[LARGE] updating all tests and API

1484d67d · thomwolf · 4f8b5f68 · 1484d67d · 1484d67d · 1484d67d
Commit 1484d67d authored Jul 02, 2019 by thomwolf
20 changed files
--- a/pytorch_pretrained_bert/model_utils.py
+++ b/pytorch_pretrained_bert/model_utils.py
@@ -41,6 +41,12 @@ class PretrainedConfig(object):
    """
    pretrained_config_archive_map = {}
+    def __init__(self, **kwargs):
+        self.finetuning_task = kwargs.pop('finetuning_task', None)
+        self.num_labels = kwargs.pop('num_labels', 2)
+        self.output_attentions = kwargs.pop('output_attentions', False)
+        self.output_hidden_states = kwargs.pop('output_hidden_states', False)
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
        """
@@ -114,6 +120,9 @@ class PretrainedConfig(object):
            text = reader.read()
        return cls.from_dict(json.loads(text))
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
    def __repr__(self):
        return str(self.to_json_string())
@@ -133,12 +142,11 @@ class PretrainedConfig(object):
 class PreTrainedModel(nn.Module):
-    """ An abstract class to handle weights initialization and
+    """ An abstract class to handle storing model config and
        a simple interface for dowloading and loading pretrained models.
    """
    config_class = PretrainedConfig
    pretrained_model_archive_map = {}
-    pretrained_config_archive_map = {}
    load_tf_weights = lambda model, config, path: None
    base_model_prefix = ""
@@ -151,8 +159,16 @@ class PreTrainedModel(nn.Module):
                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
                    self.__class__.__name__, self.__class__.__name__
                ))
+        # Save config in model
        self.config = config
+    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}
+        """
+        model_to_prune = getattr(self, self.base_model_prefix, self)  # get the base model if needed
+        model_to_prune._prune_heads(heads_to_prune)
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
        """
@@ -175,24 +191,22 @@ class PreTrainedModel(nn.Module):
            *inputs, **kwargs: additional input for the specific XLNet class
                (ex: num_labels for XLNetForSequenceClassification)
        """
-        state_dict = kwargs.get('state_dict', None)
+        state_dict = kwargs.pop('state_dict', None)
-        kwargs.pop('state_dict', None)
+        cache_dir = kwargs.pop('cache_dir', None)
-        cache_dir = kwargs.get('cache_dir', None)
+        from_tf = kwargs.pop('from_tf', None)
-        kwargs.pop('cache_dir', None)
-        from_tf = kwargs.get('from_tf', False)
+        # Load config
-        kwargs.pop('from_tf', None)
+        config = cls.config_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        # Load model
        if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
            archive_file = cls.pretrained_model_archive_map[pretrained_model_name_or_path]
-            config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
        else:
            if from_tf:
                # Directly load from a TensorFlow checkpoint
                archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
-                config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
            else:
                archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
-                config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
        # redirect to the cache, if necessary
        try:
            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir)
@@ -210,47 +224,15 @@ class PreTrainedModel(nn.Module):
                        ', '.join(cls.pretrained_model_archive_map.keys()),
                        archive_file))
            return None
-        try:
+        if resolved_archive_file == archive_file:
-            resolved_config_file = cached_path(config_file, cache_dir=cache_dir)
-        except EnvironmentError:
-            if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
-                logger.error(
-                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
-                        config_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(
-                        pretrained_model_name_or_path,
-                        ', '.join(cls.pretrained_config_archive_map.keys()),
-                        config_file))
-            return None
-        if resolved_archive_file == archive_file and resolved_config_file == config_file:
            logger.info("loading weights file {}".format(archive_file))
-            logger.info("loading configuration file {}".format(config_file))
        else:
            logger.info("loading weights file {} from cache at {}".format(
                archive_file, resolved_archive_file))
-            logger.info("loading configuration file {} from cache at {}".format(
-                config_file, resolved_config_file))
-        # Load config
-        config = cls.config_class.from_json_file(resolved_config_file)
-        # Update config with kwargs if needed
-        to_remove = []
-        for key, value in kwargs.items():
-            if hasattr(config, key):
-                setattr(config, key, value)
-                to_remove.append(key)
-        for key in to_remove:
-            kwargs.pop(key, None)
-        logger.info("Model config {}".format(config))
        # Instantiate model.
-        model = cls(config, *inputs, **kwargs)
+        model = cls(config)
        if state_dict is None and not from_tf:
            state_dict = torch.load(resolved_archive_file, map_location='cpu')
        if from_tf:
@@ -275,7 +257,7 @@ class PreTrainedModel(nn.Module):
                if child is not None:
                    load(child, prefix + name + '.')
-        # Be able to load base models as well as derived models (with heads)
+        # Make sure we are able to load base models as well as derived models (with heads)
        start_prefix = ''
        model_to_load = model
        if not hasattr(model, cls.base_model_prefix) and any(s.startswith(cls.base_model_prefix) for s in state_dict.keys()):

--- a/pytorch_pretrained_bert/modeling.py
+++ b/pytorch_pretrained_bert/modeling.py
@@ -155,7 +155,7 @@ class BertConfig(PretrainedConfig):
    pretrained_config_archive_map = PRETRAINED_CONFIG_ARCHIVE_MAP
    def __init__(self,
-                 vocab_size_or_config_json_file,
+                 vocab_size_or_config_json_file=30522,
                 hidden_size=768,
                 num_hidden_layers=12,
                 num_attention_heads=12,
@@ -167,7 +167,7 @@ class BertConfig(PretrainedConfig):
                 type_vocab_size=2,
                 initializer_range=0.02,
                 layer_norm_eps=1e-12,
-                 finetuning_task=None):
+                 **kwargs):
        """Constructs BertConfig.
        Args:
@@ -192,8 +192,8 @@ class BertConfig(PretrainedConfig):
            initializer_range: The sttdev of the truncated_normal_initializer for
                initializing all weight matrices.
            layer_norm_eps: The epsilon used by LayerNorm.
-            finetuning_task: name of the glue task on which the model was fine-tuned if any
        """
+        super(BertConfig, self).__init__(**kwargs)
        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                        and isinstance(vocab_size_or_config_json_file, unicode)):
            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
@@ -213,7 +213,6 @@ class BertConfig(PretrainedConfig):
            self.type_vocab_size = type_vocab_size
            self.initializer_range = initializer_range
            self.layer_norm_eps = layer_norm_eps
-            self.finetuning_task = finetuning_task
        else:
            raise ValueError("First argument must be either a vocabulary size (int)"
                             "or the path to a pretrained model config file (str)")
@@ -270,13 +269,13 @@ class BertEmbeddings(nn.Module):
 class BertSelfAttention(nn.Module):
-    def __init__(self, config, output_attentions=False):
+    def __init__(self, config):
        super(BertSelfAttention, self).__init__()
        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
-        self.output_attentions = output_attentions
+        self.output_attentions = config.output_attentions
        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
@@ -344,10 +343,9 @@ class BertSelfOutput(nn.Module):
 class BertAttention(nn.Module):
-    def __init__(self, config, output_attentions=False):
+    def __init__(self, config):
        super(BertAttention, self).__init__()
-        self.output_attentions = output_attentions
+        self.self = BertSelfAttention(config)
-        self.self = BertSelfAttention(config, output_attentions=output_attentions)
        self.output = BertSelfOutput(config)
    def prune_heads(self, heads):
@@ -404,10 +402,9 @@ class BertOutput(nn.Module):
 class BertLayer(nn.Module):
-    def __init__(self, config, output_attentions=False):
+    def __init__(self, config):
        super(BertLayer, self).__init__()
-        self.output_attentions = output_attentions
+        self.attention = BertAttention(config)
-        self.attention = BertAttention(config, output_attentions=output_attentions)
        self.intermediate = BertIntermediate(config)
        self.output = BertOutput(config)
@@ -421,11 +418,11 @@ class BertLayer(nn.Module):
 class BertEncoder(nn.Module):
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertEncoder, self).__init__()
-        self.output_attentions = output_attentions
+        self.output_attentions = config.output_attentions
-        self.output_hidden_states = output_hidden_states
+        self.output_hidden_states = config.output_hidden_states
-        layer = BertLayer(config, output_attentions=output_attentions)
+        layer = BertLayer(config)
        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.num_hidden_layers)])
    def forward(self, hidden_states, attention_mask, head_mask=None):
@@ -546,9 +543,6 @@ class BertPreTrainedModel(PreTrainedModel):
    load_tf_weights = load_tf_weights_in_bert
    base_model_prefix = "bert"
-    def __init__(self, *inputs, **kwargs):
-        super(BertPreTrainedModel, self).__init__(*inputs, **kwargs)
    def init_weights(self, module):
        """ Initialize the weights.
        """
@@ -612,19 +606,19 @@ class BertModel(BertPreTrainedModel):
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertModel, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
        self.embeddings = BertEmbeddings(config)
-        self.encoder = BertEncoder(config, output_attentions=output_attentions,
+        self.encoder = BertEncoder(config)
-                                           output_hidden_states=output_hidden_states)
        self.pooler = BertPooler(config)
        self.apply(self.init_weights)
-    def prune_heads(self, heads_to_prune):
+    def _prune_heads(self, heads_to_prune):
        """ Prunes heads of the model.
            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)
@@ -730,14 +724,12 @@ class BertForPreTraining(BertPreTrainedModel):
    masked_lm_logits_scores, seq_relationship_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForPreTraining, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
-        self.bert = BertModel(config, output_attentions=output_attentions,
+        self.bert = BertModel(config)
-                                      output_hidden_states=output_hidden_states)
        self.cls = BertPreTrainingHeads(config, self.bert.embeddings.word_embeddings.weight)
        self.apply(self.init_weights)
    def forward(self, input_ids, token_type_ids=None, attention_mask=None, masked_lm_labels=None,
@@ -809,13 +801,12 @@ class BertForMaskedLM(BertPreTrainedModel):
    masked_lm_logits_scores = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForMaskedLM, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
-        self.bert = BertModel(config, output_attentions=output_attentions )
+        self.bert = BertModel(config)
        self.cls = BertOnlyMLMHead(config, self.bert.embeddings.word_embeddings.weight)
        self.apply(self.init_weights)
    def forward(self, input_ids, token_type_ids=None, attention_mask=None, masked_lm_labels=None, head_mask=None):
@@ -880,12 +871,10 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
    seq_relationship_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForNextSentencePrediction, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
-        self.bert = BertModel(config, output_attentions=output_attentions)
+        self.bert = BertModel(config)
        self.cls = BertOnlyNSPHead(config)
        self.apply(self.init_weights)
@@ -954,15 +943,13 @@ class BertForSequenceClassification(BertPreTrainedModel):
    logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, num_labels=2, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForSequenceClassification, self).__init__(config)
-        self.output_attentions = output_attentions
+        self.num_labels = config.num_labels
-        self.output_hidden_states = output_hidden_states
-        self.num_labels = num_labels
-        self.bert = BertModel(config, output_attentions=output_attentions)
+        self.bert = BertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, num_labels)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
        self.apply(self.init_weights)
@@ -997,7 +984,6 @@ class BertForMultipleChoice(BertPreTrainedModel):
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `output_hidden_states`: If True, also output hidden states computed by the model at each layer. Default: False
-        `num_choices`: the number of classes for the classifier. Default = 2.
    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, num_choices, sequence_length]
@@ -1030,25 +1016,23 @@ class BertForMultipleChoice(BertPreTrainedModel):
    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)
-    num_choices = 2
+    model = BertForMultipleChoice(config)
-    model = BertForMultipleChoice(config, num_choices)
    logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, num_choices=2, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForMultipleChoice, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
-        self.num_choices = num_choices
-        self.bert = BertModel(config, output_attentions=output_attentions)
+        self.bert = BertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, 1)
        self.apply(self.init_weights)
    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None, head_mask=None):
+        """ Input shapes should be [bsz, num choices, seq length] """
+        num_choices = input_ids.shape[1]
        flat_input_ids = input_ids.view(-1, input_ids.size(-1))
        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
@@ -1057,7 +1041,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
-        reshaped_logits = logits.view(-1, self.num_choices)
+        reshaped_logits = logits.view(-1, num_choices)
        outputs = [reshaped_logits] + outputs[2:]  # add hidden states and attention if they are here
@@ -1118,15 +1102,13 @@ class BertForTokenClassification(BertPreTrainedModel):
    logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, num_labels=2, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForTokenClassification, self).__init__(config)
-        self.output_attentions = output_attentions
+        self.num_labels = config.num_labels
-        self.output_hidden_states = output_hidden_states
-        self.num_labels = num_labels
-        self.bert = BertModel(config, output_attentions=output_attentions)
+        self.bert = BertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, num_labels)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
        self.apply(self.init_weights)
@@ -1204,12 +1186,12 @@ class BertForQuestionAnswering(BertPreTrainedModel):
    start_logits, end_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(BertForQuestionAnswering, self).__init__(config)
-        self.output_attentions = output_attentions
+        self.num_labels = config.num_labels
-        self.output_hidden_states = output_hidden_states
-        self.bert = BertModel(config, output_attentions=output_attentions)
+        self.bert = BertModel(config)
-        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
        self.apply(self.init_weights)

--- a/pytorch_pretrained_bert/modeling_gpt2.py
+++ b/pytorch_pretrained_bert/modeling_gpt2.py
--- a/pytorch_pretrained_bert/modeling_openai.py
+++ b/pytorch_pretrained_bert/modeling_openai.py
@@ -147,7 +147,8 @@ class OpenAIGPTConfig(PretrainedConfig):
        attn_pdrop=0.1,
        layer_norm_epsilon=1e-5,
        initializer_range=0.02,
-        predict_special_tokens=True
+        predict_special_tokens=True,
+        **kwargs
    ):
        """Constructs OpenAIGPTConfig.
@@ -172,6 +173,8 @@ class OpenAIGPTConfig(PretrainedConfig):
                initializing all weight matrices.
            predict_special_tokens: should we predict special tokens (when the model has a LM head)
        """
+        super(OpenAIGPTConfig, self).__init__(**kwargs)
        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                        and isinstance(vocab_size_or_config_json_file, unicode)):
            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
@@ -205,7 +208,7 @@ class OpenAIGPTConfig(PretrainedConfig):
 class Attention(nn.Module):
-    def __init__(self, nx, n_ctx, config, scale=False, output_attentions=False, keep_multihead_output=False):
+    def __init__(self, nx, n_ctx, config, scale=False):
        super(Attention, self).__init__()
        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
        # [switch nx => n_state from Block to Attention to keep identical to TF implem]
@@ -215,9 +218,7 @@ class Attention(nn.Module):
        self.split_size = n_state
        self.scale = scale
-        self.output_attentions = output_attentions
+        self.output_attentions = config.output_attentions
-        self.keep_multihead_output = keep_multihead_output
-        self.multihead_output = None
        self.c_attn = Conv1D(n_state * 3, nx)
        self.c_proj = Conv1D(n_state, nx)
@@ -256,9 +257,10 @@ class Attention(nn.Module):
        if head_mask is not None:
            w = w * head_mask
+        outputs = [torch.matmul(w, v)]
        if self.output_attentions:
-            return w, torch.matmul(w, v)
+            outputs.append(w)
-        return torch.matmul(w, v)
+        return outputs
    def merge_heads(self, x):
        x = x.permute(0, 2, 1, 3).contiguous()
@@ -280,19 +282,15 @@ class Attention(nn.Module):
        key = self.split_heads(key, k=True)
        value = self.split_heads(value)
-        a = self._attn(query, key, value, head_mask)
+        attn_outputs = self._attn(query, key, value, head_mask)
-        if self.keep_multihead_output:
+        a = attn_outputs[0]
-            self.multihead_output = a
-            self.multihead_output.retain_grad()
-        if self.output_attentions:
-            attentions, a = a
        a = self.merge_heads(a)
        a = self.c_proj(a)
        a = self.resid_dropout(a)
-        if self.output_attentions:
-            return attentions, a
+        outputs = [a] + attn_outputs[1:]
-        return a
+        return outputs  # a, (attentions)
 class MLP(nn.Module):
@@ -311,25 +309,24 @@ class MLP(nn.Module):
 class Block(nn.Module):
-    def __init__(self, n_ctx, config, scale=False, output_attentions=False, keep_multihead_output=False):
+    def __init__(self, n_ctx, config, scale=False):
        super(Block, self).__init__()
        nx = config.n_embd
-        self.output_attentions = output_attentions
+        self.attn = Attention(nx, n_ctx, config, scale)
-        self.attn = Attention(nx, n_ctx, config, scale, output_attentions, keep_multihead_output)
        self.ln_1 = LayerNorm(nx, eps=config.layer_norm_epsilon)
        self.mlp = MLP(4 * nx, config)
        self.ln_2 = LayerNorm(nx, eps=config.layer_norm_epsilon)
    def forward(self, x, head_mask=None):
-        a = self.attn(x, head_mask=head_mask)
+        attn_outputs = self.attn(x, head_mask=head_mask)
-        if self.output_attentions:
+        a = attn_outputs[0]
-            attentions, a = a
        n = self.ln_1(x + a)
        m = self.mlp(n)
        h = self.ln_2(n + m)
-        if self.output_attentions:
-            return attentions, h
+        outputs = [h] + attn_outputs[1:]
-        return h
+        return outputs
 class OpenAIGPTLMHead(nn.Module):
@@ -368,10 +365,15 @@ class OpenAIGPTMultipleChoiceHead(nn.Module):
        nn.init.normal_(self.linear.weight, std=0.02)
        nn.init.normal_(self.linear.bias, 0)
-    def forward(self, hidden_states, mc_token_ids):
+    def forward(self, hidden_states, mc_token_ids=None):
-        # Classification logits
+        """ Extract classification token hidden state and project it using self.linear
-        # hidden_state (bsz, num_choices, seq_length, hidden_size)
+            hidden_state: hidden state of shape (bsz, num_choices, seq_length, hidden_size)
-        # mc_token_ids (bsz, num_choices)
+            mc_token_ids: [optional] index of the classification token, shape (bsz, num_choices)
+            if mc_token_ids=None we take the last token of the sequence as classification token
+        """
+        if mc_token_ids is None:
+            mc_token_ids = torch.full_like(hidden_states[:, :, :1, :], hidden_states.shape[2] - 1, dtype=torch.long)
+        else:
            mc_token_ids = mc_token_ids.unsqueeze(-1).unsqueeze(-1).expand(-1, -1, -1, hidden_states.size(-1))
        # (bsz, num_choices, 1, hidden_size)
        multiple_choice_h = hidden_states.gather(2, mc_token_ids).squeeze(2)
@@ -388,13 +390,9 @@ class OpenAIGPTPreTrainedModel(PreTrainedModel):
    """
    config_class = OpenAIGPTConfig
    pretrained_model_archive_map = PRETRAINED_MODEL_ARCHIVE_MAP
-    pretrained_config_archive_map = PRETRAINED_CONFIG_ARCHIVE_MAP
    load_tf_weights = load_tf_weights_in_openai_gpt
    base_model_prefix = "transformer"
-    def __init__(self, *inputs, **kwargs):
-        super(OpenAIGPTPreTrainedModel, self).__init__(*inputs, **kwargs)
    def init_weights(self, module):
        """ Initialize the weights.
        """
@@ -495,14 +493,15 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
    ```
    """
-    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
+    def __init__(self, config):
        super(OpenAIGPTModel, self).__init__(config)
-        self.output_attentions = output_attentions
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
        self.tokens_embed = nn.Embedding(config.total_tokens_embeddings, config.n_embd)
        self.positions_embed = nn.Embedding(config.n_positions, config.n_embd)
        self.drop = nn.Dropout(config.embd_pdrop)
-        block = Block(config.n_ctx, config, scale=True, output_attentions=output_attentions,
+        block = Block(config.n_ctx, config, scale=True)
-                                                        keep_multihead_output=keep_multihead_output)
        self.h = nn.ModuleList([copy.deepcopy(block) for _ in range(config.n_layer)])
        self.apply(self.init_weights)
@@ -521,19 +520,13 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
        # Copy word embeddings from the previous weights
        self.tokens_embed.weight.data[:self.config.vocab_size, :] = old_embed.weight.data[:self.config.vocab_size, :]
-    def prune_heads(self, heads_to_prune):
+    def _prune_heads(self, heads_to_prune):
        """ Prunes heads of the model.
            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
        """
        for layer, heads in heads_to_prune.items():
            self.h[layer].attn.prune_heads(heads)
-    def get_multihead_outputs(self):
-        """ Gather all multi-head outputs.
-            Return: list (layers) of multihead module outputs with gradients
-        """
-        return [h.attn.multihead_output for h in self.h]
    def forward(self, input_ids, position_ids=None, token_type_ids=None, head_mask=None):
        if position_ids is None:
            # This was used when we had a single embedding matrice from position and token embeddings
@@ -574,19 +567,26 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
        output_shape = input_shape + (hidden_states.size(-1),)
        all_attentions = []
-        all_hidden_states = [hidden_states.view(*output_shape)]
+        all_hidden_states = []
        for i, block in enumerate(self.h):
+            if self.output_hidden_states:
+                all_hidden_states.append(hidden_states.view(*output_shape))
            outputs = block(hidden_states, head_mask[i])
+            hidden_states = outputs[0]
            if self.output_attentions:
-                attentions, hidden_states = outputs
+                all_attentions.append(outputs[1])
-                all_attentions.append(attentions)
-            else:
+        # Add last layer
-                hidden_states = outputs
+        if self.output_hidden_states:
            all_hidden_states.append(hidden_states.view(*output_shape))
+        outputs = [hidden_states.view(*output_shape)]
+        if self.output_hidden_states:
+            outputs.append(all_hidden_states)
        if self.output_attentions:
-            return all_attentions, all_hidden_states
+            outputs.append(all_attentions)
-        return all_hidden_states
+        return outputs  # last hidden state, (all hidden states), (all attentions)
 class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
@@ -650,10 +650,9 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
    ```
    """
-    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
+    def __init__(self, config):
        super(OpenAIGPTLMHeadModel, self).__init__(config)
-        self.transformer = OpenAIGPTModel(config, output_attentions=output_attentions,
+        self.transformer = OpenAIGPTModel(config)
-                                             keep_multihead_output=keep_multihead_output)
        self.lm_head = OpenAIGPTLMHead(self.transformer.tokens_embed.weight, config)
        self.apply(self.init_weights)
@@ -666,12 +665,11 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
        self.lm_head.set_embeddings_weights(self.transformer.tokens_embed.weight, predict_special_tokens=predict_special_tokens)
    def forward(self, input_ids, position_ids=None, token_type_ids=None, lm_labels=None, head_mask=None):
-        hidden_states = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
+        transformer_outputs = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
-        if self.transformer.output_attentions:
+        hidden_states = transformer_outputs[0]
-            all_attentions, hidden_states = hidden_states
-        hidden_states = hidden_states[-1]
        lm_logits = self.lm_head(hidden_states)
+        outputs = [lm_logits] + transformer_outputs[1:]
        if lm_labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
@@ -680,10 +678,9 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                            shift_labels.view(-1))
-            return loss
+            outputs = [loss] + outputs
-        if self.transformer.output_attentions:
-            return all_attentions, lm_logits
+        return outputs  # (loss), lm_logits, (all hidden states), (all attentions)
-        return lm_logits
 class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
@@ -752,10 +749,9 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
    ```
    """
-    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
+    def __init__(self, config):
        super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
-        self.transformer = OpenAIGPTModel(config, output_attentions=output_attentions,
+        self.transformer = OpenAIGPTModel(config)
-                                             keep_multihead_output=keep_multihead_output)
        self.lm_head = OpenAIGPTLMHead(self.transformer.tokens_embed.weight, config)
        self.multiple_choice_head = OpenAIGPTMultipleChoiceHead(config)
        self.apply(self.init_weights)
@@ -768,26 +764,26 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
        self.transformer.set_num_special_tokens(num_special_tokens)
        self.lm_head.set_embeddings_weights(self.transformer.tokens_embed.weight, predict_special_tokens=predict_special_tokens)
-    def forward(self, input_ids, mc_token_ids, lm_labels=None, mc_labels=None, token_type_ids=None,
+    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):
-        hidden_states = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
+        transformer_outputs = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
-        if self.transformer.output_attentions:
+        hidden_states = transformer_outputs[0]
-            all_attentions, hidden_states = hidden_states
-        hidden_states = hidden_states[-1]
        lm_logits = self.lm_head(hidden_states)
        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids)
-        losses = []
+        outputs = [lm_logits, mc_logits] + transformer_outputs[1:]
+        if mc_labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(mc_logits.view(-1, mc_logits.size(-1)),
+                            mc_labels.view(-1))
+            outputs = [loss] + outputs
        if lm_labels is not None:
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = lm_labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss(ignore_index=-1)
-            losses.append(loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)))
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
-        if mc_labels is not None:
+                            shift_labels.view(-1))
-            loss_fct = CrossEntropyLoss()
+            outputs = [loss] + outputs
-            losses.append(loss_fct(mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1)))
-        if losses:
+        return outputs  # (lm loss), (mc loss), lm logits, mc logits, (all hidden_states), (attentions)
-            return losses
-        if self.transformer.output_attentions:
-            return all_attentions, lm_logits, mc_logits
-        return lm_logits, mc_logits
--- a/pytorch_pretrained_bert/modeling_transfo_xl.py
+++ b/pytorch_pretrained_bert/modeling_transfo_xl.py
--- a/pytorch_pretrained_bert/modeling_xlm.py
+++ b/pytorch_pretrained_bert/modeling_xlm.py
--- a/pytorch_pretrained_bert/modeling_xlnet.py
+++ b/pytorch_pretrained_bert/modeling_xlnet.py
@@ -198,7 +198,7 @@ class XLNetConfig(PretrainedConfig):
    pretrained_config_archive_map = PRETRAINED_CONFIG_ARCHIVE_MAP
    def __init__(self,
-                 vocab_size_or_config_json_file,
+                 vocab_size_or_config_json_file=32000,
                 d_model=1024,
                 n_layer=24,
                 n_head=16,
@@ -221,7 +221,12 @@ class XLNetConfig(PretrainedConfig):
                 bi_data=False,
                 clamp_len=-1,
                 same_length=False,
-                 finetuning_task=None):
+                 finetuning_task=None,
+                 num_labels=2,
+                 summary_type="last",
+                 use_proj=True,
+                 **kwargs):
        """Constructs XLNetConfig.
        Args:
@@ -265,6 +270,8 @@ class XLNetConfig(PretrainedConfig):
            same_length: bool, whether to use the same attention length for each token.
            finetuning_task: name of the glue task on which the model was fine-tuned if any
        """
+        super(XLNetConfig, self).__init__(**kwargs)
        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                        and isinstance(vocab_size_or_config_json_file, unicode)):
            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
@@ -297,7 +304,11 @@ class XLNetConfig(PretrainedConfig):
            self.bi_data = bi_data
            self.clamp_len = clamp_len
            self.same_length = same_length
            self.finetuning_task = finetuning_task
+            self.num_labels = num_labels
+            self.summary_type = summary_type
+            self.use_proj = use_proj
        else:
            raise ValueError("First argument must be either a vocabulary size (int)"
                             "or the path to a pretrained model config file (str)")
@@ -323,9 +334,10 @@ except ImportError:
            return self.weight * x + self.bias
 class XLNetRelativeAttention(nn.Module):
-    def __init__(self, config, output_attentions=False):
+    def __init__(self, config):
        super(XLNetRelativeAttention, self).__init__()
-        self.output_attentions = output_attentions
+        self.output_attentions = config.output_attentions
        if config.d_model % config.n_head != 0:
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
@@ -533,10 +545,9 @@ class XLNetFeedForward(nn.Module):
        return output
 class XLNetLayer(nn.Module):
-    def __init__(self, config, output_attentions=False, ):
+    def __init__(self, config):
        super(XLNetLayer, self).__init__()
-        self.output_attentions = output_attentions
+        self.rel_attn = XLNetRelativeAttention(config)
-        self.rel_attn = XLNetRelativeAttention(config, output_attentions=output_attentions)
        self.ff = XLNetFeedForward(config)
        self.dropout = nn.Dropout(config.dropout)
@@ -562,7 +573,6 @@ class XLNetPreTrainedModel(PreTrainedModel):
    """
    config_class = XLNetConfig
    pretrained_model_archive_map = PRETRAINED_MODEL_ARCHIVE_MAP
-    pretrained_config_archive_map = PRETRAINED_CONFIG_ARCHIVE_MAP
    load_tf_weights = load_tf_weights_in_xlnet
    base_model_prefix = "transformer"
@@ -589,10 +599,10 @@ class XLNetPreTrainedModel(PreTrainedModel):
 class XLNetModel(XLNetPreTrainedModel):
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(XLNetModel, self).__init__(config)
-        self.output_attentions = output_attentions
+        self.output_attentions = config.output_attentions
-        self.output_hidden_states = output_hidden_states
+        self.output_hidden_states = config.output_hidden_states
        self.mem_len = config.mem_len
        self.reuse_len = config.reuse_len
@@ -601,25 +611,17 @@ class XLNetModel(XLNetPreTrainedModel):
        self.attn_type = config.attn_type
        self.bi_data = config.bi_data
        self.clamp_len = config.clamp_len
+        self.n_layer = config.n_layer
        self.word_embedding = nn.Embedding(config.n_token, config.d_model)
        self.mask_emb = nn.Parameter(torch.Tensor(1, 1, config.d_model))
-        layer = XLNetLayer(config, output_attentions=output_attentions)
+        layer = XLNetLayer(config)
        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.n_layer)])
        self.dropout = nn.Dropout(config.dropout)
-    def prune_heads(self, heads_to_prune):
+    def _prune_heads(self, heads_to_prune):
-        """ Prunes heads of the model.
+        logger.info("Head pruning is not implemented for XLNet")
-            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        pass
-        """
-        for layer, heads in heads_to_prune.items():
-            self.layer[layer].attention.prune_heads(heads)
-    def get_multihead_outputs(self):
-        """ Gather all multi-head outputs.
-            Return: list (layers) of multihead module outputs with gradients
-        """
-        return [layer.attention.self.multihead_output for layer in self.layer]
    def create_mask(self, qlen, mlen):
        """ create causal attention mask.
@@ -708,11 +710,11 @@ class XLNetModel(XLNetPreTrainedModel):
        pos_emb = pos_emb.to(next(self.parameters()))
        return pos_emb
-    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
+    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, head_mask=None):
        """
        Args:
-            inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
+            input_ids: int32 Tensor in shape [bsz, len], the input token IDs.
            token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
            input_mask: [optional] float32 Tensor in shape [bsz, len], the input mask.
                0 for real tokens and 1 for padding.
@@ -751,7 +753,7 @@ class XLNetModel(XLNetPreTrainedModel):
        # the original code for XLNet uses shapes [len, bsz] with the batch dimension at the end
        # but we want a unified interface in the library with the batch size on the first dimension
        # so we move here the first dimension (batch) to the end
-        inp_k = inp_k.transpose(0, 1).contiguous()
+        input_ids = input_ids.transpose(0, 1).contiguous()
        token_type_ids = token_type_ids.transpose(0, 1).contiguous() if token_type_ids is not None else None
        input_mask = input_mask.transpose(0, 1).contiguous() if input_mask is not None else None
        attention_mask = attention_mask.transpose(0, 1).contiguous() if attention_mask is not None else None
@@ -759,7 +761,7 @@ class XLNetModel(XLNetPreTrainedModel):
        target_mapping = target_mapping.permute(1, 2, 0).contiguous() if target_mapping is not None else None
        inp_q = inp_q.transpose(0, 1).contiguous() if inp_q is not None else None
-        qlen, bsz = inp_k.shape[0], inp_k.shape[1]
+        qlen, bsz = input_ids.shape[0], input_ids.shape[1]
        mlen = mems[0].shape[0] if mems is not None else 0
        klen = mlen + qlen
@@ -810,7 +812,7 @@ class XLNetModel(XLNetPreTrainedModel):
            non_tgt_mask = None
        ##### Word embeddings and prepare h & g hidden states
-        word_emb_k = self.word_embedding(inp_k)
+        word_emb_k = self.word_embedding(input_ids)
        output_h = self.dropout(word_emb_k)
        if inp_q is not None:
            if target_mapping is not None:
@@ -838,20 +840,20 @@ class XLNetModel(XLNetPreTrainedModel):
        pos_emb = self.relative_positional_encoding(qlen, klen, bsz=bsz)
        pos_emb = self.dropout(pos_emb)
-        ##### Head mask if needed (for bertology/pruning)
+        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
-        # input head_mask has shape [num_heads] or [n_layer x num_heads]
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] (a head_mask for each layer)
-        # and head_mask is converted to shape [n_layer x batch x num_heads x seq_length x seq_length]
+        # and head_mask is converted to shape [num_hidden_layers x qlen x klen x bsz x n_head]
        if head_mask is not None:
            if head_mask.dim() == 1:
-                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(0).unsqueeze(0)
-                head_mask = head_mask.expand(self.config.n_layer, -1, -1, -1, -1)
+                head_mask = head_mask.expand(self.n_layer, -1, -1, -1, -1)
            elif head_mask.dim() == 2:
-                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+                head_mask = head_mask.unsqueeze(1).unsqueeze(1).unsqueeze(1)
            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
        else:
-            head_mask = [None] * self.config.n_layer
+            head_mask = [None] * self.n_layer
        new_mems = []
        if mems is None:
@@ -870,7 +872,7 @@ class XLNetModel(XLNetPreTrainedModel):
                                   head_mask=head_mask[i])
            output_h, output_g = outputs[:2]
            if self.output_attentions:
-                attentions.append(outputs[2:])
+                attentions.append(outputs[2])
        # Add last hidden state
        if self.output_hidden_states:
@@ -887,6 +889,7 @@ class XLNetModel(XLNetPreTrainedModel):
                hidden_states = [hs.permute(1, 0, 2).contiguous() for hs in hidden_states]
            outputs.append(hidden_states)
        if self.output_attentions:
+            attentions = list(t.permute(2, 3, 0, 1).contiguous() for t in attentions)
            outputs.append(attentions)
        return outputs  # outputs, new_mems, (hidden_states), (attentions)
@@ -902,7 +905,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
            This can be used to compute head importance metrics. Default: False
    Inputs:
-        inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
+        input_ids: int32 Tensor in shape [bsz, len], the input token IDs.
        token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
        input_mask: [optional] float32 Tensor in shape [bsz, len], the input mask.
            0 for real tokens and 1 for padding.
@@ -953,16 +956,12 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(XLNetLMHeadModel, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
        self.attn_type = config.attn_type
        self.same_length = config.same_length
-        self.transformer = XLNetModel(config, output_attentions=output_attentions,
+        self.transformer = XLNetModel(config)
-                                              output_hidden_states=output_hidden_states)
        self.lm_loss = nn.Linear(config.d_model, config.n_token, bias=True)
        # Tie weights
@@ -975,12 +974,12 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
        """
        self.lm_loss.weight = self.transformer.word_embedding.weight
-    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
+    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,
                labels=None, head_mask=None):
        """
        Args:
-            inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
+            input_ids: int32 Tensor in shape [bsz, len], the input token IDs.
            token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
            input_mask: float32 Tensor in shape [bsz, len], the input mask.
                0 for real tokens and 1 for padding.
@@ -1008,7 +1007,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
            summary_type: str, "last", "first", "mean", or "attn". The method
                to pool the input to get a vector representation.
        """
-        transformer_outputs = self.transformer(inp_k, token_type_ids, input_mask, attention_mask,
+        transformer_outputs = self.transformer(input_ids, token_type_ids, input_mask, attention_mask,
                                               mems, perm_mask, target_mapping, inp_q, head_mask)
        logits = self.lm_loss(transformer_outputs[0])
@@ -1025,14 +1024,14 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 class XLNetSequenceSummary(nn.Module):
-    def __init__(self, config, summary_type="last", use_proj=True):
+    def __init__(self, config):
        super(XLNetSequenceSummary, self).__init__()
-        self.summary_type = summary_type
+        self.summary_type = config.summary_type
-        if use_proj:
+        if config.use_proj:
            self.summary = nn.Linear(config.d_model, config.d_model)
        else:
            self.summary = None
-        if summary_type == 'attn':
+        if config.summary_type == 'attn':
            # We should use a standard multi-head attention module with absolute positional embedding for that.
            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
@@ -1069,7 +1068,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
            to pool the input to get a vector representation. Default: last
    Inputs:
-        inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
+        input_ids: int32 Tensor in shape [bsz, len], the input token IDs.
        token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
        input_mask: float32 Tensor in shape [bsz, len], the input mask.
            0 for real tokens and 1 for padding.
@@ -1121,30 +1120,21 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, summary_type="last", use_proj=True, num_labels=2,
+    def __init__(self, config):
-                 output_attentions=False, output_hidden_states=False):
        super(XLNetForSequenceClassification, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
-        self.attn_type = config.attn_type
-        self.same_length = config.same_length
-        self.summary_type = summary_type
-        self.num_labels = num_labels
-        self.transformer = XLNetModel(config, output_attentions=output_attentions,
+        self.transformer = XLNetModel(config)
-                                              output_hidden_states=output_hidden_states)
+        self.sequence_summary = XLNetSequenceSummary(config)
+        self.logits_proj = nn.Linear(config.d_model, config.num_labels)
-        self.sequence_summary = XLNetSequenceSummary(config, summary_type=summary_type, use_proj=use_proj)
-        self.logits_proj = nn.Linear(config.d_model, num_labels)
        self.apply(self.init_weights)
-    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
+    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,
                labels=None, head_mask=None):
        """
        Args:
-            inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
+            input_ids: int32 Tensor in shape [bsz, len], the input token IDs.
            token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
            input_mask: float32 Tensor in shape [bsz, len], the input mask.
                0 for real tokens and 1 for padding.
@@ -1169,7 +1159,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
                Only used during pretraining for two-stream attention.
                Set to None during finetuning.
        """
-        transformer_outputs = self.transformer(inp_k, token_type_ids, input_mask, attention_mask,
+        transformer_outputs = self.transformer(input_ids, token_type_ids, input_mask, attention_mask,
                                               mems, perm_mask, target_mapping, inp_q, head_mask)
        output = transformer_outputs[0]
@@ -1247,20 +1237,18 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
    start_logits, end_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """
-    def __init__(self, config, output_attentions=False, output_hidden_states=False):
+    def __init__(self, config):
        super(XLNetForQuestionAnswering, self).__init__(config)
-        self.output_attentions = output_attentions
-        self.output_hidden_states = output_hidden_states
-        self.transformer = XLNetModel(config, output_attentions=output_attentions,
+        self.transformer = XLNetModel(config)
-                                      output_hidden_states=output_hidden_states)
+        self.qa_outputs = nn.Linear(config.d_model, config.num_labels)
-        self.qa_outputs = nn.Linear(config.hidden_size, 2)
        self.apply(self.init_weights)
-    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
+    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,
                start_positions=None, end_positions=None, head_mask=None):
-        transformer_outputs = self.transformer(inp_k, token_type_ids, input_mask, attention_mask,
+        transformer_outputs = self.transformer(input_ids, token_type_ids, input_mask, attention_mask,
                                            mems, perm_mask, target_mapping, inp_q, head_mask)
        logits = self.qa_outputs(transformer_outputs[0])

--- a/pytorch_pretrained_bert/tests/__init__.py
+++ b/pytorch_pretrained_bert/tests/__init__.py
--- a/tests/conftest.py
+++ b/tests/conftest.py
--- a/samples/input.txt
+++ b/samples/input.txt
--- a/samples/sample_text.txt
+++ b/samples/sample_text.txt
--- a/samples/test_sentencepiece.model
+++ b/samples/test_sentencepiece.model
--- a/pytorch_pretrained_bert/tests/model_tests_commons.py
+++ b/pytorch_pretrained_bert/tests/model_tests_commons.py
+# coding=utf-8
+# Copyright 2019 HuggingFace Inc.
+#
+# 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 os
+import shutil
+import json
+import random
+import torch
+def create_and_check_for_headmasking(tester, model_classes, config, inputs_dict):
+    for model_class in model_classes:
+        config.output_hidden_states = True
+        model = model_class(config=config)
+        model.eval()
+        head_mask = torch.zeros(tester.num_hidden_layers, tester.num_attention_heads)
+        # Set that after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior) 
+        head_mask.requires_grad_(requires_grad=True)
+        outputs = model(**inputs_dict, head_mask=head_mask)
+        # Compute some gradients
+        output = sum(t.sum() for t in outputs[0])
+        output = output.sum()
+        output.backward()
+        multihead_outputs = head_mask.grad
+        tester.parent.assertEqual(len(multihead_outputs), tester.num_hidden_layers)
+        # self.parent.assertListEqual(
+        #     list(multihead_outputs[0].size()),
+        #     [self.batch_size, self.num_attention_heads,
+        #      self.seq_length, self.hidden_size // self.num_attention_heads])
+        # self.parent.assertEqual(
+        #     len(multihead_outputs[0][:, 1:(self.num_attention_heads-1), :, :].nonzero()),
+        #     0)
+        # self.parent.assertEqual(
+        #     len(multihead_outputs[0][:, 0, :, :].nonzero()),
+        #     self.batch_size * self.seq_length * self.hidden_size // self.num_attention_heads)
+        # self.parent.assertEqual(
+        #     len(multihead_outputs[0][:, self.num_attention_heads-1, :, :].nonzero()),
+        #     self.batch_size * self.seq_length * self.hidden_size // self.num_attention_heads)
+        # self.parent.assertListEqual(
+        #     list(multihead_outputs[1].size()),
+        #     [self.batch_size, self.num_attention_heads,
+        #      self.seq_length, self.hidden_size // self.num_attention_heads])
+        # self.parent.assertEqual(
+        #     len(multihead_outputs[1].nonzero()),
+        #     multihead_outputs[1].numel())
+        # self.parent.assertListEqual(
+        #     list(multihead_outputs[-1].size()),
+        #     [self.batch_size, self.num_attention_heads,
+        #      self.seq_length, self.hidden_size // self.num_attention_heads])
+        # self.parent.assertEqual(
+        #     len(multihead_outputs[-1][:, 1:, :, :].nonzero()),
+        #     0)
+        # self.parent.assertEqual(
+        #     len(multihead_outputs[-1][:, 0, :, :].nonzero()),
+        #     self.batch_size * self.seq_length * self.hidden_size // self.num_attention_heads)
+def create_and_check_for_head_pruning(tester, model_classes, config, inputs_dict):
+    for model_class in model_classes:
+        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)
+        # output = sum(t.sum() for t in outputs[0])
+        # output = output.sum()
+        # output.backward()
+        # multihead_outputs = bert_model.get_multihead_outputs()
+        # self.parent.assertEqual(len(multihead_outputs), self.num_hidden_layers)
+        # self.parent.assertListEqual(
+        #     list(multihead_outputs[0].size()),
+        #     [self.batch_size, 1,
+        #      self.seq_length, self.hidden_size // self.num_attention_heads])
+        # self.parent.assertListEqual(
+        #     list(multihead_outputs[1].size()),
+        #     [self.batch_size, self.num_attention_heads,
+        #      self.seq_length, self.hidden_size // self.num_attention_heads])
+        # self.parent.assertListEqual(
+        #     list(multihead_outputs[-1].size()),
+        #     [self.batch_size, self.num_attention_heads-1,
+        #      self.seq_length, self.hidden_size // self.num_attention_heads])
+def create_and_check_for_attentions(tester, model_classes, config, inputs_dict):
+    for model_class in model_classes:
+        config.output_attentions = True
+        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.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]
+        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])
+def create_and_check_for_hidden_states(tester, model_classes, config, inputs_dict):
+    for model_class in model_classes:
+        config.output_hidden_states = True
+        config.output_attentions = False
+        model = model_class(config)
+        model.eval()
+        outputs = model(**inputs_dict)
+        hidden_states = outputs[-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):
+    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_head_pruning(tester, tester.all_model_classes, config, inputs_dict)
+    create_and_check_for_hidden_states(tester, tester.all_model_classes, config, inputs_dict)
+def ids_tensor(shape, vocab_size, rng=None, name=None):
+    """Creates a random int32 tensor of the shape within the vocab size."""
+    if rng is None:
+        rng = random.Random()
+    total_dims = 1
+    for dim in shape:
+        total_dims *= dim
+    values = []
+    for _ in range(total_dims):
+        values.append(rng.randint(0, vocab_size - 1))
+    return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
+class ConfigTester(object):
+    def __init__(self, parent, config_class=None, **kwargs):
+        self.parent = parent
+        self.config_class = config_class
+        self.inputs_dict = kwargs
+    def create_and_test_config_to_json_string(self):
+        config = self.config_class(**self.inputs_dict)
+        obj = json.loads(config.to_json_string())
+        for key, value in self.inputs_dict.items():
+            self.parent.assertEqual(obj[key], value)
+    def create_and_test_config_to_json_file(self):
+        config_first = self.config_class(**self.inputs_dict)
+        json_file_path = "/tmp/config.json"
+        config_first.to_json_file(json_file_path)
+        config_second = self.config_class.from_json_file(json_file_path)
+        os.remove(json_file_path)
+        self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
+    def run_common_tests(self):
+        self.create_and_test_config_to_json_string()
+        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_special=1,
+                    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_special = n_special
+        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 + self.n_special
+        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_special=self.n_special,
+            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)
+        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.n_special + 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.n_special + 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_pretrained_bert_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):
+        config_and_inputs = self.prepare_config_and_inputs()
+        self.create_and_check_model_from_pretrained(*config_and_inputs)
--- a/pytorch_pretrained_bert/tests/model_utils_test.py
+++ b/pytorch_pretrained_bert/tests/model_utils_test.py
+# coding=utf-8
+# Copyright 2018 HuggingFace Inc..
+#
+# 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 os
+import unittest
+import json
+import random
+import shutil
+import pytest
+import torch
+from pytorch_pretrained_bert import PretrainedConfig, PreTrainedModel
+from pytorch_pretrained_bert.modeling import BertModel, BertConfig, PRETRAINED_MODEL_ARCHIVE_MAP, PRETRAINED_CONFIG_ARCHIVE_MAP
+class ModelUtilsTest(unittest.TestCase):
+    def test_model_from_pretrained(self):
+        for model_name in list(PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            config = BertConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, PretrainedConfig)
+            model = BertModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, PreTrainedModel)
+            config = BertConfig.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
+            model = BertModel.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
+            self.assertEqual(model.config.output_attentions, True)
+            self.assertEqual(model.config.output_hidden_states, True)
+            self.assertEqual(model.config, config)
+if __name__ == "__main__":
+    unittest.main()
--- a/pytorch_pretrained_bert/tests/modeling_gpt2_test.py
+++ b/pytorch_pretrained_bert/tests/modeling_gpt2_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 os
+import unittest
+import json
+import random
+import shutil
+import pytest
+import torch
+from pytorch_pretrained_bert import (GPT2Config, GPT2Model,
+                                     GPT2LMHeadModel, GPT2DoubleHeadsModel)
+from .model_tests_commons import (create_and_check_for_attentions, create_and_check_for_head_pruning,
+                                  create_and_check_for_headmasking, create_and_check_for_hidden_states,
+                                  ConfigTester, GPTModelTester)
+class GPT2ModelTest(unittest.TestCase):
+    def test_config(self):
+        config_tester = ConfigTester(self, config_class=GPT2Config, n_embd=37)
+        config_tester.run_common_tests()
+    def test_model(self):
+        model_tester = 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,
+                                           lm_head_model_class=GPT2LMHeadModel,
+                                           double_head_model_class=GPT2DoubleHeadsModel)
+        model_tester.run_slow_tests()
+if __name__ == "__main__":
+    unittest.main()
--- a/pytorch_pretrained_bert/tests/modeling_openai_test.py
+++ b/pytorch_pretrained_bert/tests/modeling_openai_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 os
+import unittest
+import json
+import random
+import shutil
+import pytest
+import torch
+from pytorch_pretrained_bert import (OpenAIGPTConfig, OpenAIGPTModel,
+                                     OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel)
+from .model_tests_commons import (create_and_check_for_attentions, create_and_check_for_head_pruning,
+                                  create_and_check_for_headmasking, create_and_check_for_hidden_states,
+                                  ConfigTester, GPTModelTester)
+class OpenAIModelTest(unittest.TestCase):
+    def test_config(self):
+        config_tester = ConfigTester(self, config_class=OpenAIGPTConfig, n_embd=37)
+        config_tester.run_common_tests()
+    def test_model(self):
+        model_tester = 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,
+                                           lm_head_model_class=OpenAIGPTLMHeadModel,
+                                           double_head_model_class=OpenAIGPTDoubleHeadsModel)
+        model_tester.run_slow_tests()
+if __name__ == "__main__":
+    unittest.main()
--- a/tests/modeling_test.py
+++ b/tests/modeling_test.py
--- a/tests/modeling_transfo_xl_test.py
+++ b/tests/modeling_transfo_xl_test.py
@@ -28,6 +28,8 @@ import torch
 from pytorch_pretrained_bert import (TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel)
 from pytorch_pretrained_bert.modeling_transfo_xl import PRETRAINED_MODEL_ARCHIVE_MAP
+from .model_tests_commons import ConfigTester, create_and_check_commons, ids_tensor
 class TransfoXLModelTest(unittest.TestCase):
    class TransfoXLModelTester(object):
@@ -41,54 +43,58 @@ class TransfoXLModelTest(unittest.TestCase):
                     use_labels=True,
                     vocab_size=99,
                     cutoffs=[10, 50, 80],
-                     d_model=32,
+                     hidden_size=32,
                     d_embed=32,
-                     n_head=4,
+                     num_attention_heads=4,
                     d_head=8,
                     d_inner=128,
                     div_val=2,
-                     n_layer=5,
+                     num_hidden_layers=5,
                     scope=None,
-                     seed=1):
+                     seed=1,
+                     all_model_classes=(TransfoXLModel, TransfoXLLMHeadModel),
+                     ):
            self.parent = parent
            self.batch_size = batch_size
            self.seq_length = seq_length
            self.mem_len = mem_len
+            self.key_len = seq_length + mem_len
            self.clamp_len = clamp_len
            self.is_training = is_training
            self.use_labels = use_labels
            self.vocab_size = vocab_size
            self.cutoffs = cutoffs
-            self.d_model = d_model
+            self.hidden_size = hidden_size
            self.d_embed = d_embed
-            self.n_head = n_head
+            self.num_attention_heads = num_attention_heads
            self.d_head = d_head
            self.d_inner = d_inner
            self.div_val = div_val
-            self.n_layer = n_layer
+            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 = TransfoXLModelTest.ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+            input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
-            input_ids_2 = TransfoXLModelTest.ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+            input_ids_2 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
            lm_labels = None
            if self.use_labels:
-                lm_labels = TransfoXLModelTest.ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+                lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
            config = TransfoXLConfig(
                vocab_size_or_config_json_file=self.vocab_size,
                mem_len=self.mem_len,
                clamp_len=self.clamp_len,
                cutoffs=self.cutoffs,
-                d_model=self.d_model,
+                d_model=self.hidden_size,
                d_embed=self.d_embed,
-                n_head=self.n_head,
+                n_head=self.num_attention_heads,
                d_head=self.d_head,
                d_inner=self.d_inner,
                div_val=self.div_val,
-                n_layer=self.n_layer)
+                n_layer=self.num_hidden_layers)
            return (config, input_ids_1, input_ids_2, lm_labels)
@@ -113,37 +119,34 @@ class TransfoXLModelTest(unittest.TestCase):
        def check_transfo_xl_model_output(self, result):
            self.parent.assertListEqual(
                list(result["hidden_states_1"].size()),
-                [self.batch_size, self.seq_length, self.d_model])
+                [self.batch_size, self.seq_length, self.hidden_size])
            self.parent.assertListEqual(
                list(result["hidden_states_2"].size()),
-                [self.batch_size, self.seq_length, self.d_model])
+                [self.batch_size, self.seq_length, self.hidden_size])
            self.parent.assertListEqual(
                list(list(mem.size()) for mem in result["mems_1"]),
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
+                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
            self.parent.assertListEqual(
                list(list(mem.size()) for mem in result["mems_2"]),
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
+                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
        def create_transfo_xl_lm_head(self, config, input_ids_1, input_ids_2, lm_labels):
            model = TransfoXLLMHeadModel(config)
            model.eval()
-            loss_1, mems_1a = model(input_ids_1, labels=lm_labels)
+            lm_logits_1, mems_1 = model(input_ids_1)
-            lm_logits_1, mems_1b = model(input_ids_1)
+            loss_1, _, mems_1 = model(input_ids_1, labels=lm_labels)
+            lm_logits_2, mems_2 = model(input_ids_2, mems=mems_1)
-            loss_2, mems_2a = model(input_ids_2, labels=lm_labels, mems=mems_1a)
+            loss_2, _, mems_2 = model(input_ids_2, labels=lm_labels, mems=mems_1)
-            lm_logits_2, mems_2b = model(input_ids_2, mems=mems_1b)
            outputs = {
                "loss_1": loss_1,
-                "mems_1a": mems_1a,
+                "mems_1": mems_1,
                "lm_logits_1": lm_logits_1,
-                "mems_1b": mems_1b,
                "loss_2": loss_2,
-                "mems_2a": mems_2a,
+                "mems_2": mems_2,
                "lm_logits_2": lm_logits_2,
-                "mems_2b": mems_2b,
            }
            return outputs
@@ -155,14 +158,8 @@ class TransfoXLModelTest(unittest.TestCase):
                list(result["lm_logits_1"].size()),
                [self.batch_size, self.seq_length, self.vocab_size])
            self.parent.assertListEqual(
-                list(list(mem.size()) for mem in result["mems_1a"]),
+                list(list(mem.size()) for mem in result["mems_1"]),
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
+                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
-            self.parent.assertListEqual(
-                list(list(mem.size()) for mem in result["mems_1b"]),
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
-            self.parent.assertListEqual(
-                list(mem[~torch.isnan(mem)].sum() for mem in result["mems_1a"]),
-                list(mem[~torch.isnan(mem)].sum() for mem in result["mems_1b"]))
            self.parent.assertListEqual(
                list(result["loss_2"].size()),
@@ -171,31 +168,19 @@ class TransfoXLModelTest(unittest.TestCase):
                list(result["lm_logits_2"].size()),
                [self.batch_size, self.seq_length, self.vocab_size])
            self.parent.assertListEqual(
-                list(list(mem.size()) for mem in result["mems_2a"]),
+                list(list(mem.size()) for mem in result["mems_2"]),
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
+                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
-            self.parent.assertListEqual(
-                list(list(mem.size()) for mem in result["mems_2b"]),
+        def create_and_check_transfo_xl_commons(self, config, input_ids_1, input_ids_2, lm_labels):
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
+            inputs_dict = {'input_ids': input_ids_1}
-            self.parent.assertListEqual(
+            create_and_check_commons(self, config, inputs_dict)
-                list(mem[~torch.isnan(mem)].sum() for mem in result["mems_2a"]),
-                list(mem[~torch.isnan(mem)].sum() for mem in result["mems_2b"]))
    def test_default(self):
        self.run_tester(TransfoXLModelTest.TransfoXLModelTester(self))
-    def test_config_to_json_string(self):
+    def test_config(self):
-        config = TransfoXLConfig(vocab_size_or_config_json_file=96, d_embed=37)
+        config_tester = ConfigTester(self, config_class=TransfoXLConfig, d_embed=37)
-        obj = json.loads(config.to_json_string())
+        config_tester.run_common_tests()
-        self.assertEqual(obj["n_token"], 96)
-        self.assertEqual(obj["d_embed"], 37)
-    def test_config_to_json_file(self):
-        config_first = TransfoXLConfig(vocab_size_or_config_json_file=96, d_embed=37)
-        json_file_path = "/tmp/config.json"
-        config_first.to_json_file(json_file_path)
-        config_second = TransfoXLConfig.from_json_file(json_file_path)
-        os.remove(json_file_path)
-        self.assertEqual(config_second.to_dict(), config_first.to_dict())
    @pytest.mark.slow
    def test_model_from_pretrained(self):
@@ -209,28 +194,18 @@ class TransfoXLModelTest(unittest.TestCase):
        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)
-    @classmethod
+        tester.set_seed()
-    def ids_tensor(cls, shape, vocab_size, rng=None, name=None):
+        config_and_inputs = tester.prepare_config_and_inputs()
-        """Creates a random int32 tensor of the shape within the vocab size."""
+        tester.create_and_check_transfo_xl_commons(*config_and_inputs)
-        if rng is None:
-            rng = random.Random()
-        total_dims = 1
-        for dim in shape:
-            total_dims *= dim
-        values = []
-        for _ in range(total_dims):
-            values.append(rng.randint(0, vocab_size - 1))
-        return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
 if __name__ == "__main__":

--- a/tests/modeling_xlnet_test.py
+++ b/tests/modeling_xlnet_test.py
@@ -25,9 +25,11 @@ import pytest
 import torch
-from pytorch_pretrained_bert import (XLNetConfig, XLNetModel, XLNetLMHeadModel)
+from pytorch_pretrained_bert import (XLNetConfig, XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering)
 from pytorch_pretrained_bert.modeling_xlnet import PRETRAINED_MODEL_ARCHIVE_MAP
+from .model_tests_commons import ConfigTester, create_and_check_commons, ids_tensor
 class XLNetModelTest(unittest.TestCase):
    class XLNetModelTester(object):
@@ -42,43 +44,48 @@ class XLNetModelTest(unittest.TestCase):
                     use_labels=True,
                     vocab_size=99,
                     cutoffs=[10, 50, 80],
-                     d_model=32,
+                     hidden_size=32,
-                     n_head=4,
+                     num_attention_heads=4,
                     d_inner=128,
-                     n_layer=5,
+                     num_hidden_layers=5,
                     max_position_embeddings=10,
                     untie_r=True,
                     bi_data=False,
                     same_length=False,
                     seed=1,
-                     type_vocab_size=2):
+                     type_vocab_size=2,
+                     all_model_classes=(XLNetModel, XLNetLMHeadModel,
+                                        XLNetForSequenceClassification, XLNetForQuestionAnswering),
+            ):
            self.parent = parent
            self.batch_size = batch_size
            self.seq_length = seq_length
            self.mem_len = mem_len
+            # self.key_len = seq_length + mem_len
            self.clamp_len = clamp_len
            self.reuse_len = reuse_len
            self.is_training = is_training
            self.use_labels = use_labels
            self.vocab_size = vocab_size
            self.cutoffs = cutoffs
-            self.d_model = d_model
+            self.hidden_size = hidden_size
-            self.n_head = n_head
+            self.num_attention_heads = num_attention_heads
            self.d_inner = d_inner
-            self.n_layer = n_layer
+            self.num_hidden_layers = num_hidden_layers
            self.max_position_embeddings = max_position_embeddings
            self.bi_data = bi_data
            self.untie_r = untie_r
            self.same_length = same_length
            self.seed = seed
            self.type_vocab_size = type_vocab_size
+            self.all_model_classes = all_model_classes
        def prepare_config_and_inputs(self):
-            input_ids_1 = XLNetModelTest.ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+            input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
-            input_ids_2 = XLNetModelTest.ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+            input_ids_2 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
-            segment_ids = XLNetModelTest.ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+            segment_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
-            input_ids_q = XLNetModelTest.ids_tensor([self.batch_size, self.seq_length + 1], self.vocab_size)
+            input_ids_q = ids_tensor([self.batch_size, self.seq_length + 1], self.vocab_size)
            perm_mask = torch.zeros(self.batch_size, self.seq_length + 1, self.seq_length + 1, dtype=torch.float)
            perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
            target_mapping = torch.zeros(self.batch_size, 1, self.seq_length + 1, dtype=torch.float)
@@ -89,8 +96,8 @@ class XLNetModelTest(unittest.TestCase):
            # token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
            # input_mask: float32 Tensor in shape [bsz, len], the input mask.
            #     0 for real tokens and 1 for padding.
-            # mems: a list of float32 Tensors in shape [bsz, mem_len, d_model], memory
+            # mems: a list of float32 Tensors in shape [bsz, mem_len, hidden_size], memory
-            #     from previous batches. The length of the list equals n_layer.
+            #     from previous batches. The length of the list equals num_hidden_layers.
            #     If None, no memory is used.
            # perm_mask: float32 Tensor in shape [bsz, len, len].
            #     If perm_mask[k, i, j] = 0, i attend to j in batch k;
@@ -108,14 +115,14 @@ class XLNetModelTest(unittest.TestCase):
            lm_labels = None
            if self.use_labels:
-                lm_labels = XLNetModelTest.ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+                lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
            config = XLNetConfig(
                vocab_size_or_config_json_file=self.vocab_size,
-                d_model=self.d_model,
+                d_model=self.hidden_size,
-                n_head=self.n_head,
+                n_head=self.num_attention_heads,
                d_inner=self.d_inner,
-                n_layer=self.n_layer,
+                n_layer=self.num_hidden_layers,
                untie_r=self.untie_r,
                max_position_embeddings=self.max_position_embeddings,
                mem_len=self.mem_len,
@@ -159,7 +166,7 @@ class XLNetModelTest(unittest.TestCase):
                [self.batch_size, self.seq_length, self.vocab_size])
            self.parent.assertListEqual(
                list(list(mem.size()) for mem in result["mems_1"]),
-                [[self.seq_length, self.batch_size, self.d_model]] * self.n_layer)
+                [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
            self.parent.assertListEqual(
                list(result["loss_2"].size()),
@@ -169,24 +176,18 @@ class XLNetModelTest(unittest.TestCase):
                [self.batch_size, self.seq_length, self.vocab_size])
            self.parent.assertListEqual(
                list(list(mem.size()) for mem in result["mems_2"]),
-                [[self.mem_len, self.batch_size, self.d_model]] * self.n_layer)
+                [[self.mem_len, 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, target_mapping, inp_q, segment_ids, lm_labels):
+            inputs_dict = {'input_ids': input_ids_1}
+            create_and_check_commons(self, config, inputs_dict)
    def test_default(self):
        self.run_tester(XLNetModelTest.XLNetModelTester(self))
-    def test_config_to_json_string(self):
+    def test_config(self):
-        config = XLNetConfig(vocab_size_or_config_json_file=96, d_model=16*4)
+        config_tester = ConfigTester(self, config_class=XLNetConfig, d_inner=37)
-        obj = json.loads(config.to_json_string())
+        config_tester.run_common_tests()
-        self.assertEqual(obj["n_token"], 96)
-        self.assertEqual(obj["d_model"], 16*4)
-    def test_config_to_json_file(self):
-        config_first = XLNetConfig(vocab_size_or_config_json_file=96, d_model=16*4)
-        json_file_path = "/tmp/config.json"
-        config_first.to_json_file(json_file_path)
-        config_second = XLNetConfig.from_json_file(json_file_path)
-        os.remove(json_file_path)
-        self.assertEqual(config_second.to_dict(), config_first.to_dict())
    @pytest.mark.slow
    def test_model_from_pretrained(self):
@@ -197,27 +198,14 @@ class XLNetModelTest(unittest.TestCase):
            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_lm_head(*config_and_inputs)
        tester.check_transfo_xl_lm_head_output(output_result)
-    @classmethod
+        tester.set_seed()
-    def ids_tensor(cls, shape, vocab_size, rng=None, name=None):
+        config_and_inputs = tester.prepare_config_and_inputs()
-        """Creates a random int32 tensor of the shape within the vocab size."""
+        tester.create_and_check_xlnet_commons(*config_and_inputs)
-        if rng is None:
-            rng = random.Random()
-        total_dims = 1
-        for dim in shape:
-            total_dims *= dim
-        values = []
-        for _ in range(total_dims):
-            values.append(rng.randint(0, vocab_size - 1))
-        return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
    @classmethod
    def mask_tensor(cls, shape, vocab_size, rng=None, name=None):

--- a/tests/optimization_test.py
+++ b/tests/optimization_test.py