SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

WIP reodering arguments for torchscript and TF

Browse Source
This commit is contained in:
thomwolf
2019-09-04 12:43:18 +02:00
parent 38b79b5a63
commit e25cba78cf
10 changed files with 331 additions and 253 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
pytorch_transformers/modeling_gpt2.py
View File

@@ -257,7 +257,7 @@ class Attention(nn.Module):
self.n_head = self.n_head - len(heads)
self.pruned_heads = self.pruned_heads.union(heads)
def _attn(self, q, k, v, head_mask=None):
def _attn(self, q, k, v, attention_mask=None, head_mask=None):
w = torch.matmul(q, k)
if self.scale:
w = w / math.sqrt(v.size(-1))
@@ -265,6 +265,10 @@ class Attention(nn.Module):
b = self.bias[:, :, ns-nd:ns, :ns]
w = w * b - 1e4 * (1 - b)
if attention_mask is not None:
# Apply the attention mask
w = w + attention_mask
w = nn.Softmax(dim=-1)(w)
w = self.attn_dropout(w)
@@ -290,7 +294,7 @@ class Attention(nn.Module):
else:
return x.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
def forward(self, x, layer_past=None, head_mask=None):
def forward(self, x, layer_past=None, attention_mask=None, head_mask=None):
x = self.c_attn(x)
query, key, value = x.split(self.split_size, dim=2)
query = self.split_heads(query)
@@ -302,7 +306,7 @@ class Attention(nn.Module):
value = torch.cat((past_value, value), dim=-2)
present = torch.stack((key.transpose(-2, -1), value)) # transpose to have same shapes for stacking
attn_outputs = self._attn(query, key, value, head_mask)
attn_outputs = self._attn(query, key, value, attention_mask, head_mask)
a = attn_outputs[0]
a = self.merge_heads(a)
@@ -337,8 +341,11 @@ class Block(nn.Module):
self.ln_2 = LayerNorm(nx, eps=config.layer_norm_epsilon)
self.mlp = MLP(4 * nx, config)
def forward(self, x, layer_past=None, head_mask=None):
output_attn = self.attn(self.ln_1(x), layer_past=layer_past, head_mask=head_mask)
def forward(self, x, layer_past=None, attention_mask=None, head_mask=None):
output_attn = self.attn(self.ln_1(x),
layer_past=layer_past,
attention_mask=attention_mask,
head_mask=head_mask)
a = output_attn[0] # output_attn: a, present, (attentions)
x = x + a
@@ -404,17 +411,21 @@ GPT2_INPUTS_DOCSTRING = r""" Inputs:
Indices can be obtained using :class:`pytorch_transformers.BPT2Tokenizer`.
See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
:func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
**position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
Indices of positions of each input sequence tokens in the position embeddings.
Selected in the range ``[0, config.max_position_embeddings - 1]``.
**token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
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).
**past**:
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.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.
**token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
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).
**position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
Indices of positions of each input sequence tokens in the position embeddings.
Selected in the range ``[0, config.max_position_embeddings - 1]``.
**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]``:
@@ -473,7 +484,7 @@ class GPT2Model(GPT2PreTrainedModel):
for layer, heads in heads_to_prune.items():
self.h[layer].attn.prune_heads(heads)
def forward(self, input_ids, position_ids=None, token_type_ids=None, past=None, head_mask=None):
def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
if past is None:
past_length = 0
past = [None] * len(self.h)
@@ -483,6 +494,23 @@ class GPT2Model(GPT2PreTrainedModel):
position_ids = torch.arange(past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device)
position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
# Attention mask.
if attention_mask is not None:
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
attention_mask = (1.0 - attention_mask) * -10000.0
# 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
@@ -520,7 +548,11 @@ class GPT2Model(GPT2PreTrainedModel):
if self.output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
outputs = block(hidden_states, layer_past, head_mask[i])
outputs = block(hidden_states,
past=layer_past,
attention_mask=attention_mask,
head_mask=head_mask[i])
hidden_states, present = outputs[:2]
presents = presents + (present,)
@@ -601,9 +633,14 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
self._tie_or_clone_weights(self.lm_head,
self.transformer.wte)
def forward(self, input_ids, position_ids=None, token_type_ids=None, labels=None, past=None, head_mask=None):
transformer_outputs = self.transformer(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
past=past, head_mask=head_mask)
def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
labels=None):
transformer_outputs = self.transformer(input_ids,
past=past,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask)
hidden_states = transformer_outputs[0]
lm_logits = self.lm_head(hidden_states)
@@ -626,33 +663,12 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
head on top e.g. for RocStories/SWAG tasks. The two heads are two linear layers.
The language modeling head has its weights tied to the input embeddings,
the classification head takes as input the input of a specified classification token index in the input sequence).
""", GPT2_START_DOCSTRING)
""", GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
r""" Inputs:
**input_ids**: ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
Indices of input sequence tokens in the vocabulary.
The second dimension of the input (`num_choices`) indicates the number of choices to score.
Indices can be obtained using :class:`pytorch_transformers.BPT2Tokenizer`.
See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
:func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
r"""
**mc_token_ids**: ``torch.LongTensor`` of shape ``(batch_size, num_choices)``:
Index of the classification token in each input sequence.
Selected in the range ``[0, input_ids.size(-1) - 1[``.
**position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
Indices of positions of each input sequence tokens in the position embeddings.
Selected in the range ``[0, config.max_position_embeddings - 1]``.
**token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
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).
**past**:
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.
**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**.
**lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
Labels for language modeling.
Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
@@ -725,10 +741,15 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
self._tie_or_clone_weights(self.lm_head,
self.transformer.wte)
def forward(self, input_ids, mc_token_ids=None, lm_labels=None, mc_labels=None, token_type_ids=None,
position_ids=None, past=None, head_mask=None):
transformer_outputs = self.transformer(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
past=past, head_mask=head_mask)
def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
mc_token_ids=None, lm_labels=None, mc_labels=None):
transformer_outputs = self.transformer(input_ids,
past=past,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask)
hidden_states = transformer_outputs[0]
lm_logits = self.lm_head(hidden_states)