xnli": 3,
+}

transformers/file_utils.py

@@ -21,7 +21,8 @@ import boto3
 from botocore.config import Config
 from botocore.exceptions import ClientError
 import requests
-from tqdm import tqdm
+from tqdm.auto import tqdm
+from contextlib import contextmanager
 
 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
@@ -72,6 +73,8 @@ TF2_WEIGHTS_NAME = 'tf_model.h5'
 TF_WEIGHTS_NAME = 'model.ckpt'
 CONFIG_NAME = "config.json"
 
+S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
+
 def is_torch_available():
     return _torch_available
 
@@ -102,6 +105,18 @@ else:
             return fn
         return docstring_decorator
 
+
+def is_remote_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ('http', 'https', 's3')
+
+def hf_bucket_url(identifier, postfix=None):
+    if postfix is None:
+        return "/".join((S3_BUCKET_PREFIX, identifier))
+    else:
+        return "/".join((S3_BUCKET_PREFIX, identifier, postfix))
+
+
 def url_to_filename(url, etag=None):
     """
     Convert `url` into a hashed filename in a repeatable way.
@@ -152,7 +167,7 @@ def filename_to_url(filename, cache_dir=None):
     return url, etag
 
 
-def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None):
+def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None, resume_download=False):
     """
     Given something that might be a URL (or might be a local path),
     determine which. If it's a URL, download the file and cache it, and
@@ -161,6 +176,7 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
     Args:
         cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
         force_download: if True, re-dowload the file even if it's already cached in the cache dir.
+        resume_download: if True, resume the download if incompletly recieved file is found.
     """
     if cache_dir is None:
         cache_dir = TRANSFORMERS_CACHE
@@ -169,15 +185,15 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
     if sys.version_info[0] == 3 and isinstance(cache_dir, Path):
         cache_dir = str(cache_dir)
 
-    parsed = urlparse(url_or_filename)
-
-    if parsed.scheme in ('http', 'https', 's3'):
+    if is_remote_url(url_or_filename):
         # URL, so get it from the cache (downloading if necessary)
-        return get_from_cache(url_or_filename, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+        return get_from_cache(url_or_filename, cache_dir=cache_dir,
+            force_download=force_download, proxies=proxies,
+            resume_download=resume_download)
     elif os.path.exists(url_or_filename):
         # File, and it exists.
         return url_or_filename
-    elif parsed.scheme == '':
+    elif urlparse(url_or_filename).scheme == '':
         # File, but it doesn't exist.
         raise EnvironmentError("file {} not found".format(url_or_filename))
     else:
@@ -234,19 +250,22 @@ def s3_get(url, temp_file, proxies=None):
     s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file)
 
 
-def http_get(url, temp_file, proxies=None):
-    req = requests.get(url, stream=True, proxies=proxies)
-    content_length = req.headers.get('Content-Length')
-    total = int(content_length) if content_length is not None else None
-    progress = tqdm(unit="B", total=total)
-    for chunk in req.iter_content(chunk_size=1024):
+def http_get(url, temp_file, proxies=None, resume_size=0):
+    headers={'Range':'bytes=%d-'%(resume_size,)} if resume_size > 0 else None
+    response = requests.get(url, stream=True, proxies=proxies, headers=headers)
+    if response.status_code == 416:  # Range not satisfiable
+        return
+    content_length = response.headers.get('Content-Length')
+    total = resume_size + int(content_length) if content_length is not None else None
+    progress = tqdm(unit="B", unit_scale=True, total=total, initial=resume_size, desc="Downloading")
+    for chunk in response.iter_content(chunk_size=1024):
         if chunk: # filter out keep-alive new chunks
             progress.update(len(chunk))
             temp_file.write(chunk)
     progress.close()
 
 
-def get_from_cache(url, cache_dir=None, force_download=False, proxies=None):
+def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag_timeout=10, resume_download=False):
     """
     Given a URL, look for the corresponding dataset in the local cache.
     If it's not there, download it. Then return the path to the cached file.
@@ -266,12 +285,12 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None):
         etag = s3_etag(url, proxies=proxies)
     else:
         try:
-            response = requests.head(url, allow_redirects=True, proxies=proxies)
+            response = requests.head(url, allow_redirects=True, proxies=proxies, timeout=etag_timeout)
             if response.status_code != 200:
                 etag = None
             else:
                 etag = response.headers.get("ETag")
-        except EnvironmentError:
+        except (EnvironmentError, requests.exceptions.Timeout):
             etag = None
 
     if sys.version_info[0] == 2 and etag is not None:
@@ -289,17 +308,35 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None):
         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
+
     if 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 tempfile.NamedTemporaryFile() as temp_file:
+        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)
             else:
-                http_get(url, temp_file, proxies=proxies)
+                http_get(url, temp_file, proxies=proxies, resume_size=resume_size)
 
             # we are copying the file before closing it, so flush to avoid truncation
             temp_file.flush()

transformers/hf_api.py

+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import, division, print_function
+
+import os
+from os.path import expanduser
+
+import requests
+import six
+from requests.exceptions import HTTPError
+from tqdm import tqdm
+
+ENDPOINT = "https://huggingface.co"
+
+class S3Obj:
+    def __init__(
+        self,
+        filename,     # type: str
+        LastModified, # type: str
+        ETag,         # type: str
+        Size,         # type: int
+        **kwargs
+    ):
+        self.filename = filename
+        self.LastModified = LastModified
+        self.ETag = ETag
+        self.Size = Size
+
+
+class PresignedUrl:
+    def __init__(
+        self,
+        write,  # type: str
+        access, # type: str
+        type,   # type: str
+        **kwargs
+    ):
+        self.write = write
+        self.access = access
+        self.type = type # mime-type to send to S3.
+
+
+class HfApi:
+    def __init__(self, endpoint=None):
+        self.endpoint = endpoint if endpoint is not None else ENDPOINT
+
+    def login(
+        self,
+        username, # type: str
+        password, # type: str
+    ):
+        # type: (...) -> str
+        """
+        Call HF API to sign in a user and get a token if credentials are valid.
+
+        Outputs:
+            token if credentials are valid
+
+        Throws:
+            requests.exceptions.HTTPError if credentials are invalid
+        """
+        path = "{}/api/login".format(self.endpoint)
+        r = requests.post(path, json={"username": username, "password": password})
+        r.raise_for_status()
+        d = r.json()
+        return d["token"]
+
+    def whoami(
+        self,
+        token, # type: str
+    ):
+        # type: (...) -> str
+        """
+        Call HF API to know "whoami"
+        """
+        path = "{}/api/whoami".format(self.endpoint)
+        r = requests.get(path, headers={"authorization": "Bearer {}".format(token)})
+        r.raise_for_status()
+        d = r.json()
+        return d["user"]
+
+    def logout(self, token):
+        # type: (...) -> void
+        """
+        Call HF API to log out.
+        """
+        path = "{}/api/logout".format(self.endpoint)
+        r = requests.post(path, headers={"authorization": "Bearer {}".format(token)})
+        r.raise_for_status()
+
+    def presign(self, token, filename):
+        # type: (...) -> PresignedUrl
+        """
+        Call HF API to get a presigned url to upload `filename` to S3.
+        """
+        path = "{}/api/presign".format(self.endpoint)
+        r = requests.post(
+            path,
+            headers={"authorization": "Bearer {}".format(token)},
+            json={"filename": filename},
+        )
+        r.raise_for_status()
+        d = r.json()
+        return PresignedUrl(**d)
+
+    def presign_and_upload(self, token, filename, filepath):
+        # type: (...) -> str
+        """
+        Get a presigned url, then upload file to S3.
+
+        Outputs:
+            url: Read-only url for the stored file on S3.
+        """
+        urls = self.presign(token, filename=filename)
+        # streaming upload:
+        # https://2.python-requests.org/en/master/user/advanced/#streaming-uploads
+        # 
+        # Even though we presign with the correct content-type,
+        # the client still has to specify it when uploading the file.
+        with open(filepath, "rb") as f:
+            pf = TqdmProgressFileReader(f)
+
+            r = requests.put(urls.write, data=f, headers={
+                "content-type": urls.type,
+            })
+            r.raise_for_status()
+            pf.close()
+        return urls.access
+
+    def list_objs(self, token):
+        # type: (...) -> List[S3Obj]
+        """
+        Call HF API to list all stored files for user.
+        """
+        path = "{}/api/listObjs".format(self.endpoint)
+        r = requests.get(path, headers={"authorization": "Bearer {}".format(token)})
+        r.raise_for_status()
+        d = r.json()
+        return [S3Obj(**x) for x in d]
+
+
+
+class TqdmProgressFileReader:
+    """
+    Wrap an io.BufferedReader `f` (such as the output of `open(…, "rb")`)
+    and override `f.read()` so as to display a tqdm progress bar.
+
+    see github.com/huggingface/transformers/pull/2078#discussion_r354739608
+    for implementation details.
+    """
+    def __init__(
+        self,
+        f   # type: io.BufferedReader
+    ):
+        self.f = f
+        self.total_size = os.fstat(f.fileno()).st_size # type: int
+        self.pbar = tqdm(total=self.total_size, leave=False)
+        if six.PY3:
+            # does not work unless PY3
+            # no big deal as the CLI does not currently support PY2 anyways.
+            self.read = f.read
+            f.read = self._read
+
+    def _read(self, n=-1):
+        self.pbar.update(n)
+        return self.read(n)
+
+    def close(self):
+        self.pbar.close()
+
+
+
+class HfFolder:
+    path_token = expanduser("~/.huggingface/token")
+
+    @classmethod
+    def save_token(cls, token):
+        """
+        Save token, creating folder as needed.
+        """
+        if six.PY3:
+            os.makedirs(os.path.dirname(cls.path_token), exist_ok=True)
+        else:
+            # Python 2
+            try:
+                os.makedirs(os.path.dirname(cls.path_token))
+            except OSError as e:
+                if e.errno != os.errno.EEXIST:
+                    raise e
+                pass
+        with open(cls.path_token, 'w+') as f:
+            f.write(token)
+
+    @classmethod
+    def get_token(cls):
+        """
+        Get token or None if not existent.
+        """
+        try:
+            with open(cls.path_token, 'r') as f:
+                return f.read()
+        except:
+            # this is too wide. When Py2 is dead use:
+            # `except FileNotFoundError:` instead
+            return None
+
+    @classmethod
+    def delete_token(cls):
+        """
+        Delete token.
+        Do not fail if token does not exist.
+        """
+        try:
+            os.remove(cls.path_token)
+        except:
+            return

transformers/modeling_albert.py

+
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch ALBERT model. """
+
+import os
+import math
+import logging
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss, MSELoss
+from transformers.modeling_utils import PreTrainedModel
+from transformers.configuration_albert import AlbertConfig
+from transformers.modeling_bert import BertEmbeddings, BertSelfAttention, prune_linear_layer, ACT2FN
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+
+ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-pytorch_model.bin",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-pytorch_model.bin",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-pytorch_model.bin",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-pytorch_model.bin",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-pytorch_model.bin",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-pytorch_model.bin",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-pytorch_model.bin",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-pytorch_model.bin",
+}
+
+
+def load_tf_weights_in_albert(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        print(name)
+    
+    for name, array in zip(names, arrays):
+        original_name = name
+
+        # If saved from the TF HUB module
+        name = name.replace("module/", "")
+
+        # Renaming and simplifying
+        name = name.replace("ffn_1", "ffn")
+        name = name.replace("bert/", "albert/")
+        name = name.replace("attention_1", "attention")   
+        name = name.replace("transform/", "")
+        name = name.replace("LayerNorm_1", "full_layer_layer_norm")    
+        name = name.replace("LayerNorm", "attention/LayerNorm")   
+        name = name.replace("transformer/", "")
+
+        # The feed forward layer had an 'intermediate' step which has been abstracted away
+        name = name.replace("intermediate/dense/", "")
+        name = name.replace("ffn/intermediate/output/dense/", "ffn_output/")
+
+        # ALBERT attention was split between self and output which have been abstracted away
+        name = name.replace("/output/", "/")
+        name = name.replace("/self/", "/")
+
+        # The pooler is a linear layer
+        name = name.replace("pooler/dense", "pooler")
+
+        # The classifier was simplified to predictions from cls/predictions
+        name = name.replace("cls/predictions", "predictions")
+        name = name.replace("predictions/attention", "predictions")
+
+        # Naming was changed to be more explicit
+        name = name.replace("embeddings/attention", "embeddings")    
+        name = name.replace("inner_group_", "albert_layers/") 
+        name = name.replace("group_", "albert_layer_groups/")   
+
+        # Classifier
+        if len(name.split("/")) == 1 and ("output_bias" in name or "output_weights" in name):
+            name = "classifier/" + name
+
+        # No ALBERT model currently handles the next sentence prediction task 
+        if "seq_relationship" in name:
+            continue
+
+        name = name.split('/')
+
+        # Ignore the gradients applied by the LAMB/ADAM optimizers.
+        if "adam_m" in name or "adam_v" in name or "global_step" in name:
+            logger.info("Skipping {}".format("/".join(name)))
+            continue
+
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+
+            if l[0] == 'kernel' or l[0] == 'gamma':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'output_bias' or l[0] == 'beta':
+                pointer = getattr(pointer, 'bias')
+            elif l[0] == 'output_weights':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'squad':
+                pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+
+        if m_name[-11:] == '_embeddings':
+            pointer = getattr(pointer, 'weight')
+        elif m_name == 'kernel':
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        print("Initialize PyTorch weight {} from {}".format(name, original_name))
+        pointer.data = torch.from_numpy(array)
+
+    return model
+
+
+class AlbertEmbeddings(BertEmbeddings):
+    """
+    Construct the embeddings from word, position and token_type embeddings.
+    """
+    def __init__(self, config):
+        super(AlbertEmbeddings, self).__init__(config)
+
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=0)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
+        self.LayerNorm = torch.nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
+
+
+class AlbertAttention(BertSelfAttention):
+    def __init__(self, config):
+        super(AlbertAttention, self).__init__(config)
+
+        self.output_attentions = config.output_attentions
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size 
+        self.attention_head_size = config.hidden_size // config.num_attention_heads
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.num_attention_heads, self.attention_head_size)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+
+        # Prune linear layers
+        self.query = prune_linear_layer(self.query, index)
+        self.key = prune_linear_layer(self.key, index)
+        self.value = prune_linear_layer(self.value, index)
+        self.dense = prune_linear_layer(self.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.num_attention_heads = self.num_attention_heads - len(heads)
+        self.all_head_size = self.attention_head_size * self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(self, input_ids, attention_mask=None, head_mask=None):
+        mixed_query_layer = self.query(input_ids)
+        mixed_key_layer = self.key(input_ids)
+        mixed_value_layer = self.value(input_ids)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        reshaped_context_layer = context_layer.view(*new_context_layer_shape)
+        
+
+        # Should find a better way to do this
+        w = self.dense.weight.t().view(self.num_attention_heads, self.attention_head_size, self.hidden_size).to(context_layer.dtype)
+        b = self.dense.bias.to(context_layer.dtype)
+
+        projected_context_layer = torch.einsum("bfnd,ndh->bfh", context_layer, w) + b
+        projected_context_layer_dropout = self.dropout(projected_context_layer)
+        layernormed_context_layer = self.LayerNorm(input_ids + projected_context_layer_dropout)
+        return (layernormed_context_layer, attention_probs) if self.output_attentions else (layernormed_context_layer,)
+
+
+class AlbertLayer(nn.Module):
+    def __init__(self, config):
+        super(AlbertLayer, self).__init__()
+        
+        self.config = config
+        self.full_layer_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = AlbertAttention(config)
+        self.ffn = nn.Linear(config.hidden_size, config.intermediate_size) 
+        self.ffn_output = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        attention_output = self.attention(hidden_states, attention_mask, head_mask)
+        ffn_output = self.ffn(attention_output[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output[0])
+
+        return (hidden_states,) + attention_output[1:]  # add attentions if we output them
+
+
+class AlbertLayerGroup(nn.Module):
+    def __init__(self, config):
+        super(AlbertLayerGroup, self).__init__()
+        
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            layer_output = albert_layer(hidden_states, attention_mask, head_mask[layer_index])
+            hidden_states = layer_output[0]
+
+            if self.output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if self.output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
+
+
+class AlbertTransformer(nn.Module):
+    def __init__(self, config):
+        super(AlbertTransformer, self).__init__()
+        
+        self.config = config
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.embedding_hidden_mapping_in = nn.Linear(config.embedding_size, config.hidden_size)
+        self.albert_layer_groups = nn.ModuleList([AlbertLayerGroup(config) for _ in range(config.num_hidden_groups)])
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+
+        all_attentions = ()
+
+        if self.output_hidden_states:
+            all_hidden_states = (hidden_states,)
+
+        for i in range(self.config.num_hidden_layers):
+            # Number of layers in a hidden group
+            layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+            # Index of the hidden group
+            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+            # Index of the layer inside the group
+            layer_idx = int(i - group_idx * layers_per_group)
+            
+            layer_group_output = self.albert_layer_groups[group_idx](hidden_states, attention_mask, head_mask[group_idx*layers_per_group:(group_idx+1)*layers_per_group])  
+            hidden_states = layer_group_output[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+
+class AlbertPreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = AlbertConfig
+    pretrained_model_archive_map = ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "albert"
+
+    def _init_weights(self, module):
+        """ Initialize the weights.
+        """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if isinstance(module, (nn.Linear)) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+ALBERT_START_DOCSTRING = r"""    The ALBERT model was proposed in
+    `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`_
+    by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut. It presents
+    two parameter-reduction techniques to lower memory consumption and increase the trainig speed of BERT.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`:
+        https://arxiv.org/abs/1909.11942
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.AlbertConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, BERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Albert is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            Indices can be obtained using :class:`transformers.AlbertTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare ALBERT Model transformer outputting raw hidden-states without any specific head on top.",
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertModel(AlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **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.
+    """
+
+    config_class = AlbertConfig
+    pretrained_model_archive_map = ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_albert
+    base_model_prefix = "albert"
+
+    def __init__(self, config):
+        super(AlbertModel, self).__init__(config)
+
+        self.config = config
+        self.embeddings = AlbertEmbeddings(config)
+        self.encoder = AlbertTransformer(config)
+        self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
+        self.pooler_activation = nn.Tanh()
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.embeddings.word_embeddings
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            ALBERT has a different architecture in that its layers are shared across groups, which then has inner groups.
+            If an ALBERT model has 12 hidden layers and 2 hidden groups, with two inner groups, there
+            is a total of 4 different layers.
+
+            These layers are flattened: the indices [0,1] correspond to the two inner groups of the first hidden layer,
+            while [2,3] correspond to the two inner groups of the second hidden layer.
+
+            Any layer with in index other than [0,1,2,3] will result in an error.
+            See base class PreTrainedModel for more information about head pruning
+        """
+        for layer, heads in heads_to_prune.items():
+            group_idx = int(layer / self.config.inner_group_num)
+            inner_group_idx = int(layer - group_idx * self.config.inner_group_num)
+            self.encoder.albert_layer_groups[group_idx].albert_layers[inner_group_idx].attention.prune_heads(heads)
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                inputs_embeds=None):
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
+                                           inputs_embeds=inputs_embeds)
+        encoder_outputs = self.encoder(embedding_output,
+                                       extended_attention_mask,
+                                       head_mask=head_mask)
+
+        sequence_output = encoder_outputs[0]
+
+        pooled_output = self.pooler_activation(self.pooler(sequence_output[:, 0]))
+
+        outputs = (sequence_output, pooled_output) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+        return outputs
+
+class AlbertMLMHead(nn.Module):
+    def __init__(self, config):
+        super(AlbertMLMHead, self).__init__()
+
+        self.LayerNorm = nn.LayerNorm(config.embedding_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.dense = nn.Linear(config.hidden_size, config.embedding_size)
+        self.decoder = nn.Linear(config.embedding_size, config.vocab_size)
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+
+        prediction_scores = hidden_states + self.bias
+
+        return prediction_scores
+
+
+@add_start_docstrings("Bert Model with a `language modeling` head on top.", ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertForMaskedLM(AlbertPreTrainedModel):
+    r"""
+        **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
+            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) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``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).
+        **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.
+    """
+
+    def __init__(self, config):
+        super(AlbertForMaskedLM, self).__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.predictions = AlbertMLMHead(config)
+
+        self.init_weights()
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.predictions.decoder,
+                                   self.albert.embeddings.word_embeddings)
+
+    def get_output_embeddings(self):
+        return self.predictions.decoder
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
+                masked_lm_labels=None):
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds
+        )
+        sequence_outputs = outputs[0]
+
+        prediction_scores = self.predictions(sequence_outputs)
+
+        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)
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
+            outputs = (masked_lm_loss,) + outputs
+
+        return outputs
+
+
+@add_start_docstrings("""Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertForSequenceClassification(AlbertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) 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 = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = AlbertForSequenceClassification.from_pretrained('albert-base-v2')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(AlbertForSequenceClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
+
+@add_start_docstrings("""Albert 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`). """,
+    ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertForQuestionAnswering(AlbertPreTrainedModel):
+    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 = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = AlbertForQuestionAnswering.from_pretrained('albert-base-v2')
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
+        input_ids = tokenizer.encode(input_text)
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
+        # a nice puppet
+
+
+    """
+    def __init__(self, config):
+        super(AlbertForQuestionAnswering, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                inputs_embeds=None, start_positions=None, end_positions=None):
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds
+        )
+
+        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_auto.py

@@ -27,6 +27,8 @@ from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassi
 from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering
 from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification
 from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, DistilBertForSequenceClassification
+from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice
+from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, AlbertForQuestionAnswering
 
 from .modeling_utils import PreTrainedModel, SequenceSummary
 
@@ -48,14 +50,16 @@ class AutoModel(object):
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertModel (DistilBERT model)
+            - contains `albert`: AlbertModel (ALBERT model)
+            - contains `camembert`: CamembertModel (CamemBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
             - contains `gpt2`: GPT2Model (OpenAI GPT-2 model)
-            - contains `ctrl`: CTRLModel (Salesforce CTRL  model)
             - contains `transfo-xl`: TransfoXLModel (Transformer-XL model)
             - contains `xlnet`: XLNetModel (XLNet model)
             - contains `xlm`: XLMModel (XLM model)
+            - contains `ctrl`: CTRLModel (Salesforce CTRL  model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -71,14 +75,16 @@ class AutoModel(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertModel (DistilBERT model)
+            - contains `albert`: AlbertModel (ALBERT model)
+            - contains `camembert`: CamembertModel (CamemBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
             - contains `gpt2`: GPT2Model (OpenAI GPT-2 model)
-            - contains `ctrl`: CTRLModel (Salesforce CTRL  model)
             - contains `transfo-xl`: TransfoXLModel (Transformer-XL model)
             - contains `xlnet`: XLNetModel (XLNet model)
             - contains `xlm`: XLMModel (XLM model)
+            - contains `ctrl`: CTRLModel (Salesforce CTRL model)
 
             The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
             To train the model, you should first set it back in training mode with `model.train()`
@@ -87,6 +93,7 @@ class AutoModel(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -112,6 +119,9 @@ class AutoModel(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
             proxies: (`optional`) dict, default None:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
@@ -138,6 +148,10 @@ class AutoModel(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return AlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -156,7 +170,7 @@ class AutoModel(object):
             return CTRLModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta, 'ctrl'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta, 'ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
 
 
 class AutoModelWithLMHead(object):
@@ -172,14 +186,16 @@ class AutoModelWithLMHead(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
+            - contains `albert`: AlbertForMaskedLM (ALBERT model)
+            - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
             - contains `gpt2`: GPT2LMHeadModel (OpenAI GPT-2 model)
-            - contains `ctrl`: CTRLLMModel (Salesforce CTRL model)
             - contains `transfo-xl`: TransfoXLLMHeadModel (Transformer-XL model)
             - contains `xlnet`: XLNetLMHeadModel (XLNet model)
             - contains `xlm`: XLMWithLMHeadModel (XLM model)
+            - contains `ctrl`: CTRLLMHeadModel (Salesforce CTRL model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -198,6 +214,8 @@ class AutoModelWithLMHead(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
+            - contains `albert`: AlbertForMaskedLM (ALBERT model)
+            - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -205,6 +223,7 @@ class AutoModelWithLMHead(object):
             - contains `transfo-xl`: TransfoXLLMHeadModel (Transformer-XL model)
             - contains `xlnet`: XLNetLMHeadModel (XLNet model)
             - contains `xlm`: XLMWithLMHeadModel (XLM model)
+            - contains `ctrl`: CTRLLMHeadModel (Salesforce CTRL model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -213,6 +232,7 @@ class AutoModelWithLMHead(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -237,6 +257,8 @@ class AutoModelWithLMHead(object):
 
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
 
             proxies: (`optional`) dict, default None:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
@@ -264,6 +286,10 @@ class AutoModelWithLMHead(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return AlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -282,7 +308,7 @@ class AutoModelWithLMHead(object):
             return CTRLLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta','ctrl'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta','ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
 
 
 class AutoModelForSequenceClassification(object):
@@ -298,6 +324,8 @@ class AutoModelForSequenceClassification(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
+            - contains `albert`: AlbertForSequenceClassification (ALBERT model)
+            - contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -320,6 +348,8 @@ class AutoModelForSequenceClassification(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
+            - contains `albert`: AlbertForSequenceClassification (ALBERT model)
+            - contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -332,6 +362,7 @@ class AutoModelForSequenceClassification(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -357,6 +388,9 @@ class AutoModelForSequenceClassification(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
             proxies: (`optional`) dict, default None:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
@@ -383,6 +417,10 @@ class AutoModelForSequenceClassification(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return AlbertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -393,7 +431,7 @@ class AutoModelForSequenceClassification(object):
             return XLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
+                         "'bert', 'xlnet', 'xlm', 'roberta', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
 
 
 class AutoModelForQuestionAnswering(object):
@@ -409,6 +447,7 @@ class AutoModelForQuestionAnswering(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForQuestionAnswering (DistilBERT model)
+            - contains `albert`: AlbertForQuestionAnswering (ALBERT model)
             - contains `bert`: BertForQuestionAnswering (Bert model)
             - contains `xlnet`: XLNetForQuestionAnswering (XLNet model)
             - contains `xlm`: XLMForQuestionAnswering (XLM model)
@@ -430,6 +469,7 @@ class AutoModelForQuestionAnswering(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForQuestionAnswering (DistilBERT model)
+            - contains `albert`: AlbertForQuestionAnswering (ALBERT model)
             - contains `bert`: BertForQuestionAnswering (Bert model)
             - contains `xlnet`: XLNetForQuestionAnswering (XLNet model)
             - contains `xlm`: XLMForQuestionAnswering (XLM model)
@@ -441,6 +481,7 @@ class AutoModelForQuestionAnswering(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -492,6 +533,8 @@ class AutoModelForQuestionAnswering(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return AlbertForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'xlnet' in pretrained_model_name_or_path:
@@ -500,4 +543,4 @@ class AutoModelForQuestionAnswering(object):
             return XLMForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
+                         "'bert', 'xlnet', 'xlm', 'distilbert', 'albert'".format(pretrained_model_name_or_path))

transformers/modeling_bert.py

@@ -17,12 +17,10 @@
 
 from __future__ import absolute_import, division, print_function, unicode_literals
 
-import json
 import logging
 import math
 import os
 import sys
-from io import open
 
 import torch
 from torch import nn
@@ -50,8 +48,13 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
     'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-pytorch_model.bin",
     'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-pytorch_model.bin",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-pytorch_model.bin",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-pytorch_model.bin",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-pytorch_model.bin",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-pytorch_model.bin"
 }
 
+
 def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
     """ Load tf checkpoints in a pytorch model.
     """
@@ -127,21 +130,28 @@ def gelu(x):
     """
     return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
 
+
 def gelu_new(x):
     """ Implementation of the gelu activation function currently in Google Bert repo (identical to OpenAI GPT).
         Also see https://arxiv.org/abs/1606.08415
     """
     return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
 
+
 def swish(x):
     return x * torch.sigmoid(x)
 
 
-ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish, "gelu_new": gelu_new}
+def mish(x):
+    return x * torch.tanh(nn.functional.softplus(x))
+
+
+ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish, "gelu_new": gelu_new, "mish": mish}
 
 
 BertLayerNorm = torch.nn.LayerNorm
 
+
 class BertEmbeddings(nn.Module):
     """Construct the embeddings from word, position and token_type embeddings.
     """
@@ -156,19 +166,26 @@ class BertEmbeddings(nn.Module):
         self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
-    def forward(self, input_ids, token_type_ids=None, position_ids=None):
-        seq_length = input_ids.size(1)
+    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_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).expand(input_shape)
         if token_type_ids is None:
-            token_type_ids = torch.zeros_like(input_ids)
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
-        words_embeddings = self.word_embeddings(input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
         position_embeddings = self.position_embeddings(position_ids)
         token_type_embeddings = self.token_type_embeddings(token_type_ids)
 
-        embeddings = words_embeddings + position_embeddings + token_type_embeddings
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
         embeddings = self.LayerNorm(embeddings)
         embeddings = self.dropout(embeddings)
         return embeddings
@@ -198,10 +215,19 @@ class BertSelfAttention(nn.Module):
         x = x.view(*new_x_shape)
         return x.permute(0, 2, 1, 3)
 
-    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
         mixed_query_layer = self.query(hidden_states)
-        mixed_key_layer = self.key(hidden_states)
-        mixed_value_layer = self.value(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        if encoder_hidden_states is not None:
+            mixed_key_layer = self.key(encoder_hidden_states)
+            mixed_value_layer = self.value(encoder_hidden_states)
+            attention_mask = encoder_attention_mask
+        else:
+            mixed_key_layer = self.key(hidden_states)
+            mixed_value_layer = self.value(hidden_states)
 
         query_layer = self.transpose_for_scores(mixed_query_layer)
         key_layer = self.transpose_for_scores(mixed_key_layer)
@@ -260,7 +286,7 @@ class BertAttention(nn.Module):
         if len(heads) == 0:
             return
         mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
-        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        heads = set(heads) - self.pruned_heads  # Convert to set and remove already pruned heads
         for head in heads:
             # Compute how many pruned heads are before the head and move the index accordingly
             head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
@@ -279,9 +305,9 @@ class BertAttention(nn.Module):
         self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
         self.pruned_heads = self.pruned_heads.union(heads)
 
-    def forward(self, input_tensor, attention_mask=None, head_mask=None):
-        self_outputs = self.self(input_tensor, attention_mask, head_mask)
-        attention_output = self.output(self_outputs[0], input_tensor)
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        self_outputs = self.self(hidden_states, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask)
+        attention_output = self.output(self_outputs[0], hidden_states)
         outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
         return outputs
 
@@ -319,15 +345,25 @@ class BertLayer(nn.Module):
     def __init__(self, config):
         super(BertLayer, self).__init__()
         self.attention = BertAttention(config)
+        self.is_decoder = config.is_decoder
+        if self.is_decoder:
+            self.crossattention = BertAttention(config)
         self.intermediate = BertIntermediate(config)
         self.output = BertOutput(config)
 
-    def forward(self, hidden_states, attention_mask=None, head_mask=None):
-        attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
-        attention_output = attention_outputs[0]
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        self_attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        if self.is_decoder and encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(attention_output, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask)
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+
         intermediate_output = self.intermediate(attention_output)
         layer_output = self.output(intermediate_output, attention_output)
-        outputs = (layer_output,) + attention_outputs[1:]  # add attentions if we output them
+        outputs = (layer_output,) + outputs
         return outputs
 
 
@@ -338,14 +374,14 @@ class BertEncoder(nn.Module):
         self.output_hidden_states = config.output_hidden_states
         self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
 
-    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
         all_hidden_states = ()
         all_attentions = ()
         for i, layer_module in enumerate(self.layer):
             if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
-            layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i])
+            layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i], encoder_hidden_states, encoder_attention_mask)
             hidden_states = layer_outputs[0]
 
             if self.output_attentions:
@@ -484,7 +520,7 @@ BERT_START_DOCSTRING = r"""    The BERT model was proposed in
         https://pytorch.org/docs/stable/nn.html#module
 
     Parameters:
-        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
             Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
@@ -498,13 +534,13 @@ BERT_INPUTS_DOCSTRING = r"""
             (a) For sequence pairs:
 
                 ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-                
+
                 ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
 
             (b) For single sequences:
 
                 ``tokens:         [CLS] the dog is hairy . [SEP]``
-                
+
                 ``token_type_ids:   0   0   0   0  0     0   0``
 
             Bert is a model with absolute position embeddings so it's usually advised to pad the inputs on
@@ -529,6 +565,18 @@ BERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        **encoder_hidden_states**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``:
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model
+            is configured as a decoder.
+        **encoder_attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask
+            is used in the cross-attention if the model is configured as a decoder.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 """
 
 @add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
@@ -557,13 +605,14 @@ class BertModel(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertModel.from_pretrained('bert-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
 
     """
     def __init__(self, config):
         super(BertModel, self).__init__(config)
+        self.config = config
 
         self.embeddings = BertEmbeddings(config)
         self.encoder = BertEncoder(config)
@@ -571,12 +620,12 @@ class BertModel(BertPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        old_embeddings = self.embeddings.word_embeddings
-        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
-        self.embeddings.word_embeddings = new_embeddings
+    def get_input_embeddings(self):
         return self.embeddings.word_embeddings
 
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -585,27 +634,87 @@ class BertModel(BertPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.encoder.layer[layer].attention.prune_heads(heads)
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None,
+                head_mask=None, inputs_embeds=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        """ Forward pass on the Model.
+
+        The model can behave as an encoder (with only self-attention) as well
+        as a decoder, in which case a layer of cross-attention is added between
+        the self-attention layers, following the architecture described in `Attention is all you need`_ by Ashish Vaswani,
+        Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+        To behave as an decoder the model needs to be initialized with the
+        `is_decoder` argument of the configuration set to `True`; an
+        `encoder_hidden_states` is expected as an input to the forward pass.
+
+        .. _`Attention is all you need`:
+            https://arxiv.org/abs/1706.03762
+
+        """
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
         if attention_mask is None:
-            attention_mask = torch.ones_like(input_ids)
+            attention_mask = torch.ones(input_shape, device=device)
         if token_type_ids is None:
-            token_type_ids = torch.zeros_like(input_ids)
-
-        # We create a 3D attention mask from a 2D tensor mask.
-        # Sizes are [batch_size, 1, 1, to_seq_length]
-        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
-        # this attention mask is more simple than the triangular masking of causal attention
-        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
-        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                batch_size, seq_length = input_shape
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = causal_mask.to(torch.long)  # not converting to long will cause errors with pytorch version < 1.3
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError("Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(input_shape, attention_mask.shape))
 
         # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
         # masked positions, this operation will create a tensor which is 0.0 for
         # positions we want to attend and -10000.0 for masked positions.
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.
-        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
         extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
 
+        # If a 2D ou 3D attention mask is provided for the cross-attention
+        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+
+            if encoder_attention_mask.dim() == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            elif encoder_attention_mask.dim() == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+            else:
+                raise ValueError("Wrong shape for encoder_hidden_shape (shape {}) or encoder_attention_mask (shape {})".format(encoder_hidden_shape,
+                                                                                                                               encoder_attention_mask.shape))
+
+            encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
+        else:
+            encoder_extended_attention_mask = None
+
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
         # attention_probs has shape bsz x n_heads x N x N
@@ -617,14 +726,16 @@ class BertModel(BertPreTrainedModel):
                 head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
             elif head_mask.dim() == 2:
                 head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
-            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype)  # switch to fload if need + fp16 compatibility
         else:
             head_mask = [None] * self.config.num_hidden_layers
 
-        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+        embedding_output = self.embeddings(input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds)
         encoder_outputs = self.encoder(embedding_output,
-                                       extended_attention_mask,
-                                       head_mask=head_mask)
+                                       attention_mask=extended_attention_mask,
+                                       head_mask=head_mask,
+                                       encoder_hidden_states=encoder_hidden_states,
+                                       encoder_attention_mask=encoder_extended_attention_mask)
         sequence_output = encoder_outputs[0]
         pooled_output = self.pooler(sequence_output)
 
@@ -633,8 +744,9 @@ class BertModel(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with two heads on top as done during the pre-training:
-    a `masked language modeling` head and a `next sentence prediction (classification)` head. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                       a `masked language modeling` head and a `next sentence prediction (classification)` head. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForPreTraining(BertPreTrainedModel):
     r"""
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
@@ -667,7 +779,7 @@ class BertForPreTraining(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForPreTraining.from_pretrained('bert-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)
         prediction_scores, seq_relationship_scores = outputs[:2]
 
@@ -679,23 +791,19 @@ class BertForPreTraining(BertPreTrainedModel):
         self.cls = BertPreTrainingHeads(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.cls.predictions.decoder,
-                                   self.bert.embeddings.word_embeddings)
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 masked_lm_labels=None, next_sentence_label=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         sequence_output, pooled_output = outputs[:2]
         prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
@@ -713,7 +821,8 @@ class BertForPreTraining(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a `language modeling` head on top. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForMaskedLM(BertPreTrainedModel):
     r"""
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
@@ -721,10 +830,17 @@ class BertForMaskedLM(BertPreTrainedModel):
             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]``
+        **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
+            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) ``torch.FloatTensor`` of shape ``(1,)``:
+        **masked_lm_loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
             Masked language modeling loss.
+        **ltr_lm_loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Next token prediction loss.
         **prediction_scores**: ``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).
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -739,7 +855,7 @@ class BertForMaskedLM(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForMaskedLM.from_pretrained('bert-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]
 
@@ -751,38 +867,52 @@ class BertForMaskedLM(BertPreTrainedModel):
         self.cls = BertOnlyMLMHead(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.cls.predictions.decoder,
-                                   self.bert.embeddings.word_embeddings)
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
-                masked_lm_labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
+                masked_lm_labels=None, encoder_hidden_states=None, encoder_attention_mask=None, lm_labels=None, ):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds,
+                            encoder_hidden_states=encoder_hidden_states,
+                            encoder_attention_mask=encoder_attention_mask)
 
         sequence_output = outputs[0]
         prediction_scores = self.cls(sequence_output)
 
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+
+        # Although this may seem awkward, BertForMaskedLM supports two scenarios:
+        # 1. If a tensor that contains the indices of masked labels is provided,
+        #    the cross-entropy is the MLM cross-entropy that measures the likelihood
+        #    of predictions for masked words.
+        # 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)
+            loss_fct = CrossEntropyLoss(ignore_index=-1)  # -1 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
 
-        return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+        if lm_labels is not None:
+            # 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)
+            ltr_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), lm_labels.view(-1))
+            outputs = (ltr_lm_loss,) + outputs
+
+        return outputs  # (masked_lm_loss), (ltr_lm_loss), prediction_scores, (hidden_states), (attentions)
 
 
 @add_start_docstrings("""Bert Model with a `next sentence prediction (classification)` head on top. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForNextSentencePrediction(BertPreTrainedModel):
     r"""
         **next_sentence_label**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -808,7 +938,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForNextSentencePrediction.from_pretrained('bert-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)
         seq_relationship_scores = outputs[0]
 
@@ -821,14 +951,15 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 next_sentence_label=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         pooled_output = outputs[1]
 
@@ -844,8 +975,9 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of
-    the pooled output) e.g. for GLUE tasks. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the pooled output) e.g. for GLUE tasks. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForSequenceClassification(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -871,7 +1003,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForSequenceClassification.from_pretrained('bert-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]
@@ -887,14 +1019,15 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         pooled_output = outputs[1]
 
@@ -917,8 +1050,9 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a multiple choice classification head on top (a linear layer on top of
-    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForMultipleChoice(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -945,7 +1079,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForMultipleChoice.from_pretrained('bert-base-uncased')
         choices = ["Hello, my dog is cute", "Hello, my cat is amazing"]
-        input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        input_ids = torch.tensor([tokenizer.encode(s, add_special_tokens=True) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
         labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, classification_scores = outputs[:2]
@@ -960,8 +1094,8 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
         num_choices = input_ids.shape[1]
 
         input_ids = input_ids.view(-1, input_ids.size(-1))
@@ -973,7 +1107,8 @@ class BertForMultipleChoice(BertPreTrainedModel):
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
                             position_ids=position_ids,
-                            head_mask=head_mask)
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         pooled_output = outputs[1]
 
@@ -992,8 +1127,9 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a token classification head on top (a linear layer on top of
-    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForTokenClassification(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
@@ -1017,7 +1153,7 @@ class BertForTokenClassification(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForTokenClassification.from_pretrained('bert-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] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, scores = outputs[:2]
@@ -1033,14 +1169,15 @@ class BertForTokenClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         sequence_output = outputs[0]
 
@@ -1064,8 +1201,9 @@ class BertForTokenClassification(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert 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`). """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the hidden-states output to compute `span start logits` and `span end logits`). """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForQuestionAnswering(BertPreTrainedModel):
     r"""
         **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -1099,9 +1237,9 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
         input_ids = tokenizer.encode(input_text)
-        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]
         start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
-        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
         print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
         # a nice puppet
 
@@ -1116,14 +1254,15 @@ class BertForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 start_positions=None, end_positions=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         sequence_output = outputs[0]
 

transformers/modeling_camembert.py

+# coding=utf-8
+# Copyright 2019 Inria, Facebook AI Research and the HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch CamemBERT model. """
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+
+from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForMultipleChoice, RobertaForTokenClassification
+from .configuration_camembert import CamembertConfig
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'camembert-base': "https://s3.amazonaws.com/models.huggingface.co/bert/camembert-base-pytorch_model.bin",
+}
+
+
+CAMEMBERT_START_DOCSTRING = r"""    The CamemBERT model was proposed in
+    `CamemBERT: a Tasty French Language Model`_
+    by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suárez, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah, and Benoît Sagot. It is based on Facebook's RoBERTa model released in 2019.
+    
+    It is a model trained on 138GB of French text.
+    
+    This implementation is the same as RoBERTa.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`CamemBERT: a Tasty French Language Model`:
+        https://arxiv.org/abs/1911.03894
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.CamembertConfig`): Model configuration class with all the parameters of the 
+            model. Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+CAMEMBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, CamemBERT input sequence should be formatted with <s> and </s> tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         <s> Is this Jacksonville ? </s> </s> No it is not . </s>``
+
+            (b) For single sequences:
+
+                ``tokens:         <s> the dog is hairy . </s>``
+
+            Fully encoded sequences or sequence pairs can be obtained using the CamembertTokenizer.encode function with 
+            the ``add_special_tokens`` parameter set to ``True``.
+
+            CamemBERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional` need to be trained) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Optional segment token indices to indicate first and second portions of the inputs.
+            This embedding matrice is not trained (not pretrained during CamemBERT pretraining), you will have to train it
+            during finetuning.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+"""
+
+@add_start_docstrings("The bare CamemBERT Model transformer outputting raw hidden-states without any specific head on top.",
+                      CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertModel(RobertaModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            eo match pre-training, CamemBERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] [SEP] no it is not . [SEP]``
+
+                ``token_type_ids:   0   0  0    0    0     0       0   0   0     1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertModel.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).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
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a `language modeling` head on top. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForMaskedLM(RobertaForMaskedLM):
+    r"""
+        **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
+            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) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``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).
+        **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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForMaskedLM.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model transformer with a sequence classification/regression head on top (a linear layer 
+    on top of the pooled output) e.g. for GLUE tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForSequenceClassification(RobertaForSequenceClassification):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) 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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForSequenceClassification.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForMultipleChoice(RobertaForMultipleChoice):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above).
+            Classification 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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForMultipleChoice.from_pretrained('camembert-base')
+        choices = ["J'aime le camembert !", "Je deteste le camembert !"]
+        input_ids = torch.tensor([tokenizer.encode(s, add_special_tokens=True) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, classification_scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForTokenClassification(RobertaForTokenClassification):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification 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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForTokenClassification.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP

transformers/modeling_ctrl.py

@@ -63,7 +63,8 @@ def scaled_dot_product_attention(q, k, v, mask, attention_mask=None, head_mask=N
     scaled_attention_logits = matmul_qk / np.sqrt(dk)
 
     if mask is not None:
-        scaled_attention_logits += (mask * -1e4)
+        nd, ns = scaled_attention_logits.size(-2), scaled_attention_logits.size(-1)
+        scaled_attention_logits += (mask[ns-nd:ns, :ns] * -1e4)
 
     if attention_mask is not None:
         # Apply the attention mask
@@ -220,7 +221,8 @@ CTRL_INPUTS_DOCSTRING = r"""    Inputs:
         **past**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `past` output below). Can be used to speed up sequential decoding.
+            (see `past` output below). Can be used to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -236,6 +238,10 @@ CTRL_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare CTRL Model transformer outputting raw hidden-states without any specific head on top.",
@@ -246,9 +252,10 @@ class CTRLModel(CTRLPreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **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)``:
@@ -289,10 +296,12 @@ class CTRLModel(CTRLPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        self.w = self._get_resized_embeddings(self.w, new_num_tokens)
+    def get_input_embeddings(self):
         return self.w
 
+    def set_input_embeddings(self, new_embeddings):
+        self.w = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
                 heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -300,17 +309,26 @@ class CTRLModel(CTRLPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.h[layer].attn.prune_heads(heads)
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
-        input_shape = input_ids.size()
-        input_ids = input_ids.view(-1, input_shape[-1])
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if past is None:
             past_length = 0
             past = [None] * len(self.h)
         else:
             past_length = past[0][0].size(-2)
         if position_ids is None:
-            position_ids = torch.arange(past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
 
         # Attention mask.
         if attention_mask is not None:
@@ -352,10 +370,11 @@ class CTRLModel(CTRLPreTrainedModel):
             token_type_embeds = 0
         position_ids = position_ids.view(-1, input_shape[-1])
 
-        inputs_embeds = self.w(input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = self.w(input_ids)
         # inputs_embeds = embedded.unsqueeze(0) if len(input_ids.shape)<2 else embedded
-        seq_len = input_ids.shape[-1]
-        mask = torch.triu(torch.ones(seq_len, seq_len), 1).to(inputs_embeds.device)
+        seq_len = input_shape[-1]
+        mask = torch.triu(torch.ones(seq_len + past_length, seq_len + past_length), 1).to(inputs_embeds.device)
 
         inputs_embeds *= np.sqrt(self.d_model_size)
 
@@ -419,9 +438,10 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         **prediction_scores**: ``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).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **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)``:
@@ -449,22 +469,19 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=True)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-                Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head, self.transformer.w)
+    def get_output_embeddings(self):
+        return self.lm_head
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                past=past,
                                                attention_mask=attention_mask,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         hidden_states = transformer_outputs[0]