fix #827

pytorch_transformers/convert_xlm_checkpoint_to_pytorch.py

@@ -36,7 +36,7 @@ def convert_xlm_checkpoint_to_pytorch(xlm_checkpoint_path, pytorch_dump_folder_p
     model = chkpt['model']
 
     config = chkpt['params']
-    config = dict((n, v) for n, v in config.items() if not isinstance(v, (torch.Tensor, numpy.ndarray)))
+    config = dict((n, v) for n, v in config.items() if not isinstance(v, (torch.FloatTensor, numpy.ndarray)))
 
     vocab = chkpt['dico_word2id']
     vocab = dict((s + '</w>' if s.find('@@') == -1 and i > 13 else s.replace('@@', ''), i) for s, i in vocab.items())

pytorch_transformers/modeling_bert.py

@@ -609,11 +609,11 @@ def init_weights(self, module):
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
             (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -1027,12 +1027,12 @@ class BertForMultipleChoice(BertPreTrainedModel):
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
             (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, num_choices, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             The second dimension of the input (`num_choices`) indicates the number of choices to score.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.

pytorch_transformers/modeling_gpt2.py

@@ -402,11 +402,11 @@ def init_weights(self, module):
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -638,11 +638,11 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, num_choices, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.

pytorch_transformers/modeling_openai.py

@@ -412,11 +412,11 @@ def init_weights(self, module):
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -624,11 +624,11 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, num_choices, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.

pytorch_transformers/modeling_transfo_xl.py

@@ -394,8 +394,8 @@ def __init__(self, n_head, d_model, d_head, dropout, dropatt=0,
         self.pre_lnorm = pre_lnorm
 
         if r_r_bias is None or r_w_bias is None: # Biases are not shared
-            self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-            self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
+            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
         else:
             self.r_r_bias = r_r_bias
             self.r_w_bias = r_w_bias
@@ -483,8 +483,8 @@ def __init__(self, n_head, d_model, d_head, dropout, dropatt=0,
         self.pre_lnorm = pre_lnorm
 
         if r_r_bias is None or r_w_bias is None: # Biases are not shared
-            self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-            self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
+            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
         else:
             self.r_r_bias = r_r_bias
             self.r_w_bias = r_w_bias
@@ -803,13 +803,13 @@ def __init__(self, n_token, d_embed, d_proj, cutoffs, div_val=1,
                 nn.Embedding(n_token, d_embed, sparse=sample_softmax>0)
             )
             if d_proj != d_embed:
-                self.emb_projs.append(nn.Parameter(torch.Tensor(d_proj, d_embed)))
+                self.emb_projs.append(nn.Parameter(torch.FloatTensor(d_proj, d_embed)))
         else:
             for i in range(len(self.cutoffs)):
                 l_idx, r_idx = self.cutoff_ends[i], self.cutoff_ends[i+1]
                 d_emb_i = d_embed // (div_val ** i)
                 self.emb_layers.append(nn.Embedding(r_idx-l_idx, d_emb_i))
-                self.emb_projs.append(nn.Parameter(torch.Tensor(d_proj, d_emb_i)))
+                self.emb_projs.append(nn.Parameter(torch.FloatTensor(d_proj, d_emb_i)))
 
     def forward(self, inp):
         if self.div_val == 1:
@@ -941,7 +941,7 @@ def set_num_special_tokens(self, num_special_tokens):
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `mems` output below). Can be used to speed up sequential decoding and attend to longer context.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -1003,8 +1003,8 @@ def __init__(self, config):
         self.attn_type = config.attn_type
 
         if not config.untie_r:
-            self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-            self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
+            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
 
         self.layers = nn.ModuleList()
         if config.attn_type == 0: # the default attention
@@ -1046,14 +1046,14 @@ def __init__(self, config):
         if self.attn_type == 0: # default attention
             self.pos_emb = PositionalEmbedding(self.d_model)
         elif self.attn_type == 1: # learnable
-            self.r_emb = nn.Parameter(torch.Tensor(
+            self.r_emb = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head, self.d_head))
-            self.r_bias = nn.Parameter(torch.Tensor(
+            self.r_bias = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head))
         elif self.attn_type == 2: # absolute standard
             self.pos_emb = PositionalEmbedding(self.d_model)
         elif self.attn_type == 3: # absolute deeper SA
-            self.r_emb = nn.Parameter(torch.Tensor(
+            self.r_emb = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head, self.d_head))
 
         self.apply(self.init_weights)

pytorch_transformers/modeling_transfo_xl_utilities.py

@@ -56,7 +56,7 @@ def __init__(self, n_token, d_embed, d_proj, cutoffs, div_val=1,
             for i in range(len(self.cutoffs)):
                 if d_proj != d_embed:
                     self.out_projs.append(
-                        nn.Parameter(torch.Tensor(d_proj, d_embed))
+                        nn.Parameter(torch.FloatTensor(d_proj, d_embed))
                     )
                 else:
                     self.out_projs.append(None)
@@ -68,7 +68,7 @@ def __init__(self, n_token, d_embed, d_proj, cutoffs, div_val=1,
                 d_emb_i = d_embed // (div_val ** i)
 
                 self.out_projs.append(
-                    nn.Parameter(torch.Tensor(d_proj, d_emb_i))
+                    nn.Parameter(torch.FloatTensor(d_proj, d_emb_i))
                 )
 
                 self.out_layers.append(nn.Linear(d_emb_i, r_idx-l_idx))

pytorch_transformers/modeling_xlm.py

@@ -436,7 +436,7 @@ def init_weights(self, module):
             A parallel sequence of tokens to be used to indicate the language of each token in the input.
             Indices are selected in the pre-trained language vocabulary,
             i.e. in the range ``[0, config.n_langs - 1[``.
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
@@ -449,7 +449,7 @@ def init_weights(self, module):
             hidden-states (key and values in the attention blocks) as computed by the model
             (see `cache` output below). Can be used to speed up sequential decoding.
             The dictionary object will be modified in-place during the forward pass to add newly computed hidden-states.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.

pytorch_transformers/modeling_xlnet.py

@@ -367,16 +367,16 @@ def __init__(self, config):
         self.d_model = config.d_model
         self.scale = 1 / (config.d_head ** 0.5)
 
-        self.q = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.k = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.v = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.o = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.r = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
+        self.q = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.k = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.v = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.o = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.r = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
 
-        self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-        self.r_s_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-        self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-        self.seg_embed = nn.Parameter(torch.Tensor(2, self.n_head, self.d_head))
+        self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+        self.r_s_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+        self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+        self.seg_embed = nn.Parameter(torch.FloatTensor(2, self.n_head, self.d_head))
 
         self.layer_norm = XLNetLayerNorm(config.d_model, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.dropout)
@@ -660,11 +660,11 @@ def init_weights(self, module):
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **input_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **input_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Negative of `attention_mask`, i.e. with 0 for real tokens and 1 for padding.
             Kept for compatibility with the original code base.
@@ -685,7 +685,7 @@ def init_weights(self, module):
             Mask to indicate the output tokens to use.
             If ``target_mapping[k, i, j] = 1``, the i-th predict in batch k is on the j-th token.
             Only used during pretraining for partial prediction or for sequential decoding (generation).
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -735,7 +735,7 @@ def __init__(self, config):
         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))
+        self.mask_emb = nn.Parameter(torch.FloatTensor(1, 1, config.d_model))
         self.layer = nn.ModuleList([XLNetLayer(config) for _ in range(config.n_layer)])
         self.dropout = nn.Dropout(config.dropout)