updating tests

pytorch_transformers/modeling_bert.py

@@ -617,6 +617,7 @@ class BertModel(BertPreTrainedModel):
         old_embeddings = self.embeddings.word_embeddings
         new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
         self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
 
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
@@ -758,11 +759,8 @@ class BertForPreTraining(BertPreTrainedModel):
         """ Make sure we are sharing the input and output embeddings.
             Export to TorchScript can't handle parameter sharing so we are cloning them instead.
         """
-        input_embeddings = self.bert.embeddings.word_embeddings.weight
+        self._tie_or_clone_weights(self.cls.predictions.decoder,
-        if self.config.torchscript:
+                                   self.bert.embeddings.word_embeddings)
-            self.cls.predictions.decoder.weight = nn.Parameter(input_embeddings.clone())
-        else:
-            self.cls.predictions.decoder.weight = input_embeddings  # Tied weights
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, masked_lm_labels=None,
                 next_sentence_label=None, head_mask=None):
@@ -864,11 +862,8 @@ class BertForMaskedLM(BertPreTrainedModel):
         """ Make sure we are sharing the input and output embeddings.
             Export to TorchScript can't handle parameter sharing so we are cloning them instead.
         """
-        input_embeddings = self.bert.embeddings.word_embeddings.weight
+        self._tie_or_clone_weights(self.cls.predictions.decoder,
-        if self.config.torchscript:
+                                   self.bert.embeddings.word_embeddings)
-            self.cls.predictions.decoder.weight = nn.Parameter(input_embeddings.clone())
-        else:
-            self.cls.predictions.decoder.weight = input_embeddings  # Tied weights
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, masked_lm_labels=None, head_mask=None):
         """

pytorch_transformers/modeling_gpt2.py

@@ -414,6 +414,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.wte = self._get_resized_embeddings(self.wte, new_num_tokens)
+        return self.wte
 
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
@@ -562,11 +563,8 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         """ Make sure we are sharing the input and output embeddings.
             Export to TorchScript can't handle parameter sharing so we are cloning them instead.
         """
-        input_embeddings = self.transformer.wte.weight
+        self._tie_or_clone_weights(self.lm_head,
-        if self.config.torchscript:
+                                   self.transformer.wte)
-            self.lm_head.weight = nn.Parameter(input_embeddings.clone())
-        else:
-            self.lm_head.weight = input_embeddings  # Tied weights
 
     def forward(self, input_ids, position_ids=None, token_type_ids=None, lm_labels=None, past=None, head_mask=None):
         """
@@ -658,11 +656,8 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         """ Make sure we are sharing the input and output embeddings.
             Export to TorchScript can't handle parameter sharing so we are cloning them instead.
         """
-        input_embeddings = self.transformer.wte.weight
+        self._tie_or_clone_weights(self.lm_head,
-        if self.config.torchscript:
+                                   self.transformer.wte)
-            self.lm_head.weight = nn.Parameter(input_embeddings.clone())
-        else:
-            self.lm_head.weight = input_embeddings  # Tied weights
 
     def forward(self, input_ids, mc_token_ids=None, lm_labels=None, mc_labels=None, token_type_ids=None,
                 position_ids=None, past=None, head_mask=None):

pytorch_transformers/modeling_openai.py

@@ -430,6 +430,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.tokens_embed = self._get_resized_embeddings(self.tokens_embed, new_num_tokens)
+        return self.tokens_embed
 
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
@@ -583,11 +584,8 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
         """ Make sure we are sharing the input and output embeddings.
             Export to TorchScript can't handle parameter sharing so we are cloning them instead.
         """
-        input_embeddings = self.transformer.tokens_embed.weight
+        self._tie_or_clone_weights(self.lm_head,
-        if self.config.torchscript:
+                                   self.transformer.tokens_embed)
-            self.lm_head.weight = nn.Parameter(input_embeddings.clone())
-        else:
-            self.lm_head.weight = input_embeddings  # Tied weights
 
     def forward(self, input_ids, position_ids=None, token_type_ids=None, lm_labels=None, head_mask=None):
         """
@@ -696,11 +694,8 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
         """ Make sure we are sharing the input and output embeddings.
             Export to TorchScript can't handle parameter sharing so we are cloning them instead.
         """
-        input_embeddings = self.transformer.tokens_embed.weight
+        self._tie_or_clone_weights(self.lm_head,
-        if self.config.torchscript:
+                                   self.transformer.tokens_embed)
-            self.lm_head.weight = nn.Parameter(input_embeddings.clone())
-        else:
-            self.lm_head.weight = input_embeddings  # Tied weights
 
     def forward(self, input_ids, mc_token_ids=None, lm_labels=None, mc_labels=None, token_type_ids=None,
                 position_ids=None, head_mask=None):

pytorch_transformers/modeling_transfo_xl.py

@@ -291,6 +291,10 @@ class TransfoXLConfig(PretrainedConfig):
     def vocab_size(self):
         return self.n_token
 
+    @vocab_size.setter
+    def vocab_size(self, value):
+        self.n_token = value
+
     @property
     def hidden_size(self):
         return self.d_model
@@ -1003,7 +1007,7 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
         self.apply(self.init_weights)
 
     def _resize_token_embeddings(self, new_num_tokens):
-        raise NotImplementedError
+        return self.word_emb
 
     def backward_compatible(self):
         self.sample_softmax = -1
@@ -1280,13 +1284,20 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
         else:
             if self.config.tie_weight:
                 for i in range(len(self.crit.out_layers)):
-                    self.crit.out_layers[i].weight = self.transformer.word_emb.emb_layers[i].weight
+                    self._tie_or_clone_weights(self.crit.out_layers[i],
+                                               self.transformer.word_emb.emb_layers[i])
             if self.config.tie_projs:
                 for i, tie_proj in enumerate(self.config.tie_projs):
                     if tie_proj and self.config.div_val == 1 and self.config.d_model != self.config.d_embed:
-                        self.crit.out_projs[i] = self.transformer.word_emb.emb_projs[0]
+                        if self.config.torchscript:
+                            self.crit.out_projs[i] = nn.Parameter(self.transformer.word_emb.emb_projs[0].clone())
+                        else:
+                            self.crit.out_projs[i] = self.transformer.word_emb.emb_projs[0]
                     elif tie_proj and self.config.div_val != 1:
-                        self.crit.out_projs[i] = self.transformer.word_emb.emb_projs[i]
+                        if self.config.torchscript:
+                            self.crit.out_projs[i] = nn.Parameter(self.transformer.word_emb.emb_projs[i].clone())
+                        else:
+                            self.crit.out_projs[i] = self.transformer.word_emb.emb_projs[i]
 
     def reset_length(self, tgt_len, ext_len, mem_len):
         self.transformer.reset_length(tgt_len, ext_len, mem_len)

pytorch_transformers/modeling_utils.py

@@ -165,9 +165,27 @@ class PreTrainedModel(nn.Module):
         # Save config in model
         self.config = config
 
-    def _get_resized_embeddings(self, old_embeddings, new_num_tokens):
+    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
-        # Build new embeddings
+        """ Build a resized Embedding Module from a provided token Embedding Module.
+            Increasing the size will add newly initialized vectors at the end
+            Reducing the size will remove vectors from the end
+
+        Args:
+            new_num_tokens: (Optional) New number of tokens in the embedding matrix.
+                Increasing the size will add newly initialized vectors at the end
+                Reducing the size will remove vectors from the end
+                If not provided or None: return the provided token Embedding Module.
+        Return:
+            Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None
+        """
+        if new_num_tokens is None:
+            return old_embeddings
+
         old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
+        if old_num_tokens == new_num_tokens:
+            return old_embeddings
+
+        # Build new embeddings
         new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
         new_embeddings.to(old_embeddings.weight.device)
 
@@ -180,18 +198,29 @@ class PreTrainedModel(nn.Module):
 
         return new_embeddings
 
-    def resize_token_embeddings(self, new_num_tokens):
+    def _tie_or_clone_weights(self, first_module, second_module):
-        """ Resize input token embeddings matrix.
+        """ Tie or clone module weights depending of weither we are using TorchScript or not
+        """
+        if self.config.torchscript:
+            first_module.weight = nn.Parameter(second_module.weight.clone())
+        else:
+            first_module.weight = second_module.weight
+
+    def resize_token_embeddings(self, new_num_tokens=None):
+        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
 
         Args:
-            new_num_tokens: New number of tokens in the embedding matrix.
+            new_num_tokens: (Optional) New number of tokens in the embedding matrix.
                 Increasing the size will add newly initialized vectors at the end
                 Reducing the size will remove vectors from the end
+                If not provided or None: does nothing.
+        Return:
+            Pointer to the input tokens Embedding Module of the model
         """
-        if new_num_tokens == self.config.vocab_size:
-            return
         base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
-        base_model._resize_token_embeddings(new_num_tokens)
+        model_embeds = base_model._resize_token_embeddings(new_num_tokens)
+        if new_num_tokens is None:
+            return model_embeds
 
         # Update base model and current model config
         self.config.vocab_size = new_num_tokens
@@ -201,6 +230,8 @@ class PreTrainedModel(nn.Module):
         if hasattr(self, 'tie_weights'):
             self.tie_weights()
 
+        return model_embeds
+
     def prune_heads(self, heads_to_prune):
         """ Prunes heads of the base model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}

pytorch_transformers/modeling_xlm.py

@@ -184,6 +184,10 @@ class XLMConfig(PretrainedConfig):
     def vocab_size(self):
         return self.n_words
 
+    @vocab_size.setter
+    def vocab_size(self, value):
+        self.n_words = value
+
     @property
     def hidden_size(self):
         return self.emb_dim
@@ -479,6 +483,7 @@ class XLMModel(XLMPreTrainedModel):
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.embeddings = self._get_resized_embeddings(self.embeddings, new_num_tokens)
+        return self.embeddings
 
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
@@ -728,10 +733,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
     def tie_weights(self):
         """ Make sure we are sharing the embeddings
         """
-        if self.config.torchscript:
+        self._tie_or_clone_weights(self.pred_layer.proj, self.transformer.embeddings)
-            self.pred_layer.proj.weight = nn.Parameter(self.transformer.embeddings.weight.clone())
-        else:
-            self.pred_layer.proj.weight = self.transformer.embeddings.weight
 
     def forward(self, input_ids, lengths=None, positions=None, langs=None, token_type_ids=None,
                 attention_mask=None, cache=None, labels=None, head_mask=None):

pytorch_transformers/modeling_xlnet.py

@@ -316,6 +316,10 @@ class XLNetConfig(PretrainedConfig):
     def vocab_size(self):
         return self.n_token
 
+    @vocab_size.setter
+    def vocab_size(self, value):
+        self.n_token = value
+
     @property
     def hidden_size(self):
         return self.d_model
@@ -660,10 +664,10 @@ class XLNetModel(XLNetPreTrainedModel):
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.word_embedding = self._get_resized_embeddings(self.word_embedding, new_num_tokens)
+        return self.word_embedding
 
     def _prune_heads(self, heads_to_prune):
-        logger.info("Head pruning is not implemented for XLNet")
+        raise NotImplementedError
-        pass
 
     def create_mask(self, qlen, mlen):
         """
@@ -987,10 +991,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
     def tie_weights(self):
         """ Make sure we are sharing the embeddings
         """
-        if self.config.torchscript:
+        self._tie_or_clone_weights(self.lm_loss, self.transformer.word_embedding)
-            self.lm_loss.weight = nn.Parameter(self.transformer.word_embedding.weight.clone())
-        else:
-            self.lm_loss.weight = self.transformer.word_embedding.weight
 
     def forward(self, input_ids, token_type_ids=None, input_mask=None, attention_mask=None,
                 mems=None, perm_mask=None, target_mapping=None, inp_q=None,

pytorch_transformers/tests/modeling_bert_test.py

@@ -26,10 +26,15 @@ from pytorch_transformers import (BertConfig, BertModel, BertForMaskedLM,
                                      BertForTokenClassification, BertForMultipleChoice)
 from pytorch_transformers.modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
-from .modeling_common_test import (create_and_check_commons, ConfigTester, ids_tensor)
+from .modeling_common_test import (CommonTestCases, ConfigTester, ids_tensor)
 
 
-class BertModelTest(unittest.TestCase):
+class BertModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (BertModel, BertForMaskedLM, BertForNextSentencePrediction,
+            BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification,
+            BertForTokenClassification)
+
     class BertModelTester(object):
 
         def __init__(self,
@@ -55,9 +60,6 @@ class BertModelTest(unittest.TestCase):
                      num_labels=3,
                      num_choices=4,
                      scope=None,
-                     all_model_classes = (BertModel, BertForMaskedLM, BertForNextSentencePrediction,
-                             BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification,
-                             BertForTokenClassification),
                     ):
             self.parent = parent
             self.batch_size = batch_size
@@ -81,7 +83,6 @@ class BertModelTest(unittest.TestCase):
             self.num_labels = num_labels
             self.num_choices = num_choices
             self.scope = scope
-            self.all_model_classes = all_model_classes
 
         def prepare_config_and_inputs(self):
             input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -253,16 +254,51 @@ class BertModelTest(unittest.TestCase):
             self.check_loss_output(result)
 
 
-        def create_and_check_bert_commons(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_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,
+             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}
-            create_and_check_commons(self, config, inputs_dict)
+            return config, inputs_dict
 
-    def test_default(self):
+    def setUp(self):
-        self.run_tester(BertModelTest.BertModelTester(self))
+        self.model_tester = BertModelTest.BertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=BertConfig, hidden_size=37)
 
     def test_config(self):
-        config_tester = ConfigTester(self, config_class=BertConfig, hidden_size=37)
+        self.config_tester.run_common_tests()
-        config_tester.run_common_tests()
+
+    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)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_bert_for_masked_lm(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_bert_for_multiple_choice(*config_and_inputs)
+
+    def test_for_next_sequence_prediction(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_bert_for_next_sequence_prediction(*config_and_inputs)
+
+    def test_for_pretraining(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_bert_for_pretraining(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_bert_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_bert_for_sequence_classification(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        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
     def test_model_from_pretrained(self):
@@ -272,33 +308,5 @@ class BertModelTest(unittest.TestCase):
             shutil.rmtree(cache_dir)
             self.assertIsNotNone(model)
 
-    def run_tester(self, tester):
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_model(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_masked_lm(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_multiple_choice(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_next_sequence_prediction(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_pretraining(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_question_answering(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_sequence_classification(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_for_token_classification(*config_and_inputs)
-
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_bert_commons(*config_and_inputs)
-
 if __name__ == "__main__":
     unittest.main()

pytorch_transformers/tests/modeling_common_test.py

@@ -39,207 +39,471 @@ def _config_zero_init(config):
             setattr(configs_no_init, key, 0.0)
     return configs_no_init
 
-def _create_and_check_torchscript_output_attentions(tester, model_classes, config, inputs_dict):
+class CommonTestCases:
-    config.output_attentions = True
-    _create_and_check_torchscript(tester, model_classes, config, inputs_dict)
 
-def _create_and_check_torchscript_output_hidden_state(tester, model_classes, config, inputs_dict):
+    class CommonModelTester(unittest.TestCase):
-    config.output_hidden_states = True
-    _create_and_check_torchscript(tester, model_classes, config, inputs_dict)
 
-def _create_and_check_torchscript(tester, model_classes, config, inputs_dict):
+        model_tester = None
-    configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        all_model_classes = ()
-    configs_no_init.torchscript = True
+        test_torchscript = True
-    for model_class in model_classes:
+        test_pruning = True
-        model = model_class(config=configs_no_init)
+        test_resize_embeddings = True
-        model.eval()
-        inputs = inputs_dict['input_ids']  # Let's keep only input_ids
 
-        try:
+        def test_initialization(self):
-            torch.jit.trace(model, inputs)
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        except RuntimeError:
-            tester.parent.fail("Couldn't trace module.")
 
-        try:
+            configs_no_init = _config_zero_init(config)
-            traced_gpt2 = torch.jit.trace(model, inputs)
+            for model_class in self.all_model_classes:
-            torch.jit.save(traced_gpt2, "traced_model.pt")
+                model = model_class(config=configs_no_init)
-        except RuntimeError:
+                for name, param in model.named_parameters():
-            tester.parent.fail("Couldn't save module.")
+                    if param.requires_grad:
+                        self.assertIn(param.data.mean().item(), [0.0, 1.0],
+                        msg="Parameter {} of model {} seems not properly initialized".format(name, model_class))
 
-        try:
+        def test_attention_outputs(self):
-            loaded_model = torch.jit.load("traced_model.pt")
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-            os.remove("traced_model.pt")
-        except ValueError:
-            tester.parent.fail("Couldn't load module.")
 
-        model.eval()
+            for model_class in self.all_model_classes:
-        loaded_model.eval()
+                config.output_attentions = True
+                config.output_hidden_states = False
+                model = model_class(config)
+                model.eval()
+                outputs = model(**inputs_dict)
+                attentions = outputs[-1]
+                self.assertEqual(model.config.output_attentions, True)
+                self.assertEqual(model.config.output_hidden_states, False)
+                self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+                self.assertListEqual(
+                    list(attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads,
+                    self.model_tester.seq_length,
+                    self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
+                out_len = len(outputs)
 
-        model_params = model.parameters()
+                # Check attention is always last and order is fine
-        loaded_model_params = loaded_model.parameters()
+                config.output_attentions = True
+                config.output_hidden_states = True
+                model = model_class(config)
+                model.eval()
+                outputs = model(**inputs_dict)
+                self.assertEqual(out_len+1, len(outputs))
+                self.assertEqual(model.config.output_attentions, True)
+                self.assertEqual(model.config.output_hidden_states, True)
 
-        models_equal = True
+                attentions = outputs[-1]
-        for p1, p2 in zip(model_params, loaded_model_params):
+                self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
-            if p1.data.ne(p2.data).sum() > 0:
+                self.assertListEqual(
-                models_equal = False
+                    list(attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads,
+                    self.model_tester.seq_length,
+                    self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
 
-        tester.parent.assertTrue(models_equal)
+        def test_torchscript(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
-def _create_and_check_initialization(tester, model_classes, config, inputs_dict):
+            self._create_and_check_torchscript(config, inputs_dict)
-    configs_no_init = _config_zero_init(config)
-    for model_class in model_classes:
-        model = model_class(config=configs_no_init)
-        for name, param in model.named_parameters():
-            if param.requires_grad:
-                tester.parent.assertIn(param.data.mean().item(), [0.0, 1.0],
-                                       msg="Parameter {} of model {} seems not properly initialized".format(name, model_class))
 
-def _create_and_check_for_headmasking(tester, model_classes, config, inputs_dict):
+        def test_torchscript_output_attentions(self):
-    configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-    for model_class in model_classes:
-        config.output_attentions = True
-        config.output_hidden_states = True
-        model = model_class(config=configs_no_init)
-        model.eval()
 
-        # Prepare head_mask
+            config.output_attentions = True
-        # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior) 
+            self._create_and_check_torchscript(config, inputs_dict)
-        head_mask = torch.ones(tester.num_hidden_layers, tester.num_attention_heads)
-        head_mask[0, 0] = 0
-        head_mask[-1, :-1] = 0
-        head_mask.requires_grad_(requires_grad=True)
-        inputs = inputs_dict.copy()
-        inputs['head_mask'] = head_mask
 
-        outputs = model(**inputs)
+        def test_torchscript_output_hidden_state(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
-        # Test that we can get a gradient back for importance score computation
+            config.output_hidden_states = True
-        output = sum(t.sum() for t in outputs[0])
+            self._create_and_check_torchscript(config, inputs_dict)
-        output = output.sum()
-        output.backward()
-        multihead_outputs = head_mask.grad
 
-        attentions = outputs[-1]
+        def _create_and_check_torchscript(self, config, inputs_dict):
-        hidden_states = outputs[-2]
+            if not self.test_torchscript:
+                return
 
-        # Remove Nan
+            configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+            configs_no_init.torchscript = True
+            for model_class in self.all_model_classes:
+                model = model_class(config=configs_no_init)
+                model.eval()
+                inputs = inputs_dict['input_ids']  # Let's keep only input_ids
 
-        tester.parent.assertIsNotNone(multihead_outputs)
+                try:
-        tester.parent.assertEqual(len(multihead_outputs), tester.num_hidden_layers)
+                    torch.jit.trace(model, inputs)
-        tester.parent.assertAlmostEqual(
+                except RuntimeError:
-            attentions[0][..., 0, :, :].flatten().sum().item(), 0.0)
+                    self.fail("Couldn't trace module.")
-        tester.parent.assertNotEqual(
+
-            attentions[0][..., -1, :, :].flatten().sum().item(), 0.0)
+                try:
-        tester.parent.assertNotEqual(
+                    traced_gpt2 = torch.jit.trace(model, inputs)
-            attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
+                    torch.jit.save(traced_gpt2, "traced_model.pt")
-        tester.parent.assertAlmostEqual(
+                except RuntimeError:
-            attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
+                    self.fail("Couldn't save module.")
-        tester.parent.assertNotEqual(
+
-            attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)
+                try:
+                    loaded_model = torch.jit.load("traced_model.pt")
+                    os.remove("traced_model.pt")
+                except ValueError:
+                    self.fail("Couldn't load module.")
+
+                model.eval()
+                loaded_model.eval()
+
+                model_params = model.parameters()
+                loaded_model_params = loaded_model.parameters()
+
+                models_equal = True
+                for p1, p2 in zip(model_params, loaded_model_params):
+                    if p1.data.ne(p2.data).sum() > 0:
+                        models_equal = False
+
+                self.assertTrue(models_equal)
 
 
-def _create_and_check_for_head_pruning(tester, model_classes, config, inputs_dict):
+        def test_headmasking(self):
-    for model_class in model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        config.output_attentions = True
-        config.output_hidden_states = False
-        model = model_class(config=config)
-        model.eval()
-        heads_to_prune = {0: list(range(1, tester.num_attention_heads)),
-                          -1: [0]}
-        model.prune_heads(heads_to_prune)
-        outputs = model(**inputs_dict)
 
-        attentions = outputs[-1]
+            config.output_attentions = True
+            config.output_hidden_states = True
+            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.eval()
 
-        tester.parent.assertEqual(
+                # Prepare head_mask
-            attentions[0].shape[-3], 1)
+                # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior) 
-        tester.parent.assertEqual(
+                head_mask = torch.ones(self.model_tester.num_hidden_layers, self.model_tester.num_attention_heads)
-            attentions[1].shape[-3], tester.num_attention_heads)
+                head_mask[0, 0] = 0
-        tester.parent.assertEqual(
+                head_mask[-1, :-1] = 0
-            attentions[-1].shape[-3], tester.num_attention_heads - 1)
+                head_mask.requires_grad_(requires_grad=True)
+                inputs = inputs_dict.copy()
+                inputs['head_mask'] = head_mask
+
+                outputs = model(**inputs)
+
+                # Test that we can get a gradient back for importance score computation
+                output = sum(t.sum() for t in outputs[0])
+                output = output.sum()
+                output.backward()
+                multihead_outputs = head_mask.grad
+
+                attentions = outputs[-1]
+                hidden_states = outputs[-2]
+
+                # Remove Nan
+
+                self.assertIsNotNone(multihead_outputs)
+                self.assertEqual(len(multihead_outputs), self.model_tester.num_hidden_layers)
+                self.assertAlmostEqual(
+                    attentions[0][..., 0, :, :].flatten().sum().item(), 0.0)
+                self.assertNotEqual(
+                    attentions[0][..., -1, :, :].flatten().sum().item(), 0.0)
+                self.assertNotEqual(
+                    attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
+                self.assertAlmostEqual(
+                    attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
+                self.assertNotEqual(
+                    attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)
 
 
-def _create_and_check_for_attentions(tester, model_classes, config, inputs_dict):
+        def test_head_pruning(self):
-    for model_class in model_classes:
+            if not self.test_pruning:
-        config.output_attentions = True
+                return
-        config.output_hidden_states = False
-        model = model_class(config)
-        model.eval()
-        outputs = model(**inputs_dict)
-        attentions = outputs[-1]
-        tester.parent.assertEqual(model.config.output_attentions, True)
-        tester.parent.assertEqual(model.config.output_hidden_states, False)
-        tester.parent.assertEqual(len(attentions), tester.num_hidden_layers)
-        tester.parent.assertListEqual(
-            list(attentions[0].shape[-3:]),
-            [tester.num_attention_heads,
-             tester.seq_length,
-             tester.key_len if hasattr(tester, 'key_len') else tester.seq_length])
-        out_len = len(outputs)
 
-        # Check attention is always last and order is fine
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        config.output_attentions = True
-        config.output_hidden_states = True
-        model = model_class(config)
-        model.eval()
-        outputs = model(**inputs_dict)
-        tester.parent.assertEqual(out_len+1, len(outputs))
-        tester.parent.assertEqual(model.config.output_attentions, True)
-        tester.parent.assertEqual(model.config.output_hidden_states, True)
 
-        attentions = outputs[-1]
+            for model_class in self.all_model_classes:
-        tester.parent.assertEqual(len(attentions), tester.num_hidden_layers)
+                config.output_attentions = True
-        tester.parent.assertListEqual(
+                config.output_hidden_states = False
-            list(attentions[0].shape[-3:]),
+                model = model_class(config=config)
-            [tester.num_attention_heads,
+                model.eval()
-             tester.seq_length,
+                heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
-             tester.key_len if hasattr(tester, 'key_len') else tester.seq_length])
+                                -1: [0]}
+                model.prune_heads(heads_to_prune)
+                outputs = model(**inputs_dict)
 
-def _create_and_check_for_hidden_states(tester, model_classes, config, inputs_dict):
+                attentions = outputs[-1]
-    for model_class in model_classes:
+
-        config.output_hidden_states = True
+                self.assertEqual(
-        config.output_attentions = False
+                    attentions[0].shape[-3], 1)
-        model = model_class(config)
+                self.assertEqual(
-        model.eval()
+                    attentions[1].shape[-3], self.model_tester.num_attention_heads)
-        outputs = model(**inputs_dict)
+                self.assertEqual(
-        hidden_states = outputs[-1]
+                    attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
-        tester.parent.assertEqual(model.config.output_attentions, False)
-        tester.parent.assertEqual(model.config.output_hidden_states, True)
-        tester.parent.assertEqual(len(hidden_states), tester.num_hidden_layers + 1)
-        tester.parent.assertListEqual(
-            list(hidden_states[0].shape[-2:]),
-            [tester.seq_length, tester.hidden_size])
 
 
-def create_and_check_commons(tester, config, inputs_dict, test_pruning=True, test_torchscript=True):
+        def test_hidden_states_output(self):
-    _create_and_check_initialization(tester, tester.all_model_classes, config, inputs_dict)
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-    _create_and_check_for_attentions(tester, tester.all_model_classes, config, inputs_dict)
-    _create_and_check_for_headmasking(tester, tester.all_model_classes, config, inputs_dict)
-    _create_and_check_for_hidden_states(tester, tester.all_model_classes, config, inputs_dict)
 
-    if test_torchscript:
+            for model_class in self.all_model_classes:
-        _create_and_check_torchscript(tester, tester.all_model_classes, config, inputs_dict)
+                config.output_hidden_states = True
-        _create_and_check_torchscript_output_attentions(tester, tester.all_model_classes, config, inputs_dict)
+                config.output_attentions = False
-        _create_and_check_torchscript_output_hidden_state(tester, tester.all_model_classes, config, inputs_dict)
+                model = model_class(config)
+                model.eval()
+                outputs = model(**inputs_dict)
+                hidden_states = outputs[-1]
+                self.assertEqual(model.config.output_attentions, False)
+                self.assertEqual(model.config.output_hidden_states, True)
+                self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
+                self.assertListEqual(
+                    list(hidden_states[0].shape[-2:]),
+                    [self.model_tester.seq_length, self.model_tester.hidden_size])
 
-    if test_pruning:
+        def test_resize_tokens_embeddings(self):
-        _create_and_check_for_head_pruning(tester, tester.all_model_classes, config, inputs_dict)
+            original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            if not self.test_resize_embeddings:
+                return
+
+            for model_class in self.all_model_classes:
+                config = copy.deepcopy(original_config)
+                model = model_class(config)
+
+                model_vocab_size = config.vocab_size
+                # Retrieve the embeddings and clone theme
+                model_embed = model.resize_token_embeddings(model_vocab_size)
+                cloned_embeddings = model_embed.weight.clone()
+
+                # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
+                model_embed = model.resize_token_embeddings(model_vocab_size + 10)
+                self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
+                # Check that it actually resizes the embeddings matrix
+                self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
+
+                # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
+                model_embed = model.resize_token_embeddings(model_vocab_size - 15)
+                self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
+                # Check that it actually resizes the embeddings matrix
+                self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
+
+                # Check that adding and removing tokens has not modified the first part of the embedding matrix.
+                models_equal = True
+                for p1, p2 in zip(cloned_embeddings, model_embed.weight):
+                    if p1.data.ne(p2.data).sum() > 0:
+                        models_equal = False
+
+                self.assertTrue(models_equal)
+
+        def test_tie_model_weights(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            def check_same_values(layer_1, layer_2):
+                equal = True
+                for p1, p2 in zip(layer_1.weight, layer_2.weight):
+                    if p1.data.ne(p2.data).sum() > 0:
+                        equal = False
+                return equal
+
+            for model_class in self.all_model_classes:
+                if not hasattr(model_class, 'tie_weights'):
+                    continue
+
+                config.torchscript = True
+                model_not_tied = model_class(config)
+                params_not_tied = list(model_not_tied.parameters())
+
+                config_tied = copy.deepcopy(config)
+                config_tied.torchscript = False
+                model_tied = model_class(config_tied)
+                params_tied = list(model_tied.parameters())
+
+                # Check that the embedding layer and decoding layer are the same in size and in value
+                self.assertGreater(len(params_not_tied), len(params_tied))
+                # self.assertTrue(check_same_values(embeddings, decoding))
+
+                # # Check that after modification, they remain the same.
+                # embeddings.weight.data.div_(2)
+                # # Check that the embedding layer and decoding layer are the same in size and in value
+                # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
+                # self.assertTrue(check_same_values(embeddings, decoding))
+
+                # # Check that after modification, they remain the same.
+                # decoding.weight.data.div_(4)
+                # # Check that the embedding layer and decoding layer are the same in size and in value
+                # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
+                # self.assertTrue(check_same_values(embeddings, decoding))
+
+                # Check that after resize they remain tied.
+                model_tied.resize_token_embeddings(config.vocab_size + 10)
+                params_tied_2 = list(model_tied.parameters())
+                self.assertGreater(len(params_not_tied), len(params_tied))
+                self.assertEqual(len(params_tied_2), len(params_tied))
+
+                # decoding.weight.data.mul_(20)
+                # # Check that the embedding layer and decoding layer are the same in size and in value
+                # self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
+                # self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))
 
 
-def ids_tensor(shape, vocab_size, rng=None, name=None):
+    class GPTModelTester(CommonModelTester):
-    """Creates a random int32 tensor of the shape within the vocab size."""
-    if rng is None:
-        rng = random.Random()
 
-    total_dims = 1
+        def __init__(self,
-    for dim in shape:
+                        parent,
-        total_dims *= dim
+                        batch_size=13,
+                        seq_length=7,
+                        is_training=True,
+                        use_position_ids=True,
+                        use_token_type_ids=True,
+                        use_labels=True,
+                        vocab_size=99,
+                        n_positions=33,
+                        hidden_size=32,
+                        num_hidden_layers=5,
+                        num_attention_heads=4,
+                        n_choices=3,
+                        type_sequence_label_size=2,
+                        initializer_range=0.02,
+                        num_labels=3,
+                        scope=None,
+                        config_class=None,
+                        base_model_class=None,
+                        lm_head_model_class=None,
+                        double_head_model_class=None,
+                        ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_position_ids = use_position_ids
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.n_positions = n_positions
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.n_choices = n_choices
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.scope = scope
+            self.config_class = config_class
+            self.base_model_class = base_model_class
+            self.lm_head_model_class = lm_head_model_class
+            self.double_head_model_class = double_head_model_class
+            self.all_model_classes = (base_model_class, lm_head_model_class, double_head_model_class)
 
-    values = []
+        def prepare_config_and_inputs(self):
-    for _ in range(total_dims):
+            total_num_tokens = self.vocab_size
-        values.append(rng.randint(0, vocab_size - 1))
+            input_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_num_tokens)
 
-    return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
+            position_ids = None
+            if self.use_position_ids:
+                position_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.n_positions)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                total_voc = self.vocab_size
+                token_type_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_voc)
+
+            mc_labels = None
+            lm_labels = None
+            mc_token_ids = None
+            if self.use_labels:
+                mc_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                lm_labels = ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.num_labels)
+                mc_token_ids = ids_tensor([self.batch_size, self.n_choices], self.seq_length)
+
+            config = self.config_class(
+                vocab_size_or_config_json_file=self.vocab_size,
+                n_positions=self.n_positions,
+                n_embd=self.hidden_size,
+                n_layer=self.num_hidden_layers,
+                n_head=self.num_attention_heads,
+                initializer_range=self.initializer_range)
+
+            return (config, input_ids, token_type_ids, position_ids,
+                    mc_labels, lm_labels, mc_token_ids)
+
+        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.eval()
+
+            outputs = model(input_ids, position_ids, token_type_ids)
+            outputs = model(input_ids, position_ids)
+            outputs = model(input_ids)
+
+            hidden_state = outputs[0]
+            self.parent.assertListEqual(
+                list(hidden_state.size()),
+                [self.batch_size, self.n_choices, self.seq_length, self.hidden_size])
+
+
+        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.eval()
+            outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
+            loss, lm_logits = outputs[:2]
+
+            total_voc = self.vocab_size
+            self.parent.assertListEqual(
+                list(lm_logits.size()),
+                [self.batch_size, self.n_choices, self.seq_length, total_voc])
+            self.parent.assertListEqual(
+                list(loss.size()),
+                [])
+
+        def create_and_check_presents(self, config, input_ids, token_type_ids, position_ids,
+                                        mc_labels, lm_labels, mc_token_ids):
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+                model.eval()
+                outputs = model(input_ids)
+                presents = outputs[-1]
+                self.parent.assertEqual(self.num_hidden_layers, len(presents))
+                self.parent.assertListEqual(
+                    list(presents[0].size()),
+                    [2, self.batch_size * self.n_choices, self.num_attention_heads,
+                        self.seq_length, self.hidden_size // self.num_attention_heads])
+
+        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.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)
+            lm_loss, mc_loss, lm_logits, mc_logits = outputs[:4]
+            loss = [lm_loss, mc_loss]
+
+            total_voc = self.vocab_size
+            self.parent.assertListEqual(
+                list(lm_logits.size()),
+                [self.batch_size, self.n_choices, self.seq_length, total_voc])
+            self.parent.assertListEqual(
+                list(mc_logits.size()),
+                [self.batch_size, self.n_choices])
+            self.parent.assertListEqual(
+                [list(l.size()) for l in loss],
+                [[], []])
+
+        def create_and_check_model_from_pretrained(self):
+            cache_dir = "/tmp/pytorch_transformers_test/"
+            for model_name in list(self.base_model_class.pretrained_model_archive_map.keys())[:1]:
+                model = self.base_model_class.from_pretrained(model_name, cache_dir=cache_dir)
+                shutil.rmtree(cache_dir)
+                self.parent.assertIsNotNone(model)
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, position_ids,
+                mc_labels, lm_labels, mc_token_ids) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids}
+            return config, inputs_dict
+
+        def run_common_tests(self, test_presents=False):
+            config_and_inputs = self.prepare_config_and_inputs()
+            self.create_and_check_base_model(*config_and_inputs)
+
+            config_and_inputs = self.prepare_config_and_inputs()
+            self.create_and_check_lm_head(*config_and_inputs)
+
+            config_and_inputs = self.prepare_config_and_inputs()
+            self.create_and_check_double_heads(*config_and_inputs)
+
+            if test_presents:
+                config_and_inputs = self.prepare_config_and_inputs()
+                self.create_and_check_presents(*config_and_inputs)
+
+        def run_slow_tests(self):
+            self.create_and_check_model_from_pretrained()
 
 
 class ConfigTester(object):
@@ -275,179 +539,22 @@ class ConfigTester(object):
         self.create_and_test_config_to_json_file()
 
 
-class GPTModelTester(object):
-    def __init__(self,
-                    parent,
-                    batch_size=13,
-                    seq_length=7,
-                    is_training=True,
-                    use_position_ids=True,
-                    use_token_type_ids=True,
-                    use_labels=True,
-                    vocab_size=99,
-                    n_positions=33,
-                    hidden_size=32,
-                    num_hidden_layers=5,
-                    num_attention_heads=4,
-                    n_choices=3,
-                    type_sequence_label_size=2,
-                    initializer_range=0.02,
-                    num_labels=3,
-                    scope=None,
-                    config_class=None,
-                    base_model_class=None,
-                    lm_head_model_class=None,
-                    double_head_model_class=None,
-                    ):
-        self.parent = parent
-        self.batch_size = batch_size
-        self.seq_length = seq_length
-        self.is_training = is_training
-        self.use_position_ids = use_position_ids
-        self.use_token_type_ids = use_token_type_ids
-        self.use_labels = use_labels
-        self.vocab_size = vocab_size
-        self.n_positions = n_positions
-        self.hidden_size = hidden_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.n_choices = n_choices
-        self.type_sequence_label_size = type_sequence_label_size
-        self.initializer_range = initializer_range
-        self.num_labels = num_labels
-        self.scope = scope
-        self.config_class = config_class
-        self.base_model_class = base_model_class
-        self.lm_head_model_class = lm_head_model_class
-        self.double_head_model_class = double_head_model_class
-        self.all_model_classes = (base_model_class, lm_head_model_class, double_head_model_class)
-
-    def prepare_config_and_inputs(self):
-        total_num_tokens = self.vocab_size
-        input_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_num_tokens)
-
-        position_ids = None
-        if self.use_position_ids:
-            position_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.n_positions)
-
-        token_type_ids = None
-        if self.use_token_type_ids:
-            total_voc = self.vocab_size
-            token_type_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_voc)
-
-        mc_labels = None
-        lm_labels = None
-        mc_token_ids = None
-        if self.use_labels:
-            mc_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
-            lm_labels = ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.num_labels)
-            mc_token_ids = ids_tensor([self.batch_size, self.n_choices], self.seq_length)
-
-        config = self.config_class(
-            vocab_size_or_config_json_file=self.vocab_size,
-            n_positions=self.n_positions,
-            n_embd=self.hidden_size,
-            n_layer=self.num_hidden_layers,
-            n_head=self.num_attention_heads,
-            initializer_range=self.initializer_range)
-
-        return (config, input_ids, token_type_ids, position_ids,
-                mc_labels, lm_labels, mc_token_ids)
-
-    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.eval()
-
-        outputs = model(input_ids, position_ids, token_type_ids)
-        outputs = model(input_ids, position_ids)
-        outputs = model(input_ids)
-
-        hidden_state = outputs[0]
-        self.parent.assertListEqual(
-            list(hidden_state.size()),
-            [self.batch_size, self.n_choices, self.seq_length, self.hidden_size])
 
 
-    def create_and_check_lm_head(self, config, input_ids, token_type_ids, position_ids,
+def ids_tensor(shape, vocab_size, rng=None, name=None):
-                                    mc_labels, lm_labels, mc_token_ids):
+    """Creates a random int32 tensor of the shape within the vocab size."""
-        model = self.lm_head_model_class(config)
+    if rng is None:
-        model.eval()
+        rng = random.Random()
-        outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
-        loss, lm_logits = outputs[:2]
 
-        total_voc = self.vocab_size
+    total_dims = 1
-        self.parent.assertListEqual(
+    for dim in shape:
-            list(lm_logits.size()),
+        total_dims *= dim
-            [self.batch_size, self.n_choices, self.seq_length, total_voc])
-        self.parent.assertListEqual(
-            list(loss.size()),
-            [])
 
-    def create_and_check_presents(self, config, input_ids, token_type_ids, position_ids,
+    values = []
-                                    mc_labels, lm_labels, mc_token_ids):
+    for _ in range(total_dims):
-        for model_class in self.all_model_classes:
+        values.append(rng.randint(0, vocab_size - 1))
-            model = model_class(config)
-            model.eval()
-            outputs = model(input_ids)
-            presents = outputs[-1]
-            self.parent.assertEqual(self.num_hidden_layers, len(presents))
-            self.parent.assertListEqual(
-                list(presents[0].size()),
-                [2, self.batch_size * self.n_choices, self.num_attention_heads,
-                    self.seq_length, self.hidden_size // self.num_attention_heads])
 
-    def create_and_check_double_heads(self, config, input_ids, token_type_ids, position_ids,
+    return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
-                                    mc_labels, lm_labels, mc_token_ids):
-        model = self.double_head_model_class(config)
-        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)
-        lm_loss, mc_loss, lm_logits, mc_logits = outputs[:4]
-        loss = [lm_loss, mc_loss]
-
-        total_voc = self.vocab_size
-        self.parent.assertListEqual(
-            list(lm_logits.size()),
-            [self.batch_size, self.n_choices, self.seq_length, total_voc])
-        self.parent.assertListEqual(
-            list(mc_logits.size()),
-            [self.batch_size, self.n_choices])
-        self.parent.assertListEqual(
-            [list(l.size()) for l in loss],
-            [[], []])
-
-    def create_and_check_model_from_pretrained(self):
-        cache_dir = "/tmp/pytorch_transformers_test/"
-        for model_name in list(self.base_model_class.pretrained_model_archive_map.keys())[:1]:
-            model = self.base_model_class.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
-            self.parent.assertIsNotNone(model)
-
-    def create_and_check_commons(self, config, input_ids, token_type_ids, position_ids,
-                                    mc_labels, lm_labels, mc_token_ids):
-        inputs_dict = {'input_ids': input_ids}
-        create_and_check_commons(self, config, inputs_dict)
-
-    def run_common_tests(self, test_presents=False):
-        config_and_inputs = self.prepare_config_and_inputs()
-        self.create_and_check_base_model(*config_and_inputs)
-
-        config_and_inputs = self.prepare_config_and_inputs()
-        self.create_and_check_lm_head(*config_and_inputs)
-
-        config_and_inputs = self.prepare_config_and_inputs()
-        self.create_and_check_double_heads(*config_and_inputs)
-
-        if test_presents:
-            config_and_inputs = self.prepare_config_and_inputs()
-            self.create_and_check_presents(*config_and_inputs)
-
-        config_and_inputs = self.prepare_config_and_inputs()
-        self.create_and_check_commons(*config_and_inputs)
-
-    def run_slow_tests(self):
-        self.create_and_check_model_from_pretrained()
 
 
 class ModelUtilsTest(unittest.TestCase):
@@ -471,79 +578,6 @@ class ModelUtilsTest(unittest.TestCase):
             self.assertEqual(model.config.output_hidden_states, True)
             self.assertEqual(model.config, config)
 
-    def test_resize_tokens_embeddings(self):
-        logging.basicConfig(level=logging.INFO)
-
-
-        for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            config = BertConfig.from_pretrained(model_name)
-            model = BertModel.from_pretrained(model_name)
-
-            model_vocab_size = config.vocab_size
-            # Retrieve the embeddings and clone theme
-            cloned_embeddings = model.embeddings.word_embeddings.weight.clone()
-
-            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
-            model.resize_token_embeddings(model_vocab_size + 10)
-            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
-            # Check that it actually resizes the embeddings matrix
-            self.assertEqual(model.embeddings.word_embeddings.weight.shape[0], cloned_embeddings.shape[0] + 10)
-
-            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
-            model.resize_token_embeddings(model_vocab_size)
-            self.assertEqual(model.config.vocab_size, model_vocab_size)
-            # Check that it actually resizes the embeddings matrix
-            self.assertEqual(model.embeddings.word_embeddings.weight.shape[0], cloned_embeddings.shape[0])
-
-            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
-            models_equal = True
-            for p1, p2 in zip(cloned_embeddings, model.embeddings.word_embeddings.weight):
-                if p1.data.ne(p2.data).sum() > 0:
-                    models_equal = False
-
-            self.assertTrue(models_equal)
-
-    def test_tie_model_weights(self):
-        logging.basicConfig(level=logging.INFO)
-
-        def check_same_values(layer_1, layer_2):
-            equal = True
-            for p1, p2 in zip(layer_1.weight, layer_2.weight):
-                if p1.data.ne(p2.data).sum() > 0:
-                    equal = False
-            return equal
-
-        for model_name in list(GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            config = GPT2Config.from_pretrained(model_name)
-            model = GPT2LMHeadModel.from_pretrained(model_name)
-
-            # Get the embeddings and decoding layer
-            embeddings = model.transformer.wte
-            decoding = model.lm_head
-
-            # Check that the embedding layer and decoding layer are the same in size and in value
-            self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
-            self.assertTrue(check_same_values(embeddings, decoding))
-
-            # Check that after modification, they remain the same.
-            embeddings.weight.data.div_(2)
-            # Check that the embedding layer and decoding layer are the same in size and in value
-            self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
-            self.assertTrue(check_same_values(embeddings, decoding))
-
-            # Check that after modification, they remain the same.
-            decoding.weight.data.div_(4)
-            # Check that the embedding layer and decoding layer are the same in size and in value
-            self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
-            self.assertTrue(check_same_values(embeddings, decoding))
-
-            # Check that after resize they remain tied.
-            model.resize_token_embeddings(config.vocab_size + 10)
-            decoding.weight.data.mul_(20)
-            # Check that the embedding layer and decoding layer are the same in size and in value
-            self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
-            self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))
-
 
 if __name__ == "__main__":
     unittest.main()

pytorch_transformers/tests/modeling_gpt2_test.py

@@ -16,19 +16,14 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import os
 import unittest
-import json
-import random
-import shutil
 import pytest
 
-import torch
 
 from pytorch_transformers import (GPT2Config, GPT2Model,
-                                     GPT2LMHeadModel, GPT2DoubleHeadsModel)
+                                  GPT2LMHeadModel, GPT2DoubleHeadsModel)
 
-from .modeling_common_test import (create_and_check_commons, ConfigTester, GPTModelTester)
+from .modeling_common_test import CommonTestCases, ConfigTester
 
 class GPT2ModelTest(unittest.TestCase):
 
@@ -37,14 +32,14 @@ class GPT2ModelTest(unittest.TestCase):
         config_tester.run_common_tests()
 
     def test_model(self):
-        model_tester = GPTModelTester(self, config_class=GPT2Config, base_model_class=GPT2Model,
+        model_tester = CommonTestCases.GPTModelTester(self, config_class=GPT2Config, base_model_class=GPT2Model,
                                             lm_head_model_class=GPT2LMHeadModel,
                                             double_head_model_class=GPT2DoubleHeadsModel)
         model_tester.run_common_tests(test_presents=True)
 
     @pytest.mark.slow
     def test_pretrained(self):
-        model_tester = GPTModelTester(self, config_class=GPT2Config, base_model_class=GPT2Model,
+        model_tester = CommonTestCases.GPTModelTester(self, config_class=GPT2Config, base_model_class=GPT2Model,
                                             lm_head_model_class=GPT2LMHeadModel,
                                             double_head_model_class=GPT2DoubleHeadsModel)
         model_tester.run_slow_tests()

pytorch_transformers/tests/modeling_openai_test.py

@@ -19,12 +19,11 @@ from __future__ import print_function
 import unittest
 import pytest
 
-import torch
 
 from pytorch_transformers import (OpenAIGPTConfig, OpenAIGPTModel,
                                      OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel)
 
-from .modeling_common_test import (create_and_check_commons, ConfigTester, GPTModelTester)
+from .modeling_common_test import CommonTestCases, ConfigTester
 
 class OpenAIModelTest(unittest.TestCase):
 
@@ -33,14 +32,14 @@ class OpenAIModelTest(unittest.TestCase):
         config_tester.run_common_tests()
 
     def test_model(self):
-        model_tester = GPTModelTester(self, config_class=OpenAIGPTConfig, base_model_class=OpenAIGPTModel,
+        model_tester = CommonTestCases.GPTModelTester(self, config_class=OpenAIGPTConfig, base_model_class=OpenAIGPTModel,
                                            lm_head_model_class=OpenAIGPTLMHeadModel,
                                            double_head_model_class=OpenAIGPTDoubleHeadsModel)
         model_tester.run_common_tests(test_presents=False)
 
     @pytest.mark.slow
     def test_pretrained(self):
-        model_tester = GPTModelTester(self, config_class=OpenAIGPTConfig, base_model_class=OpenAIGPTModel,
+        model_tester = CommonTestCases.GPTModelTester(self, config_class=OpenAIGPTConfig, base_model_class=OpenAIGPTModel,
                                            lm_head_model_class=OpenAIGPTLMHeadModel,
                                            double_head_model_class=OpenAIGPTDoubleHeadsModel)
         model_tester.run_slow_tests()

pytorch_transformers/tests/modeling_transfo_xl_test.py

@@ -28,9 +28,15 @@ import torch
 from pytorch_transformers import (TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel)
 from pytorch_transformers.modeling_transfo_xl import TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
 
-from .modeling_common_test import ConfigTester, create_and_check_commons, ids_tensor
+from .modeling_common_test import ConfigTester, CommonTestCases, ids_tensor
+
+class TransfoXLModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (TransfoXLModel, TransfoXLLMHeadModel)
+    test_pruning = False
+    test_torchscript = False
+    test_resize_embeddings = False
 
-class TransfoXLModelTest(unittest.TestCase):
     class TransfoXLModelTester(object):
 
         def __init__(self,
@@ -52,7 +58,6 @@ class TransfoXLModelTest(unittest.TestCase):
                      num_hidden_layers=5,
                      scope=None,
                      seed=1,
-                     all_model_classes=(TransfoXLModel, TransfoXLLMHeadModel),
                      ):
             self.parent = parent
             self.batch_size = batch_size
@@ -73,7 +78,6 @@ class TransfoXLModelTest(unittest.TestCase):
             self.num_hidden_layers = num_hidden_layers
             self.scope = scope
             self.seed = seed
-            self.all_model_classes = all_model_classes
 
         def prepare_config_and_inputs(self):
             input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -171,16 +175,31 @@ class TransfoXLModelTest(unittest.TestCase):
                 list(list(mem.size()) for mem in result["mems_2"]),
                 [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
 
-        def create_and_check_transfo_xl_commons(self, config, input_ids_1, input_ids_2, lm_labels):
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids_1, input_ids_2, lm_labels) = config_and_inputs
             inputs_dict = {'input_ids': input_ids_1}
-            create_and_check_commons(self, config, inputs_dict, test_pruning=False, test_torchscript=False)
+            return config, inputs_dict
 
-    def test_default(self):
+
-        self.run_tester(TransfoXLModelTest.TransfoXLModelTester(self))
+    def setUp(self):
+        self.model_tester = TransfoXLModelTest.TransfoXLModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=TransfoXLConfig, d_embed=37)
 
     def test_config(self):
-        config_tester = ConfigTester(self, config_class=TransfoXLConfig, d_embed=37)
+        self.config_tester.run_common_tests()
-        config_tester.run_common_tests()
+
+    def test_transfo_xl_model(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        output_result = self.model_tester.create_transfo_xl_model(*config_and_inputs)
+        self.model_tester.check_transfo_xl_model_output(output_result)
+
+    def test_transfo_xl_lm_head(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        output_result = self.model_tester.create_transfo_xl_lm_head(*config_and_inputs)
+        self.model_tester.check_transfo_xl_lm_head_output(output_result)
 
     @pytest.mark.slow
     def test_model_from_pretrained(self):
@@ -190,23 +209,6 @@ class TransfoXLModelTest(unittest.TestCase):
             shutil.rmtree(cache_dir)
             self.assertIsNotNone(model)
 
-    def run_tester(self, tester):
-        config_and_inputs = tester.prepare_config_and_inputs()
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        output_result = tester.create_transfo_xl_model(*config_and_inputs)
-        tester.check_transfo_xl_model_output(output_result)
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        output_result = tester.create_transfo_xl_lm_head(*config_and_inputs)
-        tester.check_transfo_xl_lm_head_output(output_result)
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_transfo_xl_commons(*config_and_inputs)
-
 
 if __name__ == "__main__":
     unittest.main()

pytorch_transformers/tests/modeling_xlm_test.py

@@ -23,10 +23,15 @@ import pytest
 from pytorch_transformers import (XLMConfig, XLMModel, XLMWithLMHeadModel, XLMForQuestionAnswering, XLMForSequenceClassification)
 from pytorch_transformers.modeling_xlm import XLM_PRETRAINED_MODEL_ARCHIVE_MAP
 
-from .modeling_common_test import (create_and_check_commons, ConfigTester, ids_tensor)
+from .modeling_common_test import (CommonTestCases, ConfigTester, ids_tensor)
 
 
-class XLMModelTest(unittest.TestCase):
+class XLMModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (XLMModel, XLMWithLMHeadModel,  
+                         XLMForQuestionAnswering, XLMForSequenceClassification) 
+                         # , XLMForSequenceClassification, XLMForTokenClassification),
+
     class XLMModelTester(object):
 
         def __init__(self,
@@ -58,8 +63,6 @@ class XLMModelTest(unittest.TestCase):
                      summary_type="last",
                      use_proj=True,
                      scope=None,
-                     all_model_classes = (XLMModel, XLMWithLMHeadModel,
-                                          XLMForQuestionAnswering, XLMForSequenceClassification),  # , XLMForSequenceClassification, XLMForTokenClassification),
                     ):
             self.parent = parent
             self.batch_size = batch_size
@@ -90,7 +93,6 @@ class XLMModelTest(unittest.TestCase):
             self.num_labels = num_labels
             self.num_choices = num_choices
             self.scope = scope
-            self.all_model_classes = all_model_classes
 
         def prepare_config_and_inputs(self):
             input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -237,28 +239,23 @@ class XLMModelTest(unittest.TestCase):
                 [self.batch_size, self.type_sequence_label_size])
 
 
-        def create_and_check_xlm_commons(self, config, input_ids, token_type_ids, input_lengths, sequence_labels, token_labels, is_impossible_labels, input_mask):
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_lengths,
+             sequence_labels, token_labels, is_impossible_labels, input_mask) = config_and_inputs
             inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'lengths': input_lengths}
-            create_and_check_commons(self, config, inputs_dict)
+            return config, inputs_dict
 
-    def test_default(self):
+    def setUp(self):
-        self.run_tester(XLMModelTest.XLMModelTester(self))
+        self.model_tester = XLMModelTest.XLMModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=XLMConfig, emb_dim=37)
 
     def test_config(self):
-        config_tester = ConfigTester(self, config_class=XLMConfig, emb_dim=37)
+        self.config_tester.run_common_tests()
-        config_tester.run_common_tests()
 
-    @pytest.mark.slow
+    def test_xlm_model(self):
-    def test_model_from_pretrained(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
-        cache_dir = "/tmp/pytorch_transformers_test/"
+        self.model_tester.create_and_check_xlm_model(*config_and_inputs)
-        for model_name in list(XLM_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XLMModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
-            self.assertIsNotNone(model)
-
-    def run_tester(self, tester):
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_xlm_model(*config_and_inputs)
 
         # config_and_inputs = tester.prepare_config_and_inputs()
         # tester.create_and_check_xlm_for_masked_lm(*config_and_inputs)
@@ -275,8 +272,14 @@ class XLMModelTest(unittest.TestCase):
         # config_and_inputs = tester.prepare_config_and_inputs()
         # tester.create_and_check_xlm_for_token_classification(*config_and_inputs)
 
-        config_and_inputs = tester.prepare_config_and_inputs()
+    @pytest.mark.slow
-        tester.create_and_check_xlm_commons(*config_and_inputs)
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/pytorch_transformers_test/"
+        for model_name in list(XLM_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = XLMModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
 
 if __name__ == "__main__":
     unittest.main()

pytorch_transformers/tests/modeling_xlnet_test.py

@@ -28,9 +28,14 @@ import torch
 from pytorch_transformers import (XLNetConfig, XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering)
 from pytorch_transformers.modeling_xlnet import XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
 
-from .modeling_common_test import ConfigTester, create_and_check_commons, ids_tensor
+from .modeling_common_test import ConfigTester, CommonTestCases, ids_tensor
+
+class XLNetModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes=(XLNetModel, XLNetLMHeadModel,
+                    XLNetForSequenceClassification, XLNetForQuestionAnswering)
+    test_pruning = False
 
-class XLNetModelTest(unittest.TestCase):
     class XLNetModelTester(object):
 
         def __init__(self,
@@ -56,8 +61,6 @@ class XLNetModelTest(unittest.TestCase):
                      initializer_range=0.05,
                      seed=1,
                      type_vocab_size=2,
-                     all_model_classes=(XLNetModel, XLNetLMHeadModel,
-                                        XLNetForSequenceClassification, XLNetForQuestionAnswering),
             ):
             self.parent = parent
             self.batch_size = batch_size
@@ -82,7 +85,6 @@ class XLNetModelTest(unittest.TestCase):
             self.seed = seed
             self.type_vocab_size = type_vocab_size
             self.type_sequence_label_size = type_sequence_label_size
-            self.all_model_classes = all_model_classes
 
         def prepare_config_and_inputs(self):
             input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -264,17 +266,41 @@ class XLNetModelTest(unittest.TestCase):
                 list(list(mem.size()) for mem in result["mems_1"]),
                 [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
 
-        def create_and_check_xlnet_commons(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
+        def prepare_config_and_inputs_for_common(self):
-                target_mapping, inp_q, segment_ids, lm_labels, sequence_labels, is_impossible_labels):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
+                target_mapping, inp_q, segment_ids, lm_labels,
+                sequence_labels, is_impossible_labels) = config_and_inputs
             inputs_dict = {'input_ids': input_ids_1}
-            create_and_check_commons(self, config, inputs_dict, test_pruning=False)
+            return config, inputs_dict
 
-    def test_default(self):
+
-        self.run_tester(XLNetModelTest.XLNetModelTester(self))
+    def setUp(self):
+        self.model_tester = XLNetModelTest.XLNetModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=XLNetConfig, d_inner=37)
 
     def test_config(self):
-        config_tester = ConfigTester(self, config_class=XLNetConfig, d_inner=37)
+        self.config_tester.run_common_tests()
-        config_tester.run_common_tests()
+
+    def test_xlnet_base_model(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xlnet_base_model(*config_and_inputs)
+
+    def test_xlnet_lm_head(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xlnet_lm_head(*config_and_inputs) 
+
+    def test_xlnet_sequence_classif(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xlnet_sequence_classif(*config_and_inputs)
+
+    def test_xlnet_qa(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xlnet_qa(*config_and_inputs)
 
     @pytest.mark.slow
     def test_model_from_pretrained(self):
@@ -284,27 +310,6 @@ class XLNetModelTest(unittest.TestCase):
             shutil.rmtree(cache_dir)
             self.assertIsNotNone(model)
 
-    def run_tester(self, tester):
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_xlnet_base_model(*config_and_inputs)
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_xlnet_lm_head(*config_and_inputs) 
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_xlnet_sequence_classif(*config_and_inputs)
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_xlnet_qa(*config_and_inputs)
-
-        tester.set_seed()
-        config_and_inputs = tester.prepare_config_and_inputs()
-        tester.create_and_check_xlnet_commons(*config_and_inputs)
-
 
 if __name__ == "__main__":
     unittest.main()