Merge branch 'master' into check-link-validity

transformers/modeling_ctrl.py

@@ -252,7 +252,7 @@ class CTRLModel(CTRLPreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -438,7 +438,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.

transformers/modeling_gpt2.py

@@ -329,7 +329,7 @@ class GPT2Model(GPT2PreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -503,7 +503,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -596,7 +596,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         **mc_prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)``
             Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.

transformers/modeling_openai.py

@@ -50,8 +50,10 @@ def load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path):
 
     logger.info("Loading weights from {}".format(openai_checkpoint_folder_path))
 
-    names = json.load(open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8'))
-    shapes = json.load(open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8'))
+    with open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8') as names_handle:
+        names = json.load(names_handle)
+    with open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8') as shapes_handle:
+        shapes = json.load(shapes_handle)
     offsets = np.cumsum([np.prod(shape) for shape in shapes])
     init_params = [np.load(openai_checkpoint_folder_path + '/params_{}.npy'.format(n)) for n in range(10)]
     init_params = np.split(np.concatenate(init_params, 0), offsets)[:-1]

transformers/modeling_tf_bert.py

@@ -48,6 +48,10 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-tf_model.h5",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-tf_model.h5",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-tf_model.h5",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-tf_model.h5",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-tf_model.h5"
 }
 
 
@@ -129,7 +133,7 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
                 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
         """
@@ -148,7 +152,7 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
             input_shape = shape_list(input_ids)
         else:
             input_shape = shape_list(inputs_embeds)[:-1]
-        
+
         seq_length = input_shape[1]
         if position_ids is None:
             position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
@@ -246,7 +250,7 @@ class TFBertSelfAttention(tf.keras.layers.Layer):
         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, 
+        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,)
@@ -591,7 +595,7 @@ BERT_START_DOCSTRING = r"""    The BERT model was proposed in
             `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
 
     Parameters:
-        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
             Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
@@ -605,13 +609,13 @@ BERT_INPUTS_DOCSTRING = r"""
             (a) For sequence pairs:
 
                 ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-                
+
                 ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
 
             (b) For single sequences:
 
                 ``tokens:         [CLS] the dog is hairy . [SEP]``
-                
+
                 ``token_type_ids:   0   0   0   0  0     0   0``
 
             Bert is a model with absolute position embeddings so it's usually advised to pad the inputs on

transformers/modeling_tf_ctrl.py

@@ -400,7 +400,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
         **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -462,7 +462,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **past**:
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)

transformers/modeling_tf_distilbert.py

@@ -704,6 +704,53 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel):
         return outputs  # logits, (hidden_states), (attentions)
 
 
+@add_start_docstrings("""DistilBert 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. """,
+    DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **scores**: ``Numpy array`` or ``tf.Tensor`` 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 ``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 DistilBertTokenizer, TFDistilBertForTokenClassification
+        tokenizer = DistilBertTokenizer.from_pretrained('bert-base-uncased')
+        model = TFDistilBertForTokenClassification.from_pretrained('bert-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        scores = outputs[0]
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFDistilBertForTokenClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.distilbert = TFDistilBertMainLayer(config, name='distilbert')
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        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.distilbert(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output, training=kwargs.get('training', False))
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # scores, (hidden_states), (attentions)
+
+
 @add_start_docstrings("""DistilBert 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`). """,
                       DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)

transformers/modeling_tf_gpt2.py

@@ -436,7 +436,7 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
         **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -476,7 +476,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
         **prediction_scores**: `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).
         **past**:
-            list of `tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of `tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -535,7 +535,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         **mc_prediction_scores**: `tf.Tensor`` of shape ``(batch_size, num_choices)``
             Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
         **past**:
-            list of `tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of `tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)

transformers/modeling_tf_utils.py

@@ -24,7 +24,8 @@ import os
 import tensorflow as tf
 
 from .configuration_utils import PretrainedConfig
-from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME, TF2_WEIGHTS_NAME
+from .file_utils import (TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, WEIGHTS_NAME,
+                         cached_path, hf_bucket_url, is_remote_url)
 from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
@@ -257,10 +258,14 @@ class TFPreTrainedModel(tf.keras.Model):
                     raise EnvironmentError("Error no file named {} found in directory {} or `from_pt` set to False".format(
                         [WEIGHTS_NAME, TF2_WEIGHTS_NAME],
                         pretrained_model_name_or_path))
-            elif os.path.isfile(pretrained_model_name_or_path):
+            elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
                 archive_file = pretrained_model_name_or_path
+            elif os.path.isfile(pretrained_model_name_or_path + ".index"):
+                archive_file = pretrained_model_name_or_path + ".index"
             else:
-                raise EnvironmentError("Error file {} not found".format(pretrained_model_name_or_path))
+                archive_file = hf_bucket_url(pretrained_model_name_or_path, postfix=TF2_WEIGHTS_NAME)
+                if from_pt:
+                    raise EnvironmentError("Loading a TF model from a PyTorch checkpoint is not supported when using a model identifier name.")
 
             # redirect to the cache, if necessary
             try:

transformers/modeling_utils.py

@@ -31,7 +31,8 @@ from torch.nn import CrossEntropyLoss
 from torch.nn import functional as F
 
 from .configuration_utils import PretrainedConfig
-from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME, TF2_WEIGHTS_NAME
+from .file_utils import (TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, WEIGHTS_NAME,
+                         cached_path, hf_bucket_url, is_remote_url)
 
 logger = logging.getLogger(__name__)
 
@@ -318,7 +319,8 @@ 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:
+        if pretrained_model_name_or_path is not None and (
+                "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.")
 
@@ -362,11 +364,16 @@ class PreTrainedModel(nn.Module):
                     raise EnvironmentError("Error no file named {} found in directory {} or `from_tf` set to False".format(
                         [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index"],
                         pretrained_model_name_or_path))
-            elif os.path.isfile(pretrained_model_name_or_path):
+            elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
                 archive_file = pretrained_model_name_or_path
-            else:
-                assert from_tf, "Error finding file {}, no file or TF 1.X checkpoint found".format(pretrained_model_name_or_path)
+            elif os.path.isfile(pretrained_model_name_or_path + ".index"):
+                assert from_tf, "We found a TensorFlow checkpoint at {}, please set from_tf to True to load from this checkpoint".format(
+                    pretrained_model_name_or_path + ".index")
                 archive_file = pretrained_model_name_or_path + ".index"
+            else:
+                archive_file = hf_bucket_url(pretrained_model_name_or_path, postfix=WEIGHTS_NAME)
+                if from_tf:
+                    raise EnvironmentError("Loading a PyTorch model from a TF checkpoint is not supported when using a model identifier name.")
 
             # redirect to the cache, if necessary
             try:

transformers/optimization_tf.py

+# Copyright 2019 The TensorFlow Authors. 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.
+# ==============================================================================
+"""Functions and classes related to optimization (weight updates)."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import re
+
+import tensorflow as tf
+
+
+class WarmUp(tf.keras.optimizers.schedules.LearningRateSchedule):
+  """Applys a warmup schedule on a given learning rate decay schedule."""
+
+  def __init__(
+      self,
+      initial_learning_rate,
+      decay_schedule_fn,
+      warmup_steps,
+      power=1.0,
+      name=None):
+    super(WarmUp, self).__init__()
+    self.initial_learning_rate = initial_learning_rate
+    self.warmup_steps = warmup_steps
+    self.power = power
+    self.decay_schedule_fn = decay_schedule_fn
+    self.name = name
+
+  def __call__(self, step):
+    with tf.name_scope(self.name or 'WarmUp') as name:
+      # Implements polynomial warmup. i.e., if global_step < warmup_steps, the
+      # learning rate will be `global_step/num_warmup_steps * init_lr`.
+      global_step_float = tf.cast(step, tf.float32)
+      warmup_steps_float = tf.cast(self.warmup_steps, tf.float32)
+      warmup_percent_done = global_step_float / warmup_steps_float
+      warmup_learning_rate = (
+          self.initial_learning_rate *
+          tf.math.pow(warmup_percent_done, self.power))
+      return tf.cond(global_step_float < warmup_steps_float,
+                     lambda: warmup_learning_rate,
+                     lambda: self.decay_schedule_fn(step),
+                     name=name)
+
+  def get_config(self):
+    return {
+        'initial_learning_rate': self.initial_learning_rate,
+        'decay_schedule_fn': self.decay_schedule_fn,
+        'warmup_steps': self.warmup_steps,
+        'power': self.power,
+        'name': self.name
+    }
+
+
+def create_optimizer(init_lr, num_train_steps, num_warmup_steps):
+  """Creates an optimizer with learning rate schedule."""
+  # Implements linear decay of the learning rate.
+  learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay(
+      initial_learning_rate=init_lr,
+      decay_steps=num_train_steps,
+      end_learning_rate=0.0)
+  if num_warmup_steps:
+    learning_rate_fn = WarmUp(initial_learning_rate=init_lr,
+                              decay_schedule_fn=learning_rate_fn,
+                              warmup_steps=num_warmup_steps)
+  optimizer = AdamWeightDecay(
+      learning_rate=learning_rate_fn,
+      weight_decay_rate=0.01,
+      beta_1=0.9,
+      beta_2=0.999,
+      epsilon=1e-6,
+      exclude_from_weight_decay=['layer_norm', 'bias'])
+  return optimizer
+
+
+class AdamWeightDecay(tf.keras.optimizers.Adam):
+  """Adam enables L2 weight decay and clip_by_global_norm on gradients.
+
+  Just adding the square of the weights to the loss function is *not* the
+  correct way of using L2 regularization/weight decay with Adam, since that will
+  interact with the m and v parameters in strange ways.
+
+  Instead we want ot decay the weights in a manner that doesn't interact with
+  the m/v parameters. This is equivalent to adding the square of the weights to
+  the loss with plain (non-momentum) SGD.
+  """
+
+  def __init__(self,
+               learning_rate=0.001,
+               beta_1=0.9,
+               beta_2=0.999,
+               epsilon=1e-7,
+               amsgrad=False,
+               weight_decay_rate=0.0,
+               include_in_weight_decay=None,
+               exclude_from_weight_decay=None,
+               name='AdamWeightDecay',
+               **kwargs):
+    super(AdamWeightDecay, self).__init__(
+        learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs)
+    self.weight_decay_rate = weight_decay_rate
+    self._include_in_weight_decay = include_in_weight_decay
+    self._exclude_from_weight_decay = exclude_from_weight_decay
+
+  @classmethod
+  def from_config(cls, config):
+    """Creates an optimizer from its config with WarmUp custom object."""
+    custom_objects = {'WarmUp': WarmUp}
+    return super(AdamWeightDecay, cls).from_config(
+        config, custom_objects=custom_objects)
+
+  def _prepare_local(self, var_device, var_dtype, apply_state):
+    super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype,
+                                                apply_state)
+    apply_state['weight_decay_rate'] = tf.constant(
+        self.weight_decay_rate, name='adam_weight_decay_rate')
+
+  def _decay_weights_op(self, var, learning_rate, apply_state):
+    do_decay = self._do_use_weight_decay(var.name)
+    if do_decay:
+      return var.assign_sub(
+          learning_rate * var *
+          apply_state['weight_decay_rate'],
+          use_locking=self._use_locking)
+    return tf.no_op()
+
+  def apply_gradients(self, grads_and_vars, clip_norm, name=None):
+    grads, tvars = list(zip(*grads_and_vars))
+    (grads, _) = tf.clip_by_global_norm(grads, clip_norm=clip_norm)
+    return super(AdamWeightDecay, self).apply_gradients(zip(grads, tvars))
+
+  def _get_lr(self, var_device, var_dtype, apply_state):
+    """Retrieves the learning rate with the given state."""
+    if apply_state is None:
+      return self._decayed_lr_t[var_dtype], {}
+
+    apply_state = apply_state or {}
+    coefficients = apply_state.get((var_device, var_dtype))
+    if coefficients is None:
+      coefficients = self._fallback_apply_state(var_device, var_dtype)
+      apply_state[(var_device, var_dtype)] = coefficients
+
+    return coefficients['lr_t'], dict(apply_state=apply_state)
+
+  def _resource_apply_dense(self, grad, var, apply_state=None):
+    lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
+    decay = self._decay_weights_op(var, lr_t, apply_state)
+    with tf.control_dependencies([decay]):
+      return super(AdamWeightDecay, self)._resource_apply_dense(
+          grad, var, **kwargs)
+
+  def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
+    lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
+    decay = self._decay_weights_op(var, lr_t, apply_state)
+    with tf.control_dependencies([decay]):
+      return super(AdamWeightDecay, self)._resource_apply_sparse(
+          grad, var, indices, **kwargs)
+
+  def get_config(self):
+    config = super(AdamWeightDecay, self).get_config()
+    config.update({
+        'weight_decay_rate': self.weight_decay_rate,
+    })
+    return config
+
+  def _do_use_weight_decay(self, param_name):
+    """Whether to use L2 weight decay for `param_name`."""
+    if self.weight_decay_rate == 0:
+      return False
+
+    if self._include_in_weight_decay:
+      for r in self._include_in_weight_decay:
+        if re.search(r, param_name) is not None:
+          return True
+
+    if self._exclude_from_weight_decay:
+      for r in self._exclude_from_weight_decay:
+        if re.search(r, param_name) is not None:
+          return False
+    return True
+
+
+## Inspired from https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/optimizers/utils.py
+class GradientAccumulator(object):
+    """Distribution strategies-aware gradient accumulation utility."""
+
+    def __init__(self):
+        """Initializes the accumulator."""
+        self._gradients = []
+        self._accum_steps = tf.Variable(
+            initial_value=0,
+            dtype=tf.int64,
+            trainable=False,
+            aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA)
+
+    @property
+    def step(self):
+        """Number of accumulated steps."""
+        return self._accum_steps.value()
+
+    @property
+    def gradients(self):
+        """The accumulated gradients."""
+        return list(gradient.value() if gradient is not None else gradient for gradient in self._get_replica_gradients())
+
+    def __call__(self, gradients):
+        """Accumulates :obj:`gradients`."""
+        if not self._gradients:
+            self._gradients.extend([tf.Variable(tf.zeros_like(gradient), trainable=False) if gradient is not None else gradient for gradient in gradients])
+
+        if len(gradients) != len(self._gradients):
+            raise ValueError("Expected %s gradients, but got %d" % (len(self._gradients), len(gradients)))
+
+        for accum_gradient, gradient in zip(self._get_replica_gradients(), gradients):
+            if accum_gradient is not None:
+                accum_gradient.assign_add(gradient)
+
+        self._accum_steps.assign_add(1)
+
+    def reset(self):
+        """Resets the accumulated gradients."""
+        if self._gradients:
+            self._accum_steps.assign(0)
+
+        for gradient in self._get_replica_gradients():
+            if gradient is not None:
+                gradient.assign(tf.zeros_like(gradient))
+
+    def _get_replica_gradients(self):
+        if tf.distribute.has_strategy():
+            # In a replica context, we want to accumulate gradients on each replica
+            # without synchronization, so we directly assign the value of the
+            # current replica.
+            replica_context = tf.distribute.get_replica_context()
+
+            if replica_context is None or tf.distribute.get_strategy().num_replicas_in_sync == 1:
+                return self._gradients
+
+            return (gradient.device_map.select_for_current_replica(gradient.values, replica_context) for gradient in self._gradients)
+        else:
+            return self._gradients

transformers/tests/conftest.py

-# content of conftest.py
-
-import pytest
-
-
-def pytest_addoption(parser):
-    parser.addoption(
-        "--runslow", action="store_true", default=False, help="run slow tests"
-    )
-    parser.addoption(
-        "--use_cuda", action="store_true", default=False, help="run tests on gpu"
-    )
-
-
-def pytest_configure(config):
-    config.addinivalue_line("markers", "slow: mark test as slow to run")
-
-
-def pytest_collection_modifyitems(config, items):
-    if config.getoption("--runslow"):
-        # --runslow given in cli: do not skip slow tests
-        return
-    skip_slow = pytest.mark.skip(reason="need --runslow option to run")
-    for item in items:
-        if "slow" in item.keywords:
-            item.add_marker(skip_slow)
-
-@pytest.fixture
-def use_cuda(request):
-    """ Run test on gpu """
-    return request.config.getoption("--use_cuda")

transformers/tests/modeling_albert_test.py

@@ -18,22 +18,21 @@ from __future__ import print_function
 
 import unittest
 import shutil
-import pytest
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (AlbertConfig, AlbertModel, AlbertForMaskedLM,
                               AlbertForSequenceClassification, AlbertForQuestionAnswering,
                               )
     from transformers.modeling_albert import ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 
+@require_torch
 class AlbertModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (AlbertModel, AlbertForMaskedLM) if is_torch_available() else ()
@@ -133,6 +132,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_albert_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = AlbertModel(config=config)
+            model.to(torch_device)
             model.eval()
             sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
             sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
@@ -150,6 +150,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_albert_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = AlbertForMaskedLM(config=config)
+            model.to(torch_device)
             model.eval()
             loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
             result = {
@@ -163,6 +164,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_albert_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = AlbertForQuestionAnswering(config=config)
+            model.to(torch_device)
             model.eval()
             loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
                                                    start_positions=sequence_labels, end_positions=sequence_labels)
@@ -183,6 +185,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_albert_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = AlbertForSequenceClassification(config)
+            model.to(torch_device)
             model.eval()
             loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
             result = {
@@ -225,7 +228,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_albert_for_sequence_classification(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         cache_dir = "/tmp/transformers_test/"
         for model_name in list(ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_auto_test.py

@@ -18,11 +18,12 @@ from __future__ import print_function
 
 import unittest
 import shutil
-import pytest
 import logging
 
 from transformers import is_torch_available
 
+from .utils import require_torch, slow, SMALL_MODEL_IDENTIFIER
+
 if is_torch_available():
     from transformers import (AutoConfig, BertConfig,
                                     AutoModel, BertModel,
@@ -33,12 +34,11 @@ if is_torch_available():
 
     from .modeling_common_test import (CommonTestCases, ids_tensor)
     from .configuration_common_test import ConfigTester
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 
+@require_torch
 class AutoModelTest(unittest.TestCase):
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
@@ -53,7 +53,7 @@ class AutoModelTest(unittest.TestCase):
             for value in loading_info.values():
                 self.assertEqual(len(value), 0)
 
-    @pytest.mark.slow
+    @slow
     def test_lmhead_model_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
@@ -66,7 +66,7 @@ class AutoModelTest(unittest.TestCase):
             self.assertIsNotNone(model)
             self.assertIsInstance(model, BertForMaskedLM)
 
-    @pytest.mark.slow
+    @slow
     def test_sequence_classification_model_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
@@ -79,7 +79,7 @@ class AutoModelTest(unittest.TestCase):
             self.assertIsNotNone(model)
             self.assertIsInstance(model, BertForSequenceClassification)
 
-    @pytest.mark.slow
+    @slow
     def test_question_answering_model_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
@@ -92,6 +92,11 @@ class AutoModelTest(unittest.TestCase):
             self.assertIsNotNone(model)
             self.assertIsInstance(model, BertForQuestionAnswering)
 
+    def test_from_pretrained_identifier(self):
+        logging.basicConfig(level=logging.INFO)
+        model = AutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER)
+        self.assertIsInstance(model, BertForMaskedLM)
+
 
 if __name__ == "__main__":
     unittest.main()

transformers/tests/modeling_bert_test.py

@@ -18,12 +18,12 @@ from __future__ import print_function
 
 import unittest
 import shutil
-import pytest
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor, floats_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (BertConfig, BertModel, BertForMaskedLM,
@@ -31,11 +31,9 @@ if is_torch_available():
                               BertForQuestionAnswering, BertForSequenceClassification,
                               BertForTokenClassification, BertForMultipleChoice)
     from transformers.modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 
-@pytest.mark.usefixtures("use_cuda")
+@require_torch
 class BertModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (BertModel, BertForMaskedLM, BertForNextSentencePrediction,
@@ -67,7 +65,6 @@ class BertModelTest(CommonTestCases.CommonModelTester):
                      num_labels=3,
                      num_choices=4,
                      scope=None,
-                     device='cpu',
                      ):
             self.parent = parent
             self.batch_size = batch_size
@@ -91,26 +88,25 @@ class BertModelTest(CommonTestCases.CommonModelTester):
             self.num_labels = num_labels
             self.num_choices = num_choices
             self.scope = scope
-            self.device = device
 
         def prepare_config_and_inputs(self):
-            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).to(self.device)
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
             input_mask = None
             if self.use_input_mask:
-                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2).to(self.device)
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
 
             token_type_ids = None
             if self.use_token_type_ids:
-                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size).to(self.device)
+                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
 
             sequence_labels = None
             token_labels = None
             choice_labels = None
             if self.use_labels:
-                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size).to(self.device)
-                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels).to(self.device)
-                choice_labels = ids_tensor([self.batch_size], self.num_choices).to(self.device)
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = BertConfig(
                 vocab_size_or_config_json_file=self.vocab_size,
@@ -144,7 +140,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = BertModel(config=config)
-            model.to(input_ids.device)
+            model.to(torch_device)
             model.eval()
             sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
             sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
@@ -161,6 +157,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_model_as_decoder(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, encoder_hidden_states, encoder_attention_mask):
             model = BertModel(config)
+            model.to(torch_device)
             model.eval()
             sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask)
             sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states)
@@ -177,6 +174,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = BertForMaskedLM(config=config)
+            model.to(torch_device)
             model.eval()
             loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
             result = {
@@ -190,6 +188,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_model_for_masked_lm_as_decoder(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, encoder_hidden_states, encoder_attention_mask):
             model = BertForMaskedLM(config=config)
+            model.to(torch_device)
             model.eval()
             loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask)
             loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels, encoder_hidden_states=encoder_hidden_states)
@@ -204,6 +203,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_for_next_sequence_prediction(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = BertForNextSentencePrediction(config=config)
+            model.to(torch_device)
             model.eval()
             loss, seq_relationship_score = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, next_sentence_label=sequence_labels)
             result = {
@@ -217,6 +217,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_for_pretraining(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = BertForPreTraining(config=config)
+            model.to(torch_device)
             model.eval()
             loss, prediction_scores, seq_relationship_score = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
                                                                     masked_lm_labels=token_labels, next_sentence_label=sequence_labels)
@@ -235,6 +236,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_bert_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = BertForQuestionAnswering(config=config)
+            model.to(torch_device)
             model.eval()
             loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
                                                    start_positions=sequence_labels, end_positions=sequence_labels)
@@ -254,6 +256,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_bert_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = BertForSequenceClassification(config)
+            model.to(torch_device)
             model.eval()
             loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
             result = {
@@ -268,6 +271,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_bert_for_token_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = BertForTokenClassification(config=config)
+            model.to(torch_device)
             model.eval()
             loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
             result = {
@@ -282,6 +286,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_bert_for_multiple_choice(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_choices = self.num_choices
             model = BertForMultipleChoice(config=config)
+            model.to(torch_device)
             model.eval()
             multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
             multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
@@ -313,10 +318,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
     def test_config(self):
         self.config_tester.run_common_tests()
 
-    def test_bert_model(self, use_cuda=False):
-        # ^^ This could be a real fixture
-        if use_cuda:
-            self.model_tester.device = "cuda"
+    def test_bert_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_bert_model(*config_and_inputs)
 
@@ -356,7 +358,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_bert_for_token_classification(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         cache_dir = "/tmp/transformers_test/"
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_common_test.py

@@ -27,10 +27,11 @@ import uuid
 
 import unittest
 import logging
-import pytest
 
 from transformers import is_torch_available
 
+from .utils import require_torch, slow, torch_device
+
 if is_torch_available():
     import torch
     import numpy as np
@@ -38,8 +39,6 @@ if is_torch_available():
     from transformers import (AdaptiveEmbedding, PretrainedConfig, PreTrainedModel,
                                     BertModel, BertConfig, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                     GPT2LMHeadModel, GPT2Config, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 if sys.version_info[0] == 2:
     import cPickle as pickle
@@ -65,6 +64,7 @@ def _config_zero_init(config):
 
 class CommonTestCases:
 
+    @require_torch
     class CommonModelTester(unittest.TestCase):
 
         model_tester = None
@@ -79,6 +79,7 @@ class CommonTestCases:
 
             for model_class in self.all_model_classes:
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
                 with torch.no_grad():
                     outputs = model(**inputs_dict)
@@ -86,12 +87,13 @@ class CommonTestCases:
                 with TemporaryDirectory() as tmpdirname:
                     model.save_pretrained(tmpdirname)
                     model = model_class.from_pretrained(tmpdirname)
+                    model.to(torch_device)
                     with torch.no_grad():
                         after_outputs = model(**inputs_dict)
 
                     # Make sure we don't have nans
-                    out_1 = after_outputs[0].numpy()
-                    out_2 = outputs[0].numpy()
+                    out_1 = after_outputs[0].cpu().numpy()
+                    out_2 = outputs[0].cpu().numpy()
                     out_1 = out_1[~np.isnan(out_1)]
                     out_2 = out_2[~np.isnan(out_2)]
                     max_diff = np.amax(np.abs(out_1 - out_2))
@@ -113,6 +115,7 @@ class CommonTestCases:
 
             for model_class in self.all_model_classes:
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
                 first, second = model(inputs_dict["input_ids"])[0], model(inputs_dict["input_ids"])[0]
                 self.assertEqual(first.ne(second).sum().item(), 0)
@@ -125,6 +128,7 @@ class CommonTestCases:
                 config.output_attentions = True
                 config.output_hidden_states = False
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
                 outputs = model(**inputs_dict)
                 attentions = outputs[-1]
@@ -142,6 +146,7 @@ class CommonTestCases:
                 config.output_attentions = True
                 config.output_hidden_states = True
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
                 outputs = model(**inputs_dict)
                 self.assertEqual(out_len+1, len(outputs))
@@ -181,6 +186,7 @@ class CommonTestCases:
             configs_no_init.torchscript = True
             for model_class in self.all_model_classes:
                 model = model_class(config=configs_no_init)
+                model.to(torch_device)
                 model.eval()
                 inputs = inputs_dict['input_ids']  # Let's keep only input_ids
 
@@ -201,7 +207,10 @@ class CommonTestCases:
                 except ValueError:
                     self.fail("Couldn't load module.")
 
+                model.to(torch_device)
                 model.eval()
+
+                loaded_model.to(torch_device)
                 loaded_model.eval()
 
                 model_params = model.parameters()
@@ -228,11 +237,12 @@ class CommonTestCases:
             configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
             for model_class in self.all_model_classes:
                 model = model_class(config=configs_no_init)
+                model.to(torch_device)
                 model.eval()
 
                 # Prepare head_mask
                 # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
-                head_mask = torch.ones(self.model_tester.num_hidden_layers, self.model_tester.num_attention_heads)
+                head_mask = torch.ones(self.model_tester.num_hidden_layers, self.model_tester.num_attention_heads, device=torch_device)
                 head_mask[0, 0] = 0
                 head_mask[-1, :-1] = 0
                 head_mask.requires_grad_(requires_grad=True)
@@ -282,6 +292,7 @@ class CommonTestCases:
                 config.output_attentions = True
                 config.output_hidden_states = False
                 model = model_class(config=config)
+                model.to(torch_device)
                 model.eval()
                 heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
                                 -1: [0]}
@@ -310,6 +321,7 @@ class CommonTestCases:
                 config.output_attentions = True
                 config.output_hidden_states = False
                 model = model_class(config=config)
+                model.to(torch_device)
                 model.eval()
                 heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
                                 -1: [0]}
@@ -319,6 +331,7 @@ class CommonTestCases:
                     os.makedirs(directory)
                 model.save_pretrained(directory)
                 model = model_class.from_pretrained(directory)
+                model.to(torch_device)
 
                 outputs = model(**inputs_dict)
                 attentions = outputs[-1]
@@ -346,6 +359,7 @@ class CommonTestCases:
                 config.pruned_heads = heads_to_prune
 
                 model = model_class(config=config)
+                model.to(torch_device)
                 model.eval()
 
                 outputs = model(**inputs_dict)
@@ -372,6 +386,7 @@ class CommonTestCases:
                 config.pruned_heads = heads_to_prune
 
                 model = model_class(config=config)
+                model.to(torch_device)
                 model.eval()
 
                 outputs = model(**inputs_dict)
@@ -388,6 +403,7 @@ class CommonTestCases:
                     os.makedirs(directory)
                 model.save_pretrained(directory)
                 model = model_class.from_pretrained(directory)
+                model.to(torch_device)
                 shutil.rmtree(directory)
 
                 outputs = model(**inputs_dict)
@@ -419,6 +435,7 @@ class CommonTestCases:
                 config.output_hidden_states = True
                 config.output_attentions = False
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
                 outputs = model(**inputs_dict)
                 hidden_states = outputs[-1]
@@ -538,6 +555,7 @@ class CommonTestCases:
 
             for model_class in self.all_model_classes:
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
 
                 wte = model.get_input_embeddings()
@@ -628,6 +646,7 @@ class CommonTestCases:
         def create_and_check_base_model(self, config, input_ids, token_type_ids, position_ids,
                                 mc_labels, lm_labels, mc_token_ids):
             model = self.base_model_class(config)
+            model.to(torch_device)
             model.eval()
 
             outputs = model(input_ids, position_ids, token_type_ids)
@@ -643,6 +662,7 @@ class CommonTestCases:
         def create_and_check_lm_head(self, config, input_ids, token_type_ids, position_ids,
                                         mc_labels, lm_labels, mc_token_ids):
             model = self.lm_head_model_class(config)
+            model.to(torch_device)
             model.eval()
             outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
             loss, lm_logits = outputs[:2]
@@ -659,6 +679,7 @@ class CommonTestCases:
                                         mc_labels, lm_labels, mc_token_ids):
             for model_class in self.all_model_classes:
                 model = model_class(config)
+                model.to(torch_device)
                 model.eval()
                 outputs = model(input_ids)
                 presents = outputs[-1]
@@ -671,6 +692,7 @@ class CommonTestCases:
         def create_and_check_double_heads(self, config, input_ids, token_type_ids, position_ids,
                                         mc_labels, lm_labels, mc_token_ids):
             model = self.double_head_model_class(config)
+            model.to(torch_device)
             model.eval()
             outputs = model(input_ids, mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels,
                             token_type_ids=token_type_ids, position_ids=position_ids)
@@ -716,7 +738,7 @@ class CommonTestCases:
                 config_and_inputs = self.prepare_config_and_inputs()
                 self.create_and_check_presents(*config_and_inputs)
 
-        @pytest.mark.slow
+        @slow
         def run_slow_tests(self):
             self.create_and_check_model_from_pretrained()
 
@@ -770,7 +792,7 @@ def ids_tensor(shape, vocab_size, rng=None, name=None):
     for _ in range(total_dims):
         values.append(rng.randint(0, vocab_size - 1))
 
-    return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
+    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
 
 
 def floats_tensor(shape, scale=1.0, rng=None, name=None):
@@ -786,11 +808,12 @@ def floats_tensor(shape, scale=1.0, rng=None, name=None):
     for _ in range(total_dims):
         values.append(rng.random() * scale)
 
-    return torch.tensor(data=values, dtype=torch.float).view(shape).contiguous()
+    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()
 
 
+@require_torch
 class ModelUtilsTest(unittest.TestCase):
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_ctrl_test.py

@@ -16,7 +16,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import pytest
 import shutil
 import pdb
 
@@ -25,13 +24,13 @@ from transformers import is_torch_available
 if is_torch_available():
     from transformers import (CTRLConfig, CTRLModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
                                     CTRLLMHeadModel)
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
 
 
+@require_torch
 class CTRLModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (CTRLModel, CTRLLMHeadModel) if is_torch_available() else ()
@@ -140,6 +139,7 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_ctrl_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = CTRLModel(config=config)
+            model.to(torch_device)
             model.eval()
 
             model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
@@ -157,6 +157,7 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = CTRLLMHeadModel(config)
+            model.to(torch_device)
             model.eval()
 
             loss, lm_logits, _ = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
@@ -202,7 +203,7 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         cache_dir = "/tmp/transformers_test/"
         for model_name in list(CTRL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_distilbert_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import pytest
 
 from transformers import is_torch_available
 
@@ -25,13 +24,13 @@ if is_torch_available():
     from transformers import (DistilBertConfig, DistilBertModel, DistilBertForMaskedLM,
                                     DistilBertForTokenClassification,
                                     DistilBertForQuestionAnswering, DistilBertForSequenceClassification)
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
 
 
+@require_torch
 class DistilBertModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (DistilBertModel, DistilBertForMaskedLM, DistilBertForQuestionAnswering,
@@ -126,6 +125,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_distilbert_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = DistilBertModel(config=config)
+            model.to(torch_device)
             model.eval()
             (sequence_output,) = model(input_ids, input_mask)
             (sequence_output,) = model(input_ids)
@@ -139,6 +139,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_distilbert_for_masked_lm(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = DistilBertForMaskedLM(config=config)
+            model.to(torch_device)
             model.eval()
             loss, prediction_scores = model(input_ids, attention_mask=input_mask, masked_lm_labels=token_labels)
             result = {
@@ -152,6 +153,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_distilbert_for_question_answering(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = DistilBertForQuestionAnswering(config=config)
+            model.to(torch_device)
             model.eval()
             loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, start_positions=sequence_labels, end_positions=sequence_labels)
             result = {
@@ -170,6 +172,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_distilbert_for_sequence_classification(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = DistilBertForSequenceClassification(config)
+            model.to(torch_device)
             model.eval()
             loss, logits = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
             result = {
@@ -184,6 +187,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_distilbert_for_token_classification(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = DistilBertForTokenClassification(config=config)
+            model.to(torch_device)
             model.eval()
 
             loss, logits = model(input_ids, attention_mask=input_mask, labels=token_labels)
@@ -229,7 +233,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_distilbert_for_token_classification(*config_and_inputs)
 
-    # @pytest.mark.slow
+    # @slow
     # def test_model_from_pretrained(self):
     #     cache_dir = "/tmp/transformers_test/"
     #     for model_name in list(DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_encoder_decoder_test.py

@@ -15,19 +15,18 @@
 
 import logging
 import unittest
-import pytest
 
 from transformers import is_torch_available
+from .utils import require_torch, slow
 
 if is_torch_available():
     from transformers import BertModel, BertForMaskedLM, Model2Model
     from transformers.modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 
+@require_torch
 class EncoderDecoderModelTest(unittest.TestCase):
-    @pytest.mark.slow
+    @slow
     def test_model2model_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_gpt2_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import pytest
 import shutil
 
 from transformers import is_torch_available
@@ -25,13 +24,13 @@ from transformers import is_torch_available
 if is_torch_available():
     from transformers import (GPT2Config, GPT2Model, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
                                     GPT2LMHeadModel, GPT2DoubleHeadsModel)
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
 
 
+@require_torch
 class GPT2ModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (GPT2Model, GPT2LMHeadModel, GPT2DoubleHeadsModel) if is_torch_available() else ()
@@ -136,6 +135,7 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_gpt2_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = GPT2Model(config=config)
+            model.to(torch_device)
             model.eval()
 
             model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
@@ -153,6 +153,7 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = GPT2LMHeadModel(config)
+            model.to(torch_device)
             model.eval()
 
             loss, lm_logits, _ = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
@@ -171,6 +172,7 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_double_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args):
             model = GPT2DoubleHeadsModel(config)
+            model.to(torch_device)
             model.eval()
 
 
@@ -235,7 +237,7 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_double_lm_head_model(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         cache_dir = "/tmp/transformers_test/"
         for model_name in list(GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:

transformers/tests/modeling_openai_test.py

@@ -17,7 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import pytest
 import shutil
 
 from transformers import is_torch_available
@@ -25,13 +24,13 @@ from transformers import is_torch_available
 if is_torch_available():
     from transformers import (OpenAIGPTConfig, OpenAIGPTModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                     OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel)
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
 
 
+@require_torch
 class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (OpenAIGPTModel, OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel) if is_torch_available() else ()
@@ -124,6 +123,7 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_openai_gpt_model(self, config, input_ids, head_mask, token_type_ids, *args):
             model = OpenAIGPTModel(config=config)
+            model.to(torch_device)
             model.eval()
 
             model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
@@ -139,6 +139,7 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_lm_head_model(self, config, input_ids, head_mask, token_type_ids, *args):
             model = OpenAIGPTLMHeadModel(config)
+            model.to(torch_device)
             model.eval()
 
             loss, lm_logits = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
@@ -157,6 +158,7 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_double_lm_head_model(self, config, input_ids, head_mask, token_type_ids, *args):
             model = OpenAIGPTDoubleHeadsModel(config)
+            model.to(torch_device)
             model.eval()
 
             loss, lm_logits, mc_logits = model(input_ids, token_type_ids=token_type_ids, lm_labels=input_ids)
@@ -203,7 +205,7 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_double_lm_head_model(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
         cache_dir = "/tmp/transformers_test/"
         for model_name in list(OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]: