Merge branch 'master' into distilbert-german

transformers/configuration_albert.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.
+""" ALBERT model configuration """
+
+from .configuration_utils import PretrainedConfig
+
+ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-config.json",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-config.json",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-config.json",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-config.json",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-config.json",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-config.json",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-config.json",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-config.json",
+}
+
+class AlbertConfig(PretrainedConfig):
+    """Configuration for `AlbertModel`.
+
+    The default settings match the configuration of model `albert_xxlarge`.
+    """
+
+    pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size_or_config_json_file=30000,
+                 embedding_size=128,
+                 hidden_size=4096,
+                 num_hidden_layers=12,
+                 num_hidden_groups=1,
+                 num_attention_heads=64,
+                 intermediate_size=16384,
+                 inner_group_num=1,
+                 hidden_act="gelu_new",
+                 hidden_dropout_prob=0,
+                 attention_probs_dropout_prob=0,
+                 max_position_embeddings=512,
+                 type_vocab_size=2,
+                 initializer_range=0.02,
+                 layer_norm_eps=1e-12, **kwargs):
+        """Constructs AlbertConfig.
+
+        Args:
+            vocab_size: Vocabulary size of `inputs_ids` in `AlbertModel`.
+            embedding_size: size of voc embeddings.
+            hidden_size: Size of the encoder layers and the pooler layer.
+            num_hidden_layers: Number of hidden layers in the Transformer encoder.
+            num_hidden_groups: Number of group for the hidden layers, parameters in
+                the same group are shared.
+            num_attention_heads: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
+                layer in the Transformer encoder.
+            inner_group_num: int, number of inner repetition of attention and ffn.
+            down_scale_factor: float, the scale to apply
+            hidden_act: The non-linear activation function (function or string) in the
+                encoder and pooler.
+            hidden_dropout_prob: The dropout probability for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob: The dropout ratio for the attention
+                probabilities.
+            max_position_embeddings: The maximum sequence length that this model might
+                ever be used with. Typically set this to something large just in case
+                (e.g., 512 or 1024 or 2048).
+            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
+                `AlbertModel`.
+            initializer_range: The stdev of the truncated_normal_initializer for
+                initializing all weight matrices.
+        """
+        super(AlbertConfig, self).__init__(**kwargs)
+
+        self.vocab_size = vocab_size_or_config_json_file
+        self.embedding_size = embedding_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_hidden_groups = num_hidden_groups
+        self.num_attention_heads = num_attention_heads
+        self.inner_group_num = inner_group_num
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
\ No newline at end of file

transformers/configuration_auto.py

@@ -95,6 +95,9 @@ class AutoConfig(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.

transformers/configuration_distilbert.py

@@ -29,6 +29,7 @@ DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-config.json",
     'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-config.json",
     'distilbert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-german-cased-config.json",
+    'distilbert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-multilingual-cased-config.json",
 }
 
 

transformers/configuration_utils.py

@@ -94,6 +94,9 @@ 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.
@@ -120,6 +123,7 @@ 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)
 
@@ -131,7 +135,8 @@ class PretrainedConfig(object):
             config_file = pretrained_model_name_or_path
         # 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(

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):

transformers/data/__init__.py

 from .processors import InputExample, InputFeatures, DataProcessor
 from .processors import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
+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/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 .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels

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

@@ -22,6 +22,7 @@ from botocore.config import Config
 from botocore.exceptions import ClientError
 import requests
 from tqdm import tqdm
+from contextlib import contextmanager
 
 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
@@ -152,7 +153,7 @@ 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 +162,7 @@ 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
@@ -173,7 +175,9 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
 
     if parsed.scheme in ('http', 'https', 's3'):
         # 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
@@ -234,19 +238,22 @@ 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", total=total, initial=resume_size)
+    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, etag_timeout=10):
+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.
@@ -289,17 +296,35 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag
         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/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
+        name = name.replace("ffn_1", "ffn")
+        name = name.replace("/bert/", "/albert/")
+        name = name.replace("ffn/intermediate/output", "ffn_output")
+        name = name.replace("attention_1", "attention")   
+        name = name.replace("cls/predictions", "predictions")
+        name = name.replace("transform/", "")
+        name = name.replace("LayerNorm_1", "full_layer_layer_norm")    
+        name = name.replace("LayerNorm", "attention/LayerNorm")    
+        name = name.replace("inner_group_", "albert_layers/") 
+        name = name.replace("group_", "albert_layer_groups/")    
+        name = name.split('/')
+        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 +27,7 @@ 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_utils import PreTrainedModel, SequenceSummary
 
@@ -48,6 +49,7 @@ 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 `camembert`: CamembertModel (CamemBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -71,6 +73,7 @@ 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 `camembert`: CamembertModel (CamemBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -112,6 +115,9 @@ 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 +144,8 @@ class AutoModel(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertModel.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:
@@ -172,6 +180,7 @@ 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 `camembert`: CamembertForMaskedLM (CamemBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -198,6 +207,7 @@ 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 `camembert`: CamembertForMaskedLM (CamemBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -237,6 +247,8 @@ 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 +276,8 @@ class AutoModelWithLMHead(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForMaskedLM.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:
@@ -298,6 +312,7 @@ 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 `camembert`: CamembertForSequenceClassification (CamemBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -320,6 +335,7 @@ 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 `camembert`: CamembertForSequenceClassification (CamemBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -357,6 +373,9 @@ 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 +402,8 @@ class AutoModelForSequenceClassification(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForSequenceClassification.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:

transformers/modeling_bert.py

@@ -278,7 +278,7 @@ 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)
@@ -597,7 +597,7 @@ 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
 
@@ -760,7 +760,7 @@ 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]
 
@@ -836,7 +836,7 @@ 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]
 
@@ -919,7 +919,7 @@ 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]
 
@@ -984,7 +984,7 @@ 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]
@@ -1060,7 +1060,7 @@ 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]
@@ -1134,7 +1134,7 @@ 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]

transformers/modeling_camembert.py

@@ -20,7 +20,7 @@ from __future__ import (absolute_import, division, print_function,
 
 import logging
 
-from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForMultipleChoice
+from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForMultipleChoice, RobertaForTokenClassification
 from .configuration_camembert import CamembertConfig
 from .file_utils import add_start_docstrings
 
@@ -255,3 +255,39 @@ class CamembertForMultipleChoice(RobertaForMultipleChoice):
     """
     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_distilbert.py

@@ -44,6 +44,7 @@ DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-pytorch_model.bin",
     'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-pytorch_model.bin",
     'distilbert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-german-cased-pytorch_model.bin",
+    'distilbert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-multilingual-cased-pytorch_model.bin",
 }
 
 

transformers/modeling_tf_albert.py

+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and 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.
+""" TF 2.0 ALBERT model. """
+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 numpy as np
+import tensorflow as tf
+
+from .configuration_albert import AlbertConfig
+from .modeling_tf_utils import TFPreTrainedModel, get_initializer
+from .modeling_tf_bert import ACT2FN, TFBertSelfAttention
+from .file_utils import add_start_docstrings
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-tf_model.h5",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-tf_model.h5",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-tf_model.h5",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-tf_model.h5",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-tf_model.h5",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-tf_model.h5",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-tf_model.h5",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-tf_model.h5",
+}
+
+
+class TFAlbertEmbeddings(tf.keras.layers.Layer):
+    """Construct the embeddings from word, position and token_type embeddings.
+    """
+
+    def __init__(self, config, **kwargs):
+        super(TFAlbertEmbeddings, self).__init__(**kwargs)
+
+        self.config = config
+        self.position_embeddings = tf.keras.layers.Embedding(config.max_position_embeddings,
+                                                             config.embedding_size,
+                                                             embeddings_initializer=get_initializer(
+                                                                 self.config.initializer_range),
+                                                             name='position_embeddings')
+        self.token_type_embeddings = tf.keras.layers.Embedding(config.type_vocab_size,
+                                                               config.embedding_size,
+                                                               embeddings_initializer=get_initializer(
+                                                                   self.config.initializer_range),
+                                                               name='token_type_embeddings')
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def build(self, input_shape):
+        """Build shared word embedding layer """
+        with tf.name_scope("word_embeddings"):
+            # Create and initialize weights. The random normal initializer was chosen
+            # arbitrarily, and works well.
+            self.word_embeddings = self.add_weight(
+                "weight",
+                shape=[self.config.vocab_size, self.config.embedding_size],
+                initializer=get_initializer(self.config.initializer_range))
+        super(TFAlbertEmbeddings, self).build(input_shape)
+
+    def call(self, inputs, mode="embedding", training=False):
+        """Get token embeddings of inputs.
+        Args:
+            inputs: list of three int64 tensors with shape [batch_size, length]: (input_ids, position_ids, token_type_ids)
+            mode: string, a valid value is one of "embedding" and "linear".
+        Returns:
+            outputs: (1) If mode == "embedding", output embedding tensor, float32 with
+                shape [batch_size, length, embedding_size]; (2) mode == "linear", output
+                linear tensor, float32 with shape [batch_size, length, vocab_size].
+        Raises:
+            ValueError: if mode is not valid.
+
+        Shared weights logic adapted from
+            https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
+        """
+        if mode == "embedding":
+            return self._embedding(inputs, training=training)
+        elif mode == "linear":
+            return self._linear(inputs)
+        else:
+            raise ValueError("mode {} is not valid.".format(mode))
+
+    def _embedding(self, inputs, training=False):
+        """Applies embedding based on inputs tensor."""
+        input_ids, position_ids, token_type_ids, inputs_embeds = inputs
+
+        if input_ids is not None:
+            input_shape = tf.shape(input_ids)
+        else:
+            input_shape = tf.shape(inputs_embeds)[:-1]
+
+        seq_length = input_shape[1]
+        if position_ids is None:
+            position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
+        if token_type_ids is None:
+            token_type_ids = tf.fill(input_shape, 0)
+
+        if inputs_embeds is None:
+            inputs_embeds = tf.gather(self.word_embeddings, input_ids)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings, training=training)
+        return embeddings
+
+    def _linear(self, inputs):
+        """Computes logits by running inputs through a linear layer.
+            Args:
+                inputs: A float32 tensor with shape [batch_size, length, embedding_size]
+            Returns:
+                float32 tensor with shape [batch_size, length, vocab_size].
+        """
+        batch_size = tf.shape(inputs)[0]
+        length = tf.shape(inputs)[1]
+        x = tf.reshape(inputs, [-1, self.config.embedding_size])
+        logits = tf.matmul(x, self.word_embeddings, transpose_b=True)
+        return tf.reshape(logits, [batch_size, length, self.config.vocab_size])
+
+
+class TFAlbertSelfAttention(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertSelfAttention, self).__init__(**kwargs)
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
+        self.output_attentions = config.output_attentions
+
+        self.num_attention_heads = config.num_attention_heads
+        assert config.hidden_size % config.num_attention_heads == 0
+        self.attention_head_size = int(
+            config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = tf.keras.layers.Dense(self.all_head_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='query')
+        self.key = tf.keras.layers.Dense(self.all_head_size,
+                                         kernel_initializer=get_initializer(
+                                             config.initializer_range),
+                                         name='key')
+        self.value = tf.keras.layers.Dense(self.all_head_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='value')
+
+        self.dropout = tf.keras.layers.Dropout(
+            config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, batch_size):
+        x = tf.reshape(
+            x, (batch_size, -1, self.num_attention_heads, self.attention_head_size))
+        return tf.transpose(x, perm=[0, 2, 1, 3])
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        batch_size = tf.shape(hidden_states)[0]
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        # scale attention_scores
+        dk = tf.cast(tf.shape(key_layer)[-1], tf.float32)
+        attention_scores = attention_scores / tf.math.sqrt(dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFAlbertModel call() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = tf.nn.softmax(attention_scores, axis=-1)
+
+        # 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, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+        context_layer = tf.reshape(context_layer,
+                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
+
+        outputs = (context_layer, attention_probs) if self.output_attentions else (
+            context_layer,)
+        return outputs
+
+
+class TFAlbertSelfOutput(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertSelfOutput, self).__init__(**kwargs)
+        self.dense = tf.keras.layers.Dense(config.hidden_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def call(self, inputs, training=False):
+        hidden_states, input_tensor = inputs
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class TFAlbertAttention(TFBertSelfAttention):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertAttention, self).__init__(config, **kwargs)
+
+        self.hidden_size = config.hidden_size
+        self.dense = tf.keras.layers.Dense(config.hidden_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        raise NotImplementedError
+
+    def call(self, inputs, training=False):
+        input_tensor, attention_mask, head_mask = inputs
+
+        batch_size = tf.shape(input_tensor)[0]
+        mixed_query_layer = self.query(input_tensor)
+        mixed_key_layer = self.key(input_tensor)
+        mixed_value_layer = self.value(input_tensor)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        # scale attention_scores
+        dk = tf.cast(tf.shape(key_layer)[-1], tf.float32)
+        attention_scores = attention_scores / tf.math.sqrt(dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFBertModel call() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = tf.nn.softmax(attention_scores, axis=-1)
+
+        # 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, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+        context_layer = tf.reshape(context_layer,
+                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
+
+        self_outputs = (context_layer, attention_probs) if self.output_attentions else (
+            context_layer,)
+
+        hidden_states = self_outputs[0]
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        attention_output = self.LayerNorm(hidden_states + input_tensor)
+
+        # add attentions if we output them
+        outputs = (attention_output,) + self_outputs[1:]
+        return outputs
+
+
+class TFAlbertLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertLayer, self).__init__(**kwargs)
+        self.attention = TFAlbertAttention(config, name='attention')
+
+        self.ffn = tf.keras.layers.Dense(config.intermediate_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='ffn')
+
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+        self.ffn_output = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='ffn_output')
+        self.full_layer_layer_norm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='full_layer_layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        attention_outputs = self.attention(
+            [hidden_states, attention_mask, head_mask], training=training)
+        ffn_output = self.ffn(attention_outputs[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = self.full_layer_layer_norm(
+            ffn_output + attention_outputs[0])
+
+        # add attentions if we output them
+        outputs = (hidden_states,) + attention_outputs[1:]
+        return outputs
+
+
+class TFAlbertLayerGroup(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertLayerGroup, self).__init__(**kwargs)
+
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.albert_layers = [TFAlbertLayer(config, name="albert_layers_._{}".format(
+            i)) for i in range(config.inner_group_num)]
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        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]], training=training)
+            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,)
+        # last-layer hidden state, (layer hidden states), (layer attentions)
+        return outputs
+
+
+class TFAlbertTransformer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertTransformer, self).__init__(**kwargs)
+
+        self.config = config
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.embedding_hidden_mapping_in = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='embedding_hidden_mapping_in')
+        self.albert_layer_groups = [TFAlbertLayerGroup(
+            config, name="albert_layer_groups_._{}".format(i)) for i in range(config.num_hidden_groups)]
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        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))
+
+            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]], training=training)
+            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,)
+
+        # last-layer hidden state, (all hidden states), (all attentions)
+        return outputs
+
+
+class TFAlbertPreTrainedModel(TFPreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = AlbertConfig
+    pretrained_model_archive_map = TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "albert"
+
+
+class TFAlbertMLMHead(tf.keras.layers.Layer):
+    def __init__(self, config, input_embeddings, **kwargs):
+        super(TFAlbertMLMHead, self).__init__(**kwargs)
+        self.vocab_size = config.vocab_size
+
+        self.dense = tf.keras.layers.Dense(config.embedding_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = input_embeddings
+
+    def build(self, input_shape):
+        self.bias = self.add_weight(shape=(self.vocab_size,),
+                                    initializer='zeros',
+                                    trainable=True,
+                                    name='bias')
+        self.decoder_bias = self.add_weight(shape=(self.vocab_size,),
+                                    initializer='zeros',
+                                    trainable=True,
+                                    name='decoder/bias')
+        super(TFAlbertMLMHead, self).build(input_shape)
+
+    def call(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, mode="linear") + self.decoder_bias
+        hidden_states = hidden_states + self.bias
+        return hidden_states
+
+
+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 tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
+    refer to the TF 2.0 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
+
+    .. _`tf.keras.Model`:
+        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
+
+    Note on the model inputs:
+        TF 2.0 models accepts two formats as inputs:
+
+            - having all inputs as keyword arguments (like PyTorch models), or
+            - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
+
+        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
+            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
+
+    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**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, ALBERT 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`) ``Numpy array`` or ``tf.Tensor`` 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`) ``Numpy array`` or ``tf.Tensor`` 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 `ALBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` 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`) ``Numpy array`` or ``tf.Tensor`` 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 outputing raw hidden-states without any specific head on top.",
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertModel(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``tf.Tensor`` 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 Albert 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 ``tf.Tensor`` (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 ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertModel
+
+        tokenizer = AlbertTokenizer.from_pretrained('bert-base-uncased')
+        model = TFAlbertModel.from_pretrained('bert-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+
+    def __init__(self, config, **kwargs):
+        super(TFAlbertModel, self).__init__(config, **kwargs)
+        self.num_hidden_layers = config.num_hidden_layers
+
+        self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
+        self.encoder = TFAlbertTransformer(config, name="encoder")
+        self.pooler = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), activation='tanh', name='pooler')
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        raise NotImplementedError
+
+    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}
+            See base class PreTrainedModel
+        """
+        raise NotImplementedError
+
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
+        if isinstance(inputs, (tuple, list)):
+            input_ids = inputs[0]
+            attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
+            token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
+            position_ids = inputs[3] if len(inputs) > 3 else position_ids
+            head_mask = inputs[4] if len(inputs) > 4 else head_mask
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
+        elif isinstance(inputs, dict):
+            input_ids = inputs.get('input_ids')
+            attention_mask = inputs.get('attention_mask', attention_mask)
+            token_type_ids = inputs.get('token_type_ids', token_type_ids)
+            position_ids = inputs.get('position_ids', position_ids)
+            head_mask = inputs.get('head_mask', head_mask)
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
+        else:
+            input_ids = inputs
+
+        if input_ids is not None 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 = tf.shape(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.shape[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if attention_mask is None:
+            attention_mask = tf.fill(input_shape, 1)
+        if token_type_ids is None:
+            token_type_ids = tf.fill(input_shape, 0)
+
+        # 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[:, tf.newaxis, tf.newaxis, :]
+
+        # 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 = tf.cast(extended_attention_mask, tf.float32)
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # 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
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if not head_mask is None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            [input_ids, position_ids, token_type_ids, inputs_embeds], training=training)
+        encoder_outputs = self.encoder(
+            [embedding_output, extended_attention_mask, head_mask], training=training)
+
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output[:, 0])
+
+        # add hidden_states and attentions if they are here
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]
+        # sequence_output, pooled_output, (hidden_states), (attentions)
+        return outputs
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top. """,
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertForMaskedLM(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **prediction_scores**: ``Numpy array`` or ``tf.Tensor`` 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 ``Numpy array`` or ``tf.Tensor`` (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 ``Numpy array`` or ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForMaskedLM
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForMaskedLM.from_pretrained('albert-base-v2')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores = outputs[0]
+
+    """
+
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFAlbertForMaskedLM, self).__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertModel(config, name='albert')
+        self.predictions = TFAlbertMLMHead(
+            config, self.albert.embeddings, name='predictions')
+
+    def get_output_embeddings(self):
+        return self.albert.embeddings
+
+    def call(self, inputs, **kwargs):
+        outputs = self.albert(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+        prediction_scores = self.predictions(
+            sequence_output, training=kwargs.get('training', False))
+
+        # Add hidden states and attention if they are here
+        outputs = (prediction_scores,) + outputs[2:]
+
+        return outputs  # prediction_scores, (hidden_states), (attentions)
+
+
+@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 TFAlbertForSequenceClassification(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **logits**: ``Numpy array`` or ``tf.Tensor`` 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 ``Numpy array`` or ``tf.Tensor`` (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 ``Numpy array`` or ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForSequenceClassification
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForSequenceClassification.from_pretrained('albert-base-v2')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        logits = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFAlbertForSequenceClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertModel(config, name='albert')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.albert(inputs, **kwargs)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output, training=kwargs.get('training', False))
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # logits, (hidden_states), (attentions)
\ No newline at end of file

transformers/modeling_tf_auto.py

@@ -109,6 +109,9 @@ class TFAutoModel(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.
@@ -237,6 +240,9 @@ class TFAutoModelWithLMHead(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.
@@ -360,6 +366,9 @@ class TFAutoModelForSequenceClassification(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.
@@ -472,6 +481,9 @@ class TFAutoModelForQuestionAnswering(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.

transformers/modeling_tf_utils.py

@@ -191,6 +191,9 @@ class TFPreTrainedModel(tf.keras.Model):
             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.
@@ -216,6 +219,7 @@ class TFPreTrainedModel(tf.keras.Model):
         cache_dir = kwargs.pop('cache_dir', None)
         from_pt = kwargs.pop('from_pt', False)
         force_download = kwargs.pop('force_download', False)
+        resume_download = kwargs.pop('resume_download', False)
         proxies = kwargs.pop('proxies', None)
 
         # Load config
@@ -224,6 +228,7 @@ class TFPreTrainedModel(tf.keras.Model):
                 pretrained_model_name_or_path, *model_args,
                 cache_dir=cache_dir, return_unused_kwargs=True,
                 force_download=force_download,
+                resume_download=resume_download,
                 **kwargs
             )
         else:
@@ -251,7 +256,8 @@ class TFPreTrainedModel(tf.keras.Model):
 
             # redirect to the cache, if necessary
             try:
-                resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+                resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download,
+                                                    resume_download=resume_download, proxies=proxies)
             except EnvironmentError as e:
                 if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
                     logger.error(

transformers/modeling_utils.py

@@ -291,6 +291,9 @@ class PreTrainedModel(nn.Module):
             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.
@@ -315,11 +318,16 @@ class PreTrainedModel(nn.Module):
             model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
+        if "albert" in pretrained_model_name_or_path and "v2" in pretrained_model_name_or_path:
+            logger.warning("There is currently an upstream reproducibility issue with ALBERT v2 models. Please see " +
+                           "https://github.com/google-research/google-research/issues/119 for more information.")
+
         config = kwargs.pop('config', None)
         state_dict = kwargs.pop('state_dict', None)
         cache_dir = kwargs.pop('cache_dir', None)
         from_tf = kwargs.pop('from_tf', False)
         force_download = kwargs.pop('force_download', False)
+        resume_download = kwargs.pop('resume_download', False)
         proxies = kwargs.pop('proxies', None)
         output_loading_info = kwargs.pop('output_loading_info', False)
 
@@ -329,6 +337,7 @@ class PreTrainedModel(nn.Module):
                 pretrained_model_name_or_path, *model_args,
                 cache_dir=cache_dir, return_unused_kwargs=True,
                 force_download=force_download,
+                resume_download=resume_download,
                 proxies=proxies,
                 **kwargs
             )
@@ -361,7 +370,8 @@ class PreTrainedModel(nn.Module):
 
             # redirect to the cache, if necessary
             try:
-                resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+                resolved_archive_file = cached_path(archive_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_model_archive_map:
                     msg = "Couldn't reach server at '{}' to download pretrained weights.".format(