Merge branch 'master' into conditional-generation

transformers/modeling_tf_xlnet.py

@@ -30,7 +30,6 @@ import tensorflow as tf
 from .configuration_xlnet import XLNetConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 
 logger = logging.getLogger(__name__)
@@ -41,13 +40,6 @@ TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_xlnet_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)  # build the network
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """ Implementation of the gelu activation function.
         XLNet is using OpenAI GPT's gelu
@@ -362,6 +354,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         super(TFXLNetMainLayer, self).__init__(**kwargs)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.output_past = config.output_past
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -421,9 +414,6 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
 
     def cache_mem(self, curr_out, prev_mem):
         """cache hidden states into memory."""
-        if self.mem_len is None or self.mem_len == 0:
-            return None
-        else:
         if self.reuse_len is not None and self.reuse_len > 0:
             curr_out = curr_out[:self.reuse_len]
 
@@ -546,7 +536,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             raise ValueError('Unsupported attention type: {}'.format(self.attn_type))
 
         # data mask: input mask & perm mask
-        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) "
+        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) " \
             "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
         if input_mask is None and attention_mask is not None:
             input_mask = 1.0 - attention_mask
@@ -632,6 +622,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         hidden_states = []
         for i, layer_module in enumerate(self.layer):
             # cache new mems
+            if self.mem_len is not None and self.mem_len > 0 and self.output_past:
                 new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
             if self.output_hidden_states:
                 hidden_states.append((output_h, output_g) if output_g is not None else output_h)
@@ -650,7 +641,11 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         output = self.dropout(output_g if output_g is not None else output_h, training=training)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (tf.transpose(output, perm=(1, 0, 2)), new_mems)
+        outputs = (tf.transpose(output, perm=(1, 0, 2)),)
+
+        if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+            outputs = outputs + (new_mems,)
+
         if self.output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(tf.transpose(h, perm=(1, 0, 2)) for hs in hidden_states for h in hs)
@@ -661,7 +656,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
             outputs = outputs + (attentions,)
 
-        return outputs  # outputs, new_mems, (hidden_states), (attentions)
+        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
 
 
 class TFXLNetPreTrainedModel(TFPreTrainedModel):
@@ -670,7 +665,6 @@ class TFXLNetPreTrainedModel(TFPreTrainedModel):
     """
     config_class = XLNetConfig
     pretrained_model_archive_map = TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_xlnet_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 
@@ -777,7 +771,7 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
     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 last layer of the model.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -819,7 +813,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **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).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -863,7 +857,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
 
         outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # return logits, mems, (hidden states), (attentions)
+        return outputs  # return logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
@@ -874,7 +868,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **logits**: ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -918,7 +912,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
 
         outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # return logits, mems, (hidden states), (attentions)
+        return outputs  # return logits, (mems), (hidden states), (attentions)
 
 
 # @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -932,6 +926,11 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
             Span-start scores (before SoftMax).
         **end_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
+            list of ``tf.Tensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **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)``:
@@ -971,7 +970,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
 
         outputs = (start_logits, end_logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # start_logits, end_logits, (hidden_states), (attentions)
+        return outputs  # start_logits, end_logits, (mems), (hidden_states), (attentions)
 
 # @add_start_docstrings("""XLNet 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`). """,

transformers/modeling_utils.py

@@ -316,20 +316,20 @@ 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)
-            except EnvironmentError as e:
+            except EnvironmentError:
                 if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
-                    logger.error(
-                        "Couldn't reach server at '{}' to download pretrained weights.".format(
-                            archive_file))
+                    msg = "Couldn't reach server at '{}' to download pretrained weights.".format(
+                            archive_file)
                 else:
-                    logger.error(
-                        "Model name '{}' was not found in model name list ({}). "
-                        "We assumed '{}' was a path or url but couldn't find any file "
-                        "associated to this path or url.".format(
+                    msg = "Model name '{}' was not found in model name list ({}). " \
+                        "We assumed '{}' was a path or url to model weight files named one of {} but " \
+                        "couldn't find any such file at this path or url.".format(
                             pretrained_model_name_or_path,
                             ', '.join(cls.pretrained_model_archive_map.keys()),
-                            archive_file))
-                raise e
+                            archive_file,
+                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME])
+                raise EnvironmentError(msg)
+
             if resolved_archive_file == archive_file:
                 logger.info("loading weights file {}".format(archive_file))
             else:

transformers/modeling_xlnet.py

@@ -188,11 +188,8 @@ def swish(x):
 ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish}
 
 
-try:
-    from apex.normalization.fused_layer_norm import FusedLayerNorm as XLNetLayerNorm
-except (ImportError, AttributeError) as e:
-    logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
-    from torch.nn import LayerNorm as XLNetLayerNorm
+XLNetLayerNorm = nn.LayerNorm
+
 
 class XLNetRelativeAttention(nn.Module):
     def __init__(self, config):
@@ -239,45 +236,60 @@ class XLNetRelativeAttention(nn.Module):
 
         return x
 
+    @staticmethod
+    def rel_shift_bnij(x, klen=-1):
+        x_size = x.shape
+
+        x = x.reshape(x_size[0], x_size[1], x_size[3], x_size[2])
+        x = x[:, :, 1:, :]
+        x = x.reshape(x_size[0], x_size[1], x_size[2], x_size[3]-1)
+        # Note: the tensor-slice form was faster in my testing than torch.index_select
+        #       However, tracing doesn't like the nature of the slice, and if klen changes
+        #       during the run then it'll fail, whereas index_select will be fine.
+        x = torch.index_select(x, 3, torch.arange(klen, device=x.device, dtype=torch.long))
+        # x = x[:, :, :, :klen]
+
+        return x
+
     def rel_attn_core(self, q_head, k_head_h, v_head_h, k_head_r, seg_mat=None, attn_mask=None, head_mask=None):
         """Core relative positional attention operations."""
 
         # content based attention score
-        ac = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_w_bias, k_head_h)
+        ac = torch.einsum('ibnd,jbnd->bnij', q_head + self.r_w_bias, k_head_h)
 
         # position based attention score
-        bd = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_r_bias, k_head_r)
-        bd = self.rel_shift(bd, klen=ac.shape[1])
+        bd = torch.einsum('ibnd,jbnd->bnij', q_head + self.r_r_bias, k_head_r)
+        bd = self.rel_shift_bnij(bd, klen=ac.shape[3])
 
         # segment based attention score
         if seg_mat is None:
             ef = 0
         else:
             ef = torch.einsum('ibnd,snd->ibns', q_head + self.r_s_bias, self.seg_embed)
-            ef = torch.einsum('ijbs,ibns->ijbn', seg_mat, ef)
+            ef = torch.einsum('ijbs,ibns->bnij', seg_mat, ef)
 
         # merge attention scores and perform masking
         attn_score = (ac + bd + ef) * self.scale
         if attn_mask is not None:
             # attn_score = attn_score * (1 - attn_mask) - 1e30 * attn_mask
             if attn_mask.dtype == torch.float16:
-                attn_score = attn_score - 65500 * attn_mask
+                attn_score = attn_score - 65500 * torch.einsum('ijbn->bnij', attn_mask)
             else:
-                attn_score = attn_score - 1e30 * attn_mask
+                attn_score = attn_score - 1e30 * torch.einsum('ijbn->bnij', attn_mask)
 
         # attention probability
-        attn_prob = F.softmax(attn_score, dim=1)
+        attn_prob = F.softmax(attn_score, dim=3)
         attn_prob = self.dropout(attn_prob)
 
         # Mask heads if we want to
         if head_mask is not None:
-            attn_prob = attn_prob * head_mask
+            attn_prob = attn_prob * torch.einsum('ijbn->bnij', head_mask)
 
         # attention output
-        attn_vec = torch.einsum('ijbn,jbnd->ibnd', attn_prob, v_head_h)
+        attn_vec = torch.einsum('bnij,jbnd->ibnd', attn_prob, v_head_h)
 
         if self.output_attentions:
-            return attn_vec, attn_prob
+            return attn_vec, torch.einsum('bnij->ijbn', attn_prob)
 
         return attn_vec
 
@@ -555,7 +567,7 @@ class XLNetModel(XLNetPreTrainedModel):
     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 last layer of the model.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -581,6 +593,7 @@ class XLNetModel(XLNetPreTrainedModel):
         super(XLNetModel, self).__init__(config)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.output_past = config.output_past
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -637,9 +650,6 @@ class XLNetModel(XLNetPreTrainedModel):
 
     def cache_mem(self, curr_out, prev_mem):
         """cache hidden states into memory."""
-        if self.mem_len is None or self.mem_len == 0:
-            return None
-        else:
         if self.reuse_len is not None and self.reuse_len > 0:
             curr_out = curr_out[:self.reuse_len]
 
@@ -817,6 +827,7 @@ class XLNetModel(XLNetPreTrainedModel):
         attentions = []
         hidden_states = []
         for i, layer_module in enumerate(self.layer):
+            if self.mem_len is not None and self.mem_len > 0 and self.output_past:
                 # cache new mems
                 new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
             if self.output_hidden_states:
@@ -836,7 +847,11 @@ class XLNetModel(XLNetPreTrainedModel):
         output = self.dropout(output_g if output_g is not None else output_h)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (output.permute(1, 0, 2).contiguous(), new_mems)
+        outputs = (output.permute(1, 0, 2).contiguous(),)
+
+        if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+            outputs = outputs + (new_mems,)
+
         if self.output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(h.permute(1, 0, 2).contiguous() for hs in hidden_states for h in hs)
@@ -847,7 +862,7 @@ class XLNetModel(XLNetPreTrainedModel):
             attentions = tuple(t.permute(2, 3, 0, 1).contiguous() for t in attentions)
             outputs = outputs + (attentions,)
 
-        return outputs  # outputs, new_mems, (hidden_states), (attentions)
+        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a language modeling head on top
@@ -867,7 +882,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
             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).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -932,7 +947,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
                             labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
@@ -951,7 +966,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
             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).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -1011,7 +1026,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 @add_start_docstrings("""XLNet Model with a multiple choice classification head on top (a linear layer on top of
     the pooled output and a softmax) e.g. for RACE/SWAG tasks. """,
@@ -1046,6 +1061,11 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
         **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above).
             Classification scores (before SoftMax).
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
+            list of ``torch.FloatTensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **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)``:
@@ -1102,7 +1122,7 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
             loss = loss_fct(reshaped_logits, labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -1126,7 +1146,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             Span-start scores (before SoftMax).
         **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -1197,7 +1217,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             total_loss = (start_loss + end_loss) / 2
             outputs = (total_loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        return outputs  # (loss), start_logits, end_logits, (mems), (hidden_states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -1239,7 +1259,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
         **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
             ``torch.FloatTensor`` of shape ``(batch_size,)``
             Log probabilities for the ``is_impossible`` label of the answers.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.

transformers/tests/modeling_common_test.py

@@ -17,8 +17,10 @@ from __future__ import division
 from __future__ import print_function
 
 import copy
+import sys
 import os
 import shutil
+import tempfile
 import json
 import random
 import uuid
@@ -31,6 +33,7 @@ from transformers import is_torch_available
 
 if is_torch_available():
     import torch
+    import numpy as np
 
     from transformers import (PretrainedConfig, PreTrainedModel,
                                     BertModel, BertConfig, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -38,6 +41,20 @@ if is_torch_available():
 else:
     pytestmark = pytest.mark.skip("Require Torch")
 
+if sys.version_info[0] == 2:
+    import cPickle as pickle
+
+    class TemporaryDirectory(object):
+        """Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
+        def __enter__(self):
+            self.name = tempfile.mkdtemp()
+            return self.name
+        def __exit__(self, exc_type, exc_value, traceback):
+            shutil.rmtree(self.name)
+else:
+    import pickle
+    TemporaryDirectory = tempfile.TemporaryDirectory
+    unicode = str
 
 def _config_zero_init(config):
     configs_no_init = copy.deepcopy(config)
@@ -57,6 +74,29 @@ class CommonTestCases:
         test_resize_embeddings = True
         test_head_masking = True
 
+        def test_save_load(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+                model.eval()
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
+
+                with TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname)
+                    model = model_class.from_pretrained(tmpdirname)
+                    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 = out_1[~np.isnan(out_1)]
+                    out_2 = out_2[~np.isnan(out_2)]
+                    max_diff = np.amax(np.abs(out_1 - out_2))
+                    self.assertLessEqual(max_diff, 1e-5)
+
         def test_initialization(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

transformers/tests/modeling_ctrl_test.py

+# coding=utf-8
+# Copyright 2018 Salesforce and HuggingFace Inc. team.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import pytest
+import shutil
+import pdb
+
+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
+
+
+class CTRLModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (CTRLModel, CTRLLMHeadModel) if is_torch_available() else ()
+    test_pruning = False
+    test_torchscript = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    class CTRLModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_token_type_ids=True,
+                     use_input_mask=True,
+                     use_labels=True,
+                     use_mc_token_ids=True,
+                     vocab_size=99,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                     ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_token_type_ids = use_token_type_ids
+            self.use_input_mask = use_input_mask
+            self.use_labels = use_labels
+            self.use_mc_token_ids = use_mc_token_ids
+            self.vocab_size = vocab_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            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.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            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)
+
+            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)
+
+            mc_token_ids = None
+            if self.use_mc_token_ids:
+                mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)
+
+            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)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = CTRLConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                n_embd=self.hidden_size,
+                n_layer=self.num_hidden_layers,
+                n_head=self.num_attention_heads,
+                # intermediate_size=self.intermediate_size,
+                # hidden_act=self.hidden_act,
+                # hidden_dropout_prob=self.hidden_dropout_prob,
+                # attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                n_positions=self.max_position_embeddings,
+                n_ctx=self.max_position_embeddings
+                # type_vocab_size=self.type_vocab_size,
+                # initializer_range=self.initializer_range
+            )
+
+            head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
+
+            return config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_ctrl_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+            model = CTRLModel(config=config)
+            model.eval()
+
+            model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
+            model(input_ids, token_type_ids=token_type_ids)
+            sequence_output, presents = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output,
+                "presents": presents,
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].size()),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertEqual(len(result["presents"]), config.n_layer)
+
+        def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+            model = CTRLLMHeadModel(config)
+            model.eval()
+
+            loss, lm_logits, _ = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
+
+            result = {
+                "loss": loss,
+                "lm_logits": lm_logits
+            }
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+            self.parent.assertListEqual(
+                list(result["lm_logits"].size()),
+                [self.batch_size, self.seq_length, self.vocab_size])
+
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+
+            (config, input_ids, input_mask, head_mask, token_type_ids,
+             mc_token_ids, sequence_labels, token_labels, choice_labels) = config_and_inputs
+
+            inputs_dict = {
+                'input_ids': input_ids,
+                'token_type_ids': token_type_ids,
+                'head_mask': head_mask
+            }
+
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = CTRLModelTest.CTRLModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CTRLConfig, n_embd=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_ctrl_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_ctrl_model(*config_and_inputs)
+
+    def test_ctrl_lm_head_model(self):
+        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
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in list(CTRL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = CTRLModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+
+if __name__ == "__main__":
+    unittest.main()

transformers/tests/modeling_roberta_test.py

@@ -24,7 +24,8 @@ from transformers import is_torch_available
 
 if is_torch_available():
     import torch
-    from transformers import (RobertaConfig, RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification)
+    from transformers import (RobertaConfig, RobertaModel, RobertaForMaskedLM,
+                              RobertaForSequenceClassification, RobertaForTokenClassification)
     from transformers.modeling_roberta import ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
 else:
     pytestmark = pytest.mark.skip("Require Torch")
@@ -156,6 +157,22 @@ class RobertaModelTest(CommonTestCases.CommonModelTester):
                 [self.batch_size, self.seq_length, self.vocab_size])
             self.check_loss_output(result)
 
+        def create_and_check_roberta_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 = RobertaForTokenClassification(config=config)
+            model.eval()
+            loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
+                                 labels=token_labels)
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.seq_length, self.num_labels])
+            self.check_loss_output(result)
+
         def prepare_config_and_inputs_for_common(self):
             config_and_inputs = self.prepare_config_and_inputs()
             (config, input_ids, token_type_ids, input_mask,

transformers/tests/modeling_tf_common_test.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 from __future__ import absolute_import, division, print_function
 
+import os
 import copy
 import json
 import logging
@@ -22,6 +23,7 @@ import random
 import shutil
 import unittest
 import uuid
+import tempfile
 
 import pytest
 import sys
@@ -36,6 +38,20 @@ if is_tf_available():
 else:
     pytestmark = pytest.mark.skip("Require TensorFlow")
 
+if sys.version_info[0] == 2:
+    import cPickle as pickle
+
+    class TemporaryDirectory(object):
+        """Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
+        def __enter__(self):
+            self.name = tempfile.mkdtemp()
+            return self.name
+        def __exit__(self, exc_type, exc_value, traceback):
+            shutil.rmtree(self.name)
+else:
+    import pickle
+    TemporaryDirectory = tempfile.TemporaryDirectory
+    unicode = str
 
 def _config_zero_init(config):
     configs_no_init = copy.deepcopy(config)
@@ -66,11 +82,31 @@ class TFCommonTestCases:
             #             self.assertIn(param.data.mean().item(), [0.0, 1.0],
             #             msg="Parameter {} of model {} seems not properly initialized".format(name, model_class))
 
+        def test_save_load(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+                outputs = model(inputs_dict)
+
+                with TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname)
+                    model = model_class.from_pretrained(tmpdirname)
+                    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 = out_1[~np.isnan(out_1)]
+                    out_2 = out_2[~np.isnan(out_2)]
+                    max_diff = np.amax(np.abs(out_1 - out_2))
+                    self.assertLessEqual(max_diff, 1e-5)
 
         def test_pt_tf_model_equivalence(self):
             if not is_torch_available():
                 return
 
+            import torch
             import transformers
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -79,12 +115,71 @@ class TFCommonTestCases:
                 pt_model_class_name = model_class.__name__[2:]  # Skip the "TF" at the beggining
                 pt_model_class = getattr(transformers, pt_model_class_name)
 
+                config.output_hidden_states = True
                 tf_model = model_class(config)
                 pt_model = pt_model_class(config)
 
+                # Check we can load pt model in tf and vice-versa with model => model functions
                 tf_model = transformers.load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=inputs_dict)
                 pt_model = transformers.load_tf2_model_in_pytorch_model(pt_model, tf_model)
 
+                # Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
+                pt_model.eval()
+                pt_inputs_dict = dict((name, torch.from_numpy(key.numpy()).to(torch.long))
+                                      for name, key in inputs_dict.items())
+                with torch.no_grad():
+                    pto = pt_model(**pt_inputs_dict)
+                tfo = tf_model(inputs_dict)
+                max_diff = np.amax(np.abs(tfo[0].numpy() - pto[0].numpy()))
+                self.assertLessEqual(max_diff, 2e-2)
+
+                # Check we can load pt model in tf and vice-versa with checkpoint => model functions
+                with TemporaryDirectory() as tmpdirname:
+                    pt_checkpoint_path = os.path.join(tmpdirname, 'pt_model.bin')
+                    torch.save(pt_model.state_dict(), pt_checkpoint_path)
+                    tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(tf_model, pt_checkpoint_path)
+
+                    tf_checkpoint_path = os.path.join(tmpdirname, 'tf_model.h5')
+                    tf_model.save_weights(tf_checkpoint_path)
+                    pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path)
+
+                # Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
+                pt_model.eval()
+                pt_inputs_dict = dict((name, torch.from_numpy(key.numpy()).to(torch.long))
+                                      for name, key in inputs_dict.items())
+                with torch.no_grad():
+                    pto = pt_model(**pt_inputs_dict)
+                tfo = tf_model(inputs_dict)
+                max_diff = np.amax(np.abs(tfo[0].numpy() - pto[0].numpy()))
+                self.assertLessEqual(max_diff, 2e-2)
+
+        def test_compile_tf_model(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            input_ids = tf.keras.Input(batch_shape=(2, 2000), name='input_ids', dtype='int32')
+            optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
+            loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+            metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
+
+            for model_class in self.all_model_classes:
+                # Prepare our model
+                model = model_class(config)
+                
+                # Let's load it from the disk to be sure we can use pretrained weights
+                with TemporaryDirectory() as tmpdirname:
+                    outputs = model(inputs_dict)  # build the model
+                    model.save_pretrained(tmpdirname)
+                    model = model_class.from_pretrained(tmpdirname)
+
+                outputs_dict = model(input_ids)
+                hidden_states = outputs_dict[0]
+
+                # Add a dense layer on top to test intetgration with other keras modules
+                outputs = tf.keras.layers.Dense(2, activation='softmax', name='outputs')(hidden_states)
+
+                # Compile extended model
+                extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
+                extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
 
         def test_keyword_and_dict_args(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

transformers/tests/modeling_tf_ctrl_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+import sys
+
+from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+
+from transformers import CTRLConfig, is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+    from transformers.modeling_tf_ctrl import (TFCTRLModel, TFCTRLLMHeadModel,
+                                                TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
+else:
+    pytestmark = pytest.mark.skip("Require TensorFlow")
+
+
+class TFCTRLModelTest(TFCommonTestCases.TFCommonModelTester):
+
+    all_model_classes = (TFCTRLModel, TFCTRLLMHeadModel) if is_tf_available() else ()
+
+    class TFCTRLModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_token_type_ids=True,
+                     use_input_mask=True,
+                     use_labels=True,
+                     use_mc_token_ids=True,
+                     vocab_size=99,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                     ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_token_type_ids = use_token_type_ids
+            self.use_input_mask = use_input_mask
+            self.use_labels = use_labels
+            self.use_mc_token_ids = use_mc_token_ids
+            self.vocab_size = vocab_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            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.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            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)
+
+            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)
+
+            mc_token_ids = None
+            if self.use_mc_token_ids:
+                mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)
+
+            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)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = CTRLConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                n_embd=self.hidden_size,
+                n_layer=self.num_hidden_layers,
+                n_head=self.num_attention_heads,
+                # intermediate_size=self.intermediate_size,
+                # hidden_act=self.hidden_act,
+                # hidden_dropout_prob=self.hidden_dropout_prob,
+                # attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                n_positions=self.max_position_embeddings,
+                n_ctx=self.max_position_embeddings
+                # type_vocab_size=self.type_vocab_size,
+                # initializer_range=self.initializer_range
+            )
+
+            head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
+
+            return config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, token_labels, choice_labels
+
+        def create_and_check_ctrl_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+            model = TFCTRLModel(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            sequence_output = model(inputs)[0]
+
+            inputs = [input_ids, None, input_mask]  # None is the input for 'past'
+            sequence_output = model(inputs)[0]
+
+            sequence_output = model(input_ids)[0]
+
+            result = {
+                "sequence_output": sequence_output.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].shape),
+                [self.batch_size, self.seq_length, self.hidden_size])
+
+
+        def create_and_check_ctrl_lm_head(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+            model = TFCTRLLMHeadModel(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            prediction_scores = model(inputs)[0]
+            result = {
+                "prediction_scores": prediction_scores.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].shape),
+                [self.batch_size, self.seq_length, self.vocab_size])
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+
+            (config, input_ids, input_mask, head_mask, token_type_ids,
+             mc_token_ids, sequence_labels, token_labels, choice_labels) = config_and_inputs
+
+            inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = TFCTRLModelTest.TFCTRLModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CTRLConfig, n_embd=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_ctrl_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_ctrl_model(*config_and_inputs)
+
+    def test_ctrl_lm_head(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_ctrl_lm_head(*config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in list(TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = TFCTRLModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()
+

transformers/tests/modeling_tf_gpt2_test.py

@@ -222,7 +222,7 @@ class TFGPT2ModelTest(TFCommonTestCases.TFCommonModelTester):
     @pytest.mark.slow
     def test_model_from_pretrained(self):
         cache_dir = "/tmp/transformers_test/"
-        for model_name in list(TF_gpt2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+        for model_name in list(TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
             model = TFGPT2Model.from_pretrained(model_name, cache_dir=cache_dir)
             shutil.rmtree(cache_dir)
             self.assertIsNotNone(model)

transformers/tests/modeling_tf_roberta_test.py

@@ -30,6 +30,7 @@ if is_tf_available():
     import numpy
     from transformers.modeling_tf_roberta import (TFRobertaModel, TFRobertaForMaskedLM,
                                                           TFRobertaForSequenceClassification,
+                                                          TFRobertaForTokenClassification,
                                                           TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
 else:
     pytestmark = pytest.mark.skip("Require TensorFlow")
@@ -154,6 +155,20 @@ class TFRobertaModelTest(TFCommonTestCases.TFCommonModelTester):
                 list(result["prediction_scores"].shape),
                 [self.batch_size, self.seq_length, self.vocab_size])
 
+        def create_and_check_roberta_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 = TFRobertaForTokenClassification(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            logits, = model(inputs)
+            result = {
+                "logits": logits.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].shape),
+                [self.batch_size, self.seq_length, self.num_labels])
+
         def prepare_config_and_inputs_for_common(self):
             config_and_inputs = self.prepare_config_and_inputs()
             (config, input_ids, token_type_ids, input_mask,

transformers/tests/modeling_tf_xlnet_test.py

@@ -161,6 +161,11 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                 "outputs": outputs.numpy(),
             }
 
+            config.mem_len = 0
+            model = TFXLNetModel(config)
+            no_mems_outputs = model(inputs)
+            self.parent.assertEqual(len(no_mems_outputs), 1)
+
             self.parent.assertListEqual(
                 list(result["outputs"].shape),
                 [self.batch_size, self.seq_length, self.hidden_size])

transformers/tests/modeling_xlnet_test.py

@@ -150,6 +150,12 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 "outputs": outputs,
             }
 
+            config.mem_len = 0
+            model = XLNetModel(config)
+            model.eval()
+            no_mems_outputs = model(input_ids_1)
+            self.parent.assertEqual(len(no_mems_outputs), 1)
+
             self.parent.assertListEqual(
                 list(result["outputs"].size()),
                 [self.batch_size, self.seq_length, self.hidden_size])

transformers/tests/tokenization_bert_test.py

@@ -131,8 +131,8 @@ class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
         text = tokenizer.encode("sequence builders")
         text_2 = tokenizer.encode("multi-sequence build")
 
-        encoded_sentence = tokenizer.add_special_tokens_single_sequence(text)
-        encoded_pair = tokenizer.add_special_tokens_sequence_pair(text, text_2)
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
 
         assert encoded_sentence == [101] + text + [102]
         assert encoded_pair == [101] + text + [102] + text_2 + [102]

transformers/tests/tokenization_ctrl_test.py

+# coding=utf-8
+# Copyright 2018 Salesforce and HuggingFace Inc. team.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+import unittest
+import json
+from io import open
+
+from transformers.tokenization_ctrl import CTRLTokenizer, VOCAB_FILES_NAMES
+
+from .tokenization_tests_commons import CommonTestCases
+
+class CTRLTokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = CTRLTokenizer
+
+    def setUp(self):
+        super(CTRLTokenizationTest, self).setUp()
+
+        # Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt
+        vocab = ['adapt', 're@@', 'a@@', 'apt', 'c@@', 't', '<unk>']
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+        merges = ["#version: 0.2", 'a p', 'ap t</w>', 'r e', 'a d', 'ad apt</w>', '']
+        self.special_tokens_map = {"unk_token": "<unk>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES['vocab_file'])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES['merges_file'])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(self.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+    def get_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return CTRLTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self):
+        input_text = u"adapt react readapt apt"
+        output_text = u"adapt react readapt apt"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = CTRLTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
+        text = "adapt react readapt apt"
+        bpe_tokens = 'adapt re@@ a@@ c@@ t re@@ adapt apt'.split()
+        tokens = tokenizer.tokenize(text)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens + [tokenizer.unk_token]
+
+        input_bpe_tokens = [0, 1, 2, 4, 5, 1, 0, 3, 6]
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+
+if __name__ == '__main__':
+    unittest.main()

transformers/tests/tokenization_distilbert_test.py

@@ -36,8 +36,8 @@ class DistilBertTokenizationTest(BertTokenizationTest):
         text = tokenizer.encode("sequence builders")
         text_2 = tokenizer.encode("multi-sequence build")
 
-        encoded_sentence = tokenizer.add_special_tokens_single_sequence(text)
-        encoded_pair = tokenizer.add_special_tokens_sequence_pair(text, text_2)
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
 
         assert encoded_sentence == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id]
         assert encoded_pair == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id] + \

transformers/tests/tokenization_roberta_test.py

@@ -87,8 +87,8 @@ class RobertaTokenizationTest(CommonTestCases.CommonTokenizerTester):
         encoded_text_from_decode = tokenizer.encode("sequence builders", add_special_tokens=True)
         encoded_pair_from_decode = tokenizer.encode("sequence builders", "multi-sequence build", add_special_tokens=True)
 
-        encoded_sentence = tokenizer.add_special_tokens_single_sequence(text)
-        encoded_pair = tokenizer.add_special_tokens_sequence_pair(text, text_2)
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
 
         assert encoded_sentence == encoded_text_from_decode
         assert encoded_pair == encoded_pair_from_decode

transformers/tests/tokenization_tests_commons.py

@@ -193,12 +193,12 @@ class CommonTestCases:
 
             tokenizer = self.get_tokenizer()
 
-            if tokenizer.add_special_tokens_sequence_pair.__qualname__.split('.')[0] != "PreTrainedTokenizer":
+            if tokenizer.build_inputs_with_special_tokens.__qualname__.split('.')[0] != "PreTrainedTokenizer":
                 seq_0 = "Test this method."
                 seq_1 = "With these inputs."
                 information = tokenizer.encode_plus(seq_0, seq_1, add_special_tokens=True)
                 sequences, mask = information["input_ids"], information["token_type_ids"]
-                assert len(sequences) == len(mask)
+                self.assertEqual(len(sequences), len(mask))
 
         def test_number_of_added_tokens(self):
             tokenizer = self.get_tokenizer()
@@ -211,7 +211,7 @@ class CommonTestCases:
 
             # Method is implemented (e.g. not GPT-2)
             if len(attached_sequences) != 2:
-                assert tokenizer.num_added_tokens(pair=True) == len(attached_sequences) - len(sequences)
+                self.assertEqual(tokenizer.num_added_tokens(pair=True), len(attached_sequences) - len(sequences))
 
         def test_maximum_encoding_length_single_input(self):
             tokenizer = self.get_tokenizer()
@@ -227,10 +227,10 @@ class CommonTestCases:
             truncated_sequence = information["input_ids"]
             overflowing_tokens = information["overflowing_tokens"]
 
-            assert len(overflowing_tokens) == 2 + stride
-            assert overflowing_tokens == sequence[-(2 + stride):]
-            assert len(truncated_sequence) == total_length - 2
-            assert truncated_sequence == tokenizer.add_special_tokens_single_sequence(sequence[:-2])
+            self.assertEqual(len(overflowing_tokens), 2 + stride)
+            self.assertEqual(overflowing_tokens, sequence[-(2 + stride):])
+            self.assertEqual(len(truncated_sequence), total_length - 2)
+            self.assertEqual(truncated_sequence, tokenizer.build_inputs_with_special_tokens(sequence[:-2]))
 
         def test_maximum_encoding_length_pair_input(self):
             tokenizer = self.get_tokenizer()
@@ -243,26 +243,26 @@ class CommonTestCases:
             sequence_1_no_special_tokens = tokenizer.encode(seq_1)
 
             sequence = tokenizer.encode(seq_0, seq_1, add_special_tokens=True)
-            truncated_second_sequence = tokenizer.add_special_tokens_sequence_pair(
+            truncated_second_sequence = tokenizer.build_inputs_with_special_tokens(
                 tokenizer.encode(seq_0),
                 tokenizer.encode(seq_1)[:-2]
             )
 
             information = tokenizer.encode_plus(seq_0, seq_1, max_length=len(sequence) - 2, add_special_tokens=True,
-                                                stride=stride, truncate_first_sequence=False)
+                                                stride=stride, truncation_strategy='only_second')
             information_first_truncated = tokenizer.encode_plus(seq_0, seq_1, max_length=len(sequence) - 2,
                                                                 add_special_tokens=True, stride=stride,
-                                                                truncate_first_sequence=True)
+                                                                truncation_strategy='only_first')
 
             truncated_sequence = information["input_ids"]
             overflowing_tokens = information["overflowing_tokens"]
             overflowing_tokens_first_truncated = information_first_truncated["overflowing_tokens"]
 
-            assert len(overflowing_tokens) == 2 + stride
-            assert overflowing_tokens == sequence_1_no_special_tokens[-(2 + stride):]
-            assert overflowing_tokens_first_truncated == sequence_0_no_special_tokens[-(2 + stride):]
-            assert len(truncated_sequence) == len(sequence) - 2
-            assert truncated_sequence == truncated_second_sequence
+            self.assertEqual(len(overflowing_tokens), 2 + stride)
+            self.assertEqual(overflowing_tokens, sequence_1_no_special_tokens[-(2 + stride):])
+            self.assertEqual(overflowing_tokens_first_truncated, sequence_0_no_special_tokens[-(2 + stride):])
+            self.assertEqual(len(truncated_sequence), len(sequence) - 2)
+            self.assertEqual(truncated_sequence, truncated_second_sequence)
 
         def test_encode_input_type(self):
             tokenizer = self.get_tokenizer()
@@ -273,5 +273,43 @@ class CommonTestCases:
             input_ids = tokenizer.convert_tokens_to_ids(tokens)
             formatted_input = tokenizer.encode(sequence, add_special_tokens=True)
 
-            assert tokenizer.encode(tokens, add_special_tokens=True) == formatted_input
-            assert tokenizer.encode(input_ids, add_special_tokens=True) == formatted_input
+            self.assertEqual(tokenizer.encode(tokens, add_special_tokens=True), formatted_input)
+            self.assertEqual(tokenizer.encode(input_ids, add_special_tokens=True), formatted_input)
+
+        def test_special_tokens_mask(self):
+            tokenizer = self.get_tokenizer()
+
+            sequence_0 = "Encode this."
+            sequence_1 = "This one too please."
+
+            # Testing single inputs
+            encoded_sequence = tokenizer.encode(sequence_0)
+            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, add_special_tokens=True)
+            encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
+            special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
+            self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
+
+            filtered_sequence = [(x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)]
+            filtered_sequence = [x for x in filtered_sequence if x is not None]
+            self.assertEqual(encoded_sequence, filtered_sequence)
+
+            # Testing inputs pairs
+            encoded_sequence = tokenizer.encode(sequence_0) + tokenizer.encode(sequence_1)
+            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, sequence_1, add_special_tokens=True)
+            encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
+            special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
+            self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
+
+            filtered_sequence = [(x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)]
+            filtered_sequence = [x for x in filtered_sequence if x is not None]
+            self.assertEqual(encoded_sequence, filtered_sequence)
+
+            # Testing with already existing special tokens
+            if tokenizer.cls_token_id == tokenizer.unk_token_id and tokenizer.cls_token_id == tokenizer.unk_token_id:
+                tokenizer.add_special_tokens({'cls_token': '</s>', 'sep_token': '<s>'})
+            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, add_special_tokens=True)
+            encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
+            special_tokens_mask_orig = encoded_sequence_dict["special_tokens_mask"]
+            special_tokens_mask = tokenizer.get_special_tokens_mask(encoded_sequence_w_special, already_has_special_tokens=True)
+            self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
+            self.assertEqual(special_tokens_mask_orig, special_tokens_mask)

transformers/tests/tokenization_xlm_test.py

@@ -72,8 +72,8 @@ class XLMTokenizationTest(CommonTestCases.CommonTokenizerTester):
         text = tokenizer.encode("sequence builders")
         text_2 = tokenizer.encode("multi-sequence build")
 
-        encoded_sentence = tokenizer.add_special_tokens_single_sequence(text)
-        encoded_pair = tokenizer.add_special_tokens_sequence_pair(text, text_2)
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
 
         assert encoded_sentence == [1] + text + [1]
         assert encoded_pair == [1] + text + [1] + text_2 + [1]

transformers/tests/tokenization_xlnet_test.py

@@ -95,8 +95,8 @@ class XLNetTokenizationTest(CommonTestCases.CommonTokenizerTester):
         text = tokenizer.encode("sequence builders")
         text_2 = tokenizer.encode("multi-sequence build")
 
-        encoded_sentence = tokenizer.add_special_tokens_single_sequence(text)
-        encoded_pair = tokenizer.add_special_tokens_sequence_pair(text, text_2)
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
 
         assert encoded_sentence == text + [4, 3]
         assert encoded_pair == text + [4] + text_2 + [4, 3]

transformers/tokenization_auto.py

@@ -21,6 +21,7 @@ import logging
 from .tokenization_bert import BertTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer
+from .tokenization_ctrl import CTRLTokenizer
 from .tokenization_transfo_xl import TransfoXLTokenizer
 from .tokenization_xlnet import XLNetTokenizer
 from .tokenization_xlm import XLMTokenizer
@@ -45,6 +46,7 @@ class AutoTokenizer(object):
             - contains `bert`: BertTokenizer (Bert model)
             - contains `openai-gpt`: OpenAIGPTTokenizer (OpenAI GPT model)
             - contains `gpt2`: GPT2Tokenizer (OpenAI GPT-2 model)
+            - contains `ctrl`: CTRLTokenizer (Salesforce CTRL model)
             - contains `transfo-xl`: TransfoXLTokenizer (Transformer-XL model)
             - contains `xlnet`: XLNetTokenizer (XLNet model)
             - contains `xlm`: XLMTokenizer (XLM model)
@@ -67,6 +69,7 @@ class AutoTokenizer(object):
             - contains `bert`: BertTokenizer (Bert model)
             - contains `openai-gpt`: OpenAIGPTTokenizer (OpenAI GPT model)
             - contains `gpt2`: GPT2Tokenizer (OpenAI GPT-2 model)
+            - contains `ctrl`: CTRLTokenizer (Salesforce CTRL model)
             - contains `transfo-xl`: TransfoXLTokenizer (Transformer-XL model)
             - contains `xlnet`: XLNetTokenizer (XLNet model)
             - contains `xlm`: XLMTokenizer (XLM model)
@@ -114,7 +117,8 @@ class AutoTokenizer(object):
             return XLNetTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'xlm' in pretrained_model_name_or_path:
             return XLMTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
-
+        elif 'ctrl' in pretrained_model_name_or_path:
+            return CTRLTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))