def test_add_positional_features(self):
# This is hard to test, so we check that we get the same result as the
# original tensorflow implementation:
# https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/layers/common_attention.py#L270
tensor2tensor_result = numpy.asarray([[0.00000000e+00, 0.00000000e+00, 1.00000000e+00, 1.00000000e+00],
[8.41470957e-01, 9.99999902e-05, 5.40302277e-01, 1.00000000e+00],
[9.09297407e-01, 1.99999980e-04, -4.16146845e-01, 1.00000000e+00]])
tensor = torch.zeros([2, 3, 4])
result = util.add_positional_features(tensor, min_timescale=1.0, max_timescale=1.0e4)
numpy.testing.assert_almost_equal(result[0].detach().cpu().numpy(), tensor2tensor_result)
numpy.testing.assert_almost_equal(result[1].detach().cpu().numpy(), tensor2tensor_result)
# Check case with odd number of dimensions.
tensor2tensor_result = numpy.asarray([[0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 1.00000000e+00,
1.00000000e+00, 1.00000000e+00, 0.00000000e+00],
[8.41470957e-01, 9.99983307e-03, 9.99999902e-05, 5.40302277e-01,
9.99949992e-01, 1.00000000e+00, 0.00000000e+00],
[9.09297407e-01, 1.99986659e-02, 1.99999980e-04, -4.16146815e-01,
9.99800026e-01, 1.00000000e+00, 0.00000000e+00]])
tensor = torch.zeros([2, 3, 7])
result = util.add_positional_features(tensor, min_timescale=1.0, max_timescale=1.0e4)
numpy.testing.assert_almost_equal(result[0].detach().cpu().numpy(), tensor2tensor_result)
numpy.testing.assert_almost_equal(result[1].detach().cpu().numpy(), tensor2tensor_result)
def test_add_positional_features(self):
# This is hard to test, so we check that we get the same result as the
# original tensorflow implementation:
# https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/layers/common_attention.py#L270
tensor2tensor_result = numpy.asarray([[0.00000000e+00, 0.00000000e+00, 1.00000000e+00, 1.00000000e+00],
[8.41470957e-01, 9.99999902e-05, 5.40302277e-01, 1.00000000e+00],
[9.09297407e-01, 1.99999980e-04, -4.16146845e-01, 1.00000000e+00]])
tensor = torch.zeros([2, 3, 4])
result = util.add_positional_features(tensor, min_timescale=1.0, max_timescale=1.0e4)
numpy.testing.assert_almost_equal(result[0].detach().cpu().numpy(), tensor2tensor_result)
numpy.testing.assert_almost_equal(result[1].detach().cpu().numpy(), tensor2tensor_result)
# Check case with odd number of dimensions.
tensor2tensor_result = numpy.asarray([[0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 1.00000000e+00,
1.00000000e+00, 1.00000000e+00, 0.00000000e+00],
[8.41470957e-01, 9.99983307e-03, 9.99999902e-05, 5.40302277e-01,
9.99949992e-01, 1.00000000e+00, 0.00000000e+00],
[9.09297407e-01, 1.99986659e-02, 1.99999980e-04, -4.16146815e-01,
9.99800026e-01, 1.00000000e+00, 0.00000000e+00]])
tensor = torch.zeros([2, 3, 7])
result = util.add_positional_features(tensor, min_timescale=1.0, max_timescale=1.0e4)
numpy.testing.assert_almost_equal(result[0].detach().cpu().numpy(), tensor2tensor_result)
numpy.testing.assert_almost_equal(result[1].detach().cpu().numpy(), tensor2tensor_result)
def _encode(self, source_tokens: Dict[str, torch.Tensor], segments: Dict[str, torch.Tensor],
source_entity_length: torch.Tensor, edge_mask: torch.Tensor, ) -> Dict[str, torch.Tensor]:
"""
:param source_tokens:
:param segments:
:param merge_indicators:
:return:
"""
# shape: (batch_size, encode_length, embedding_dim)
source_embedded_input = self._embed_source(source_tokens, source_entity_length)
# shape: (batch_size, encode_length, embedding_dim)
segments_embedded_input = self._segment_embedder(segments)
encode_length = segments_embedded_input.size(1)
assert source_embedded_input.size(1) == segments_embedded_input.size(1)
# token_mask = (segments['tokens'] == self._token_index).unsqueeze(-1).float()
# valid_token_embedded_input = batched_embedded_input * token_mask
# valid_token_embedded_input = util.add_positional_features(valid_token_embedded_input)
# valid_token_embedded_input = batched_embedded_input * (1 - token_mask) + valid_token_embedded_input * token_mask
if self._source_embedding_dim == self._encoder_d_model:
batched_embedded_input = segments_embedded_input + source_embedded_input
final_embedded_input = util.add_positional_features(batched_embedded_input)
else:
batched_embedded_input = torch.cat([source_embedded_input, segments_embedded_input], dim=-1)
final_embedded_input = util.add_positional_features(batched_embedded_input)
# shape: (encode_length, batch_size, d_model)
final_embedded_input = final_embedded_input.permute(1, 0, 2)
# shape: (batch_size, encode_length)
source_mask = util.get_text_field_mask(segments)
source_key_padding_mask = (1 - source_mask.byte()).bool()
if not self._use_gnn_encoder:
# shape: (encode_length, batch_size, d_model)
encoder_outputs = self._encoder(final_embedded_input, src_key_padding_mask=source_key_padding_mask)
else:
# GNN encoders
encoder_outputs = self._encoder(src=final_embedded_input, edge_mask=edge_mask.permute(0, 2, 3, 1),
padding_mask=source_key_padding_mask)
source_token_mask = (segments['tokens'] == self._token_index).float()
return {
"source_mask": source_mask,
"source_key_padding_mask": source_key_padding_mask,
"source_token_mask": source_token_mask,
"encoder_outputs": encoder_outputs,
"source_embedded": batched_embedded_input,
"source_raw_embedded": source_embedded_input,
}
def forward(
self, # type: ignore
abstract_text: Dict[str, torch.LongTensor],
label: torch.LongTensor = None) -> Dict[str, torch.Tensor]:
# pylint: disable=arguments-differ
embedded_abstract_text = self.text_field_embedder(abstract_text,
num_wrapping_dims=1)
abstract_text_mask = util.get_text_field_mask(abstract_text,
num_wrapping_dims=1)
num_instance, num_sentence, num_word, _ = embedded_abstract_text.size()
embedded_abstract_text = embedded_abstract_text.view(
num_instance * num_sentence, num_word, -1)
abstract_text_mask = abstract_text_mask.view(
num_instance * num_sentence, -1)
if self.use_positional_encoding:
embedded_abstract_text = util.add_positional_features(
embedded_abstract_text)
encoded_abstract_text = self.word_encoder(embedded_abstract_text,
abstract_text_mask)
attended_sentences = self.word_level_attention(encoded_abstract_text,
abstract_text_mask)
attended_sentences = attended_sentences.view(num_instance,
num_sentence, -1)
abstract_text_mask = abstract_text_mask.view(num_instance,
num_sentence,
-1).sum(2).ge(1).long()
if self.use_positional_encoding:
attended_sentences = util.add_positional_features(
attended_sentences)
attended_sentences = self.sentence_encoder(attended_sentences,
abstract_text_mask)
attended_abstract_text = self.sentence_level_attention(
attended_sentences, abstract_text_mask)
outputs = self.classifier_feedforward(attended_abstract_text)
logits = torch.sigmoid(outputs)
logits = logits.unsqueeze(0) if logits.dim() < 2 else logits
output_dict = {'logits': logits}
if label is not None:
outputs = outputs.unsqueeze(0) if outputs.dim() < 2 else outputs
loss = self.loss(outputs, label.squeeze(-1))
for metric in self.metrics.values():
metric(logits, label.squeeze(-1))
output_dict["loss"] = loss
return output_dict
def forward(self, **inputs):
x = self.aa_embedder(inputs['aa']) + self.pssm_embedder(inputs['pssm'])
if 'ss' in inputs and self.use_ss:
x += self.ss_embedder(inputs['ss'])
if self.use_positional_encoding:
add_positional_features(x)
x = self.input_dropout(x)
mask = get_text_field_mask(inputs['aa'])
x = self.encoder(x, mask)
if 'msa' in inputs and self.msa_encoder is not None:
x += self.msa_encoder(inputs['msa'])
x = self.decoder(self.feedforward(x))
outputs = {
'predictions': x,
'protein_id': inputs['protein_id'],
'length': inputs['length']
}
if 'dcalpha' in inputs:
mask = mask.unsqueeze(-1).float()
if self.target == 'dcalpha':
target = inputs['dcalpha']
mask = mask.matmul(mask.transpose(dim0=-2, dim1=-1))
mask_triu = torch.triu(torch.ones_like(mask[0]),
diagonal=1).unsqueeze(0).to(mask.device)
mask *= mask_triu
elif self.target == 'angles':
target = torch.stack([inputs['psi'], inputs['phi']], dim=2)
else:
target = inputs['coords']
mse = ((mask * (x - target))**2).sum() / mask.sum()
if torch.isnan(mse):
while True:
expr = input('\nInput = ')
if expr == 'q':
exit(0)
try:
print(eval(expr))
except Exception as e:
print(e)
self.metrics['rmse']((mse**0.5).detach().item())
outputs['loss'] = mse
return outputs
def forward(self, inputs: torch.Tensor, mask: torch.BoolTensor):
if self._use_positional_encoding:
output = add_positional_features(inputs)
else:
output = inputs
for i in range(len(self._attention_layers)):
# It's necessary to use `getattr` here because the elements stored
# in the lists are not replicated by torch.nn.parallel.replicate
# when running on multiple GPUs. Please use `ModuleList` in new
# code. It handles this issue transparently. We've kept `add_module`
# (in conjunction with `getattr`) solely for backwards compatibility
# with existing serialized models.
attention = getattr(self, f"self_attention_{i}")
feedforward = getattr(self, f"feedforward_{i}")
feedforward_layer_norm = getattr(self,
f"feedforward_layer_norm_{i}")
layer_norm = getattr(self, f"layer_norm_{i}")
cached_input = output
# Project output of attention encoder through a feedforward
# network and back to the input size for the next layer.
# shape (batch_size, timesteps, input_size)
feedforward_output = feedforward(output)
feedforward_output = self.dropout(feedforward_output)
if feedforward_output.size() == cached_input.size():
# First layer might have the wrong size for highway
# layers, so we exclude it here.
feedforward_output = feedforward_layer_norm(
feedforward_output + cached_input)
# shape (batch_size, sequence_length, hidden_dim)
attention_output = attention(feedforward_output, mask)
output = layer_norm(
self.dropout(attention_output) + feedforward_output)
return output
def forward(self, # type: ignore
abstract_text: Dict[str, torch.LongTensor],
local_label: torch.LongTensor = None,
global_label: torch.LongTensor = None) -> Dict[str, torch.Tensor]:
# pylint: disable=arguments-differ
embedded_abstract_text = self.text_field_embedder(abstract_text)
abstract_text_mask = util.get_text_field_mask(abstract_text)
if self.use_positional_encoding:
embedded_abstract_text = util.add_positional_features(embedded_abstract_text)
encoded_abstract_text = self.abstract_text_encoder(embedded_abstract_text, abstract_text_mask)
attended_abstract_text = self.attention_encoder(encoded_abstract_text, abstract_text_mask)
local_outputs, global_outputs = self.HMCN_recurrent(attended_abstract_text)
logits = self.local_globel_tradeoff * global_outputs + (1 - self.local_globel_tradeoff) * local_outputs
logits = torch.sigmoid(logits)
logits = logits.unsqueeze(0) if logits.dim() < 2 else logits
output_dict = {'logits': logits}
if local_label is not None and global_label is not None:
local_outputs = local_outputs.unsqueeze(0) if local_outputs.dim() < 2 else local_outputs
global_outputs = global_outputs.unsqueeze(0) if global_outputs.dim() < 2 else global_outputs
loss = self.loss(local_outputs, global_outputs, local_label.squeeze(-1), global_label.squeeze(-1))
for metric in self.metrics.values():
metric(logits, global_label.squeeze(-1))
output_dict["loss"] = loss
return output_dict
def forward(
self, # type: ignore
abstract_text: Dict[str, torch.LongTensor],
label: torch.LongTensor = None) -> Dict[str, torch.Tensor]:
# pylint: disable=arguments-differ
embedded_abstract_text = self.text_field_embedder(abstract_text)
abstract_text_mask = util.get_text_field_mask(abstract_text)
encoded_abstract_text = self.abstract_text_encoder(
embedded_abstract_text, abstract_text_mask)
if self.use_positional_encoding:
encoded_abstract_text = util.add_positional_features(
encoded_abstract_text)
attended_abstract_text = self.attention_encoder(
encoded_abstract_text, abstract_text_mask)
outputs = self.classifier_feedforward(attended_abstract_text)
logits = torch.sigmoid(outputs)
logits = logits.unsqueeze(0) if len(logits.size()) < 2 else logits
output_dict = {'logits': logits}
if label is not None:
outputs = outputs.unsqueeze(0) if len(
outputs.size()) < 2 else outputs
loss = self.loss(outputs, label.squeeze(-1))
for metric in self.metrics.values():
metric(logits, label.squeeze(-1))
output_dict["loss"] = loss
return output_dict
def forward(self, inputs: torch.Tensor, mask: torch.BoolTensor):
output = inputs
if self._sinusoidal_positional_encoding:
output = add_positional_features(output)
if self._positional_embedding is not None:
position_ids = torch.arange(inputs.size(1),
dtype=torch.long,
device=output.device)
position_ids = position_ids.unsqueeze(0).expand(inputs.shape[:-1])
output = output + self._positional_embedding(position_ids)
# print()
# print(sum(output[0][4]), sum(output[0][100]))
# For some reason the torch transformer expects the shape (sequence, batch, features), not the more
# familiar (batch, sequence, features), so we have to fix it.
output = output.permute(1, 0, 2)
# For some other reason, the torch transformer takes the mask backwards.
mask = ~mask
output = self._transformer(output, src_key_padding_mask=mask)
output = output.permute(1, 0, 2)
# print(sum(inputs[0][4]), sum(inputs[0][100]))
# print(sum(output[0][4]), sum(output[0][100]))
# print()
return output
def forward(self, inputs, mask): # pylint: disable=arguments-differ
if self._use_positional_encoding:
output = add_positional_features(inputs)
else:
output = inputs
for (attention, feedforward, feedforward_layer_norm,
layer_norm) in zip(
self._attention_layers,
self._feedfoward_layers,
self._feed_forward_layer_norm_layers,
self._layer_norm_layers,
):
cached_input = output
# Project output of attention encoder through a feedforward
# network and back to the input size for the next layer.
# shape (batch_size, timesteps, input_size)
feedforward_output = feedforward(feedforward_layer_norm(output))
feedforward_output = self.dropout(feedforward_output)
if feedforward_output.size() == cached_input.size():
# First layer might have the wrong size for highway
# layers, so we exclude it here.
feedforward_output += cached_input
# shape (batch_size, sequence_length, hidden_dim)
attention_output = attention(layer_norm(feedforward_output), mask)
output = self.dropout(attention_output) + feedforward_output
return self._output_layer_norm(output)
def forward(self, inputs: torch.Tensor, mask: torch.Tensor): # pylint: disable=arguments-differ
if self._use_positional_encoding:
output = add_positional_features(inputs)
else:
output = inputs
for (attention,
feedforward,
feedforward_layer_norm,
layer_norm) in zip(self._attention_layers,
self._feedfoward_layers,
self._feed_forward_layer_norm_layers,
self._layer_norm_layers):
cached_input = output
# Project output of attention encoder through a feedforward
# network and back to the input size for the next layer.
# shape (batch_size, timesteps, input_size)
feedforward_output = feedforward(output)
feedforward_output = self.dropout(feedforward_output)
if feedforward_output.size() == cached_input.size():
# First layer might have the wrong size for highway
# layers, so we exclude it here.
feedforward_output = feedforward_layer_norm(feedforward_output + cached_input)
# shape (batch_size, sequence_length, hidden_dim)
attention_output = attention(feedforward_output, mask)
output = layer_norm(self.dropout(attention_output) + feedforward_output)
return output
def forward(self, inputs: torch.Tensor, mask: torch.Tensor = None) -> torch.Tensor:
if self._use_positional_encoding:
output = add_positional_features(inputs)
else:
output = inputs
total_sublayers = len(self._conv_layers) + 2
sublayer_count = 0
for conv_norm_layer, conv_layer in zip(self._conv_norm_layers, self._conv_layers):
conv_norm_out = self.dropout(conv_norm_layer(output))
conv_out = self.dropout(conv_layer(conv_norm_out.transpose_(1, 2)).transpose_(1, 2))
sublayer_count += 1
output = self.residual_with_layer_dropout(
output, conv_out, sublayer_count, total_sublayers
)
attention_norm_out = self.dropout(self.attention_norm_layer(output))
attention_out = self.dropout(self.attention_layer(attention_norm_out, mask))
sublayer_count += 1
output = self.residual_with_layer_dropout(
output, attention_out, sublayer_count, total_sublayers
)
feedforward_norm_out = self.dropout(self.feedforward_norm_layer(output))
feedforward_out = self.dropout(self.feedforward(feedforward_norm_out))
sublayer_count += 1
output = self.residual_with_layer_dropout(
output, feedforward_out, sublayer_count, total_sublayers
)
return output
def forward(self,
inputs: torch.Tensor,
mask: Optional[torch.Tensor] = None
) -> torch.Tensor:
output = add_positional_features(inputs)
output = self.projection(output)
for block in self.blocks:
output = block(output, mask)
return output
def _run_transformer_decoder(self, context_output: torch.Tensor,
query_output: torch.Tensor,
context_mask: torch.Tensor,
query_mask: torch.Tensor,
rewrite_embed: torch.Tensor,
rewrite_mask: torch.Tensor):
"""
实现Transformer解码器的decoder过程
:param _output: [B, _len, d_model]
:param _mask: [B, _len]
:param rewrite_embed: [B, cur_dec_len, d_model]
:param rewrite_mask: [B, cur_dec_len]，这里只是pad的mask，上三角mask在decoder内部实现
"""
if self._share_decoder_params:
rewrite_embed = add_positional_features(rewrite_embed)
previous_state = None
encoder_outputs = {
"context_output": context_output,
"query_output": query_output
}
source_mask = {"context_mask": context_mask, "query_mask": query_mask}
# dec_output: [B, dec_len, d_model]
# context_attn: [B, num_heads, dec_len, context_len]
# query_attn: [B, num_heads, dec_len, query_len]
# x_context: [B, dec_len, d_model]
# x_query: [B, dec_len, d_model]
dec_output, context_attn, query_attn, x_context, x_query = self.decoder(
previous_state, encoder_outputs, source_mask, rewrite_embed,
rewrite_mask)
# 如果共享解码器的参数
if self._share_decoder_params:
for _ in range(self.decoder_num_layers - 1):
dec_output, context_attn, query_attn, x_context, x_query = self.decoder(
previous_state, encoder_outputs, source_mask, dec_output,
rewrite_mask)
# sum the attention dists of different heads
context_attn = torch.sum(context_attn, dim=1, keepdim=False)
query_attn = torch.sum(query_attn, dim=1, keepdim=False)
# mask softmax get the final attention dists
context_attn = masked_softmax(context_attn, context_mask, dim=-1)
query_attn = masked_softmax(query_attn, query_mask, dim=-1)
# compute lambda
# [B, dec_len, 2]
# 注意这里和LSTM解码器的区别
# Transformer解码器是一次解码全部输出的，所以需要包含len维度
lamb = self._compute_lambda(dec_output,
dec_context=x_context,
dec_query=x_query)
return dec_output, context_attn, query_attn, lamb
def forward(self,
src: torch.Tensor,
kb: torch.Tensor,
src_mask: Optional[torch.Tensor] = None,
kb_mask: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
src = add_positional_features(src)
src = self.src_projection(src)
kb = self.kb_projection(kb)
for i in range(self.num_layers):
src = self.src_self_blocks[i](src, src_mask)
kb = self.kb_self_blocks[i](kb, kb_mask)
src, kb = self.rel_blocks[i](kb, src, kb_mask, src_mask)
if self.return_kb:
return src, kb
return src
def _forward(self,
embedding_sequence: torch.Tensor,
state: Dict[str, torch.Tensor],
apply_target_subsequent_mask: bool = True) -> torch.Tensor:
if self.use_position_encoding:
embedding_sequence = add_positional_features(embedding_sequence)
output = embedding_sequence
for layer in self.layers:
output = layer(
output,
state['target_mask'] if 'target_mask' in state else None,
apply_target_subsequent_mask=apply_target_subsequent_mask)
if self.pre_norm:
# We need to add an additional function of layer normalization to the top layer
# to prevent the excessively increased value caused by the sum of unnormalized output
# (https://arxiv.org/pdf/1906.01787.pdf)
output = self.output_layer_norm(output)
return output
def forward(
self,
previous_state: Dict[str, torch.Tensor],
encoder_outputs: torch.Tensor,
source_mask: torch.Tensor,
previous_steps_predictions: torch.Tensor,
previous_steps_mask: Optional[torch.Tensor] = None
) -> Tuple[Dict[str, torch.Tensor], torch.Tensor]:
seq_len = previous_steps_predictions.size(-2)
future_mask = torch.triu(
torch.ones(seq_len,
seq_len,
device=source_mask.device,
dtype=torch.float)).transpose(0, 1)
future_mask = future_mask.masked_fill(future_mask == 0,
float('-inf')).masked_fill(
future_mask == 1, float(0.0))
future_mask = Variable(future_mask)
if self._use_positional_encoding:
previous_steps_predictions = add_positional_features(
previous_steps_predictions)
previous_steps_predictions = self._dropout(previous_steps_predictions)
source_mask = ~(source_mask.bool())
if previous_steps_mask is not None:
previous_steps_mask = ~(previous_steps_mask.bool())
previous_steps_predictions = previous_steps_predictions.permute(
1, 0, 2)
encoder_outputs = encoder_outputs.permute(1, 0, 2)
output = self._decoder(previous_steps_predictions,
encoder_outputs,
tgt_mask=future_mask,
tgt_key_padding_mask=previous_steps_mask,
memory_key_padding_mask=source_mask)
return {}, output.permute(1, 0, 2)
def forward(self, embedding_sequence: torch.LongTensor) -> torch.Tensor:
"""
Parameters
----------
embedding_sequence : (batch_size, sequence_length, embedding_size)
Returns
-------
(batch_size, sequence_length, embedding_size)
"""
if self.learned:
batch_size, sequence_length, _ = embedding_sequence.size()
position_indices = torch.arange(sequence_length).to(
self.embedding.weight.device)
position_embeddings = self.embedding(position_indices)
position_embeddings = position_embeddings.unsqueeze(0).expand(
batch_size, sequence_length, -1)
return embedding_sequence + position_embeddings
else:
return util.add_positional_features(embedding_sequence)
def forward(self,
query_embeddings: torch.Tensor,
document_embeddings: torch.Tensor,
query_pad_oov_mask: torch.Tensor,
document_pad_oov_mask: torch.Tensor,
output_secondary_output: bool = False) -> torch.Tensor:
# pylint: disable=arguments-differ
query_embeddings = query_embeddings * query_pad_oov_mask.unsqueeze(-1)
document_embeddings = document_embeddings * document_pad_oov_mask.unsqueeze(
-1)
query_embeddings_context = self.contextualizer(
add_positional_features(query_embeddings).transpose(1, 0),
src_key_padding_mask=~query_pad_oov_mask.bool()).transpose(1, 0)
document_embeddings_context = self.contextualizer(
add_positional_features(document_embeddings).transpose(1, 0),
src_key_padding_mask=~document_pad_oov_mask.bool()).transpose(
1, 0)
query_embeddings = (self.mixer * query_embeddings +
(1 - self.mixer) * query_embeddings_context
) * query_pad_oov_mask.unsqueeze(-1)
document_embeddings = (self.mixer * document_embeddings +
(1 - self.mixer) * document_embeddings_context
) * document_pad_oov_mask.unsqueeze(-1)
#
# prepare embedding tensors & paddings masks
# -------------------------------------------------------
query_by_doc_mask = torch.bmm(
query_pad_oov_mask.unsqueeze(-1),
document_pad_oov_mask.unsqueeze(-1).transpose(-1, -2))
query_by_doc_mask_view = query_by_doc_mask.unsqueeze(-1)
#
# cosine matrix
# -------------------------------------------------------
# shape: (batch, query_max, doc_max)
cosine_matrix = self.cosine_module.forward(query_embeddings,
document_embeddings)
cosine_matrix_masked = cosine_matrix * query_by_doc_mask
cosine_matrix_extradim = cosine_matrix_masked.unsqueeze(-1)
#
# gaussian kernels & soft-TF
#
# first run through kernel, then sum on doc dim then sum on query dim
# -------------------------------------------------------
raw_kernel_results = torch.exp(
-torch.pow(cosine_matrix_extradim - self.mu, 2) /
(2 * torch.pow(self.sigma, 2)))
kernel_results_masked = raw_kernel_results * query_by_doc_mask_view
#
# mean kernels
#
#kernel_results_masked2 = kernel_results_masked.clone()
doc_lengths = torch.sum(document_pad_oov_mask, 1)
#kernel_results_masked2_mean = kernel_results_masked / doc_lengths.unsqueeze(-1)
per_kernel_query = torch.sum(kernel_results_masked, 2)
log_per_kernel_query = torch.log2(
torch.clamp(per_kernel_query, min=1e-10)) * self.nn_scaler
log_per_kernel_query_masked = log_per_kernel_query * query_pad_oov_mask.unsqueeze(
-1) # make sure we mask out padding values
per_kernel = torch.sum(log_per_kernel_query_masked, 1)
#per_kernel_query_mean = torch.sum(kernel_results_masked2_mean, 2)
per_kernel_query_mean = per_kernel_query / (
doc_lengths.view(-1, 1, 1) + 1
) # well, that +1 needs an explanation, sometimes training data is just broken ... (and nans all the things!)
log_per_kernel_query_mean = per_kernel_query_mean * self.nn_scaler
log_per_kernel_query_masked_mean = log_per_kernel_query_mean * query_pad_oov_mask.unsqueeze(
-1) # make sure we mask out padding values
per_kernel_mean = torch.sum(log_per_kernel_query_masked_mean, 1)
##
## "Learning to rank" layer - connects kernels with learned weights
## -------------------------------------------------------
dense_out = self.dense(per_kernel)
dense_mean_out = self.dense_mean(per_kernel_mean)
dense_comb_out = self.dense_comb(
torch.cat([dense_out, dense_mean_out], dim=1))
score = torch.squeeze(dense_comb_out, 1) #torch.tanh(dense_out), 1)
if output_secondary_output:
query_mean_vector = query_embeddings.sum(
dim=1) / query_pad_oov_mask.sum(dim=1).unsqueeze(-1)
return score, {
"score": score,
"dense_out": dense_out,
"dense_mean_out": dense_mean_out,
"per_kernel": per_kernel,
"per_kernel_mean": per_kernel_mean,
"query_mean_vector": query_mean_vector,
"cosine_matrix_masked": cosine_matrix_masked
}
else:
return score
def forward(self, source: torch.Tensor, target: torch.Tensor, metadata: dict, seq_labels: torch.Tensor = None,
reg_labels: torch.Tensor = None, source_mask: Optional[torch.Tensor]=None,
target_mask: Optional[torch.Tensor]=None, memory_mask: Optional[torch.Tensor]=None,
src_key_padding_mask: Optional[torch.Tensor]=None, tgt_key_padding_mask: Optional[torch.Tensor]=None,
memory_key_padding_mask: Optional[torch.Tensor]=None) -> Dict[str, torch.Tensor]:
r"""Take in and process masked source/target sequences.
Args:
source: the sequence to the encoder (required).
target: the sequence to the decoder (required).
metadata: the metadata of the samples (required).
seq_labels: the labels of each round (optional).
reg_labels: the labels of the total future payoff (optional).
source_mask: the additive mask for the src sequence (optional).
target_mask: the additive mask for the tgt sequence (optional).
memory_mask: the additive mask for the encoder output (optional).
src_key_padding_mask: the ByteTensor mask for src keys per batch (optional).
tgt_key_padding_mask: the ByteTensor mask for tgt keys per batch (optional).
memory_key_padding_mask: the ByteTensor mask for memory keys per batch (optional).
Shape:
- source: :math:`(S, N, E)`.
- target: :math:`(T, N, E)`.
- source_mask: :math:`(S, S)`.
- target_mask: :math:`(T, T)`.
- memory_mask: :math:`(T, S)`.
- src_key_padding_mask: :math:`(N, S)`.
- tgt_key_padding_mask: :math:`(N, T)`.
- memory_key_padding_mask: :math:`(N, S)`.
Note: [src/tgt/memory]_mask should be filled with
float('-inf') for the masked positions and float(0.0) else. These masks
ensure that predictions for position i depend only on the unmasked positions
j and are applied identically for each sequence in a batch.
[src/tgt/memory]_key_padding_mask should be a ByteTensor where True values are positions
that should be masked with float('-inf') and False values will be unchanged.
This mask ensures that no information will be taken from position i if
it is masked, and has a separate mask for each sequence in a batch.
- output: :math:`(T, N, E)`.
Note: Due to the multi-head attention architecture in the transformer model,
the output sequence length of a transformer is same as the input sequence
(i.e. target) length of the decode.
where S is the source sequence length, T is the target sequence length, N is the
batch size, E is the feature number
"""
if self._first_pair is not None:
if self._first_pair == metadata[0]['pair_id']:
self._epoch += 1
else:
self._first_pair = metadata[0]['pair_id']
output = dict()
src_key_padding_mask = get_text_field_mask({'tokens': source})
tgt_key_padding_mask = get_text_field_mask({'tokens': target})
# The torch transformer takes the mask backwards.
src_key_padding_mask_byte = ~src_key_padding_mask.bool()
tgt_key_padding_mask_byte = ~tgt_key_padding_mask.bool()
# create mask where only the first round is not masked --> need to be the same first round for all sequances
if self.only_raisha:
temp_mask = torch.ones(tgt_key_padding_mask_byte.shape, dtype=torch.bool)
temp_mask[:, 0] = False
tgt_key_padding_mask_byte = temp_mask
if self._sinusoidal_positional_encoding:
source = add_positional_features(source)
target = add_positional_features(target)
if torch.cuda.is_available(): # change to cuda
source = source.cuda()
target = target.cuda()
tgt_key_padding_mask = tgt_key_padding_mask.cuda()
src_key_padding_mask = src_key_padding_mask.cuda()
tgt_key_padding_mask_byte = tgt_key_padding_mask_byte.cuda()
src_key_padding_mask_byte = src_key_padding_mask_byte.cuda()
if seq_labels is not None:
seq_labels = seq_labels.cuda()
if reg_labels is not None:
reg_labels = reg_labels.cuda()
# The torch transformer expects the shape (sequence, batch, features), not the more
# familiar (batch, sequence, features), so we have to fix it.
source = source.permute(1, 0, 2)
target = target.permute(1, 0, 2)
if source.size(1) != target.size(1):
raise RuntimeError("the batch number of src and tgt must be equal")
if source.size(2) != self._input_dim or target.size(2) != self._input_dim:
raise RuntimeError("the feature number of src and tgt must be equal to d_model")
encoder_out = self.encoder(source, src_key_padding_mask=src_key_padding_mask_byte)
decoder_output = self.decoder(target, encoder_out, tgt_key_padding_mask=tgt_key_padding_mask_byte,
memory_key_padding_mask=src_key_padding_mask_byte)
decoder_output = decoder_output.permute(1, 0, 2)
if self.predict_seq:
if self.linear_layer is not None:
decoder_output = self.linear_layer(decoder_output) # add linear layer before hidden2tag
decision_logits = self.hidden2tag(decoder_output)
output['decision_logits'] = masked_softmax(decision_logits, tgt_key_padding_mask.unsqueeze(2))
self.seq_predictions = save_predictions_seq_models(prediction_df=self.seq_predictions,
mask=tgt_key_padding_mask,
predictions=output['decision_logits'],
gold_labels=seq_labels, metadata=metadata,
epoch=self._epoch, is_train=self.training,)
if self.predict_avg_total_payoff:
# (batch_size, seq_len, dimensions) * (batch_size, dimensions, 1) -> (batch_size, seq_len)
attention_output = self.attention(self.attention_vector, decoder_output, tgt_key_padding_mask)
# (batch_size, 1, seq_len) * (batch_size, seq_len, dimensions) -> (batch_size, dimensions)
attention_output = torch.bmm(attention_output.unsqueeze(1), decoder_output).squeeze()
# (batch_size, dimensions) -> (batch_size, batch_size)
linear_out = self.linear_after_attention_layer(attention_output)
# (batch_size, batch_size) -> (batch_size, 1)
regression_output = self.regressor(linear_out)
output['regression_output'] = regression_output
self.reg_predictions = save_predictions(prediction_df=self.reg_predictions,
predictions=output['regression_output'], gold_labels=reg_labels,
metadata=metadata, epoch=self._epoch, is_train=self.training,
int_label=False)
if seq_labels is not None or reg_labels is not None:
temp_loss = 0
if self.predict_seq and seq_labels is not None:
for metric_name, metric in self.metrics_dict_seq.items():
metric(decision_logits, seq_labels, tgt_key_padding_mask)
output['seq_loss'] = sequence_cross_entropy_with_logits(decision_logits, seq_labels,
tgt_key_padding_mask)
temp_loss += self.seq_weight_loss * output['seq_loss']
if self.predict_avg_total_payoff and reg_labels is not None:
for metric_name, metric in self.metrics_dict_reg.items():
metric(regression_output, reg_labels, tgt_key_padding_mask)
output['reg_loss'] = self.mse_loss(regression_output, reg_labels.view(reg_labels.shape[0], -1))
temp_loss += self.reg_weight_loss * output['reg_loss']
output['loss'] = temp_loss
return output
def forward(self,
content_id: torch.LongTensor,
bundle_id: torch.LongTensor,
feature: torch.FloatTensor,
user_id: torch.FloatTensor,
mask: torch.Tensor,
initial_state: Optional[TensorPair] = None,
ans_prev_correctly: Optional[torch.Tensor] = None):
batch_size, seq_len = content_id.shape
# content_emb: (batch, seq, dim)
content_emb = self.content_id_emb(content_id)
if self.hparams["emb_dropout"] > 0:
content_emb = self.emb_dropout(content_emb)
# content_emb: (batch, seq, dim)
feature = torch.cat([content_emb, feature], dim=-1)
if not self.hparams.get("no_prev_ans", False):
if ans_prev_correctly is None:
ans_prev_correctly = torch.ones(batch_size,
seq_len,
1,
device=self.device,
dtype=torch.float)
feature = torch.cat([ans_prev_correctly, feature], dim=-1)
if hasattr(self, "lstm_in_proj"):
feature = self.lstm_in_proj(feature)
feature = F.relu(feature)
# Apply LSTM
sequence_lengths = self.__class__.get_lengths_from_seq_mask(mask)
clamped_sequence_lengths = sequence_lengths.clamp(min=1)
if self.encoder_type == "augmented_lstm":
sorted_feature, sorted_sequence_lengths, restoration_indices, sorting_indices = \
self.__class__.sort_batch_by_length(feature, clamped_sequence_lengths)
packed_sequence_input = pack_padded_sequence(
sorted_feature,
sorted_sequence_lengths.data.tolist(),
enforce_sorted=False,
batch_first=True)
# encoder_out: (batch, seq_len, num_directions * hidden_size):
# h_t: (num_layers * num_directions, batch, hidden_size)
# - this dimension is valid regardless of batch_first=True!!
packed_lstm_out, (h_t, c_t) = self.encoder(packed_sequence_input)
# lstm_out: (batch, seq, num_directions * hidden_size)
lstm_out, _ = pad_packed_sequence(packed_lstm_out,
batch_first=True)
lstm_out = lstm_out.index_select(0, restoration_indices)
h_t = h_t.index_select(1, restoration_indices)
c_t = c_t.index_select(1, restoration_indices)
state = (h_t, c_t)
elif self.encoder_type == "attention":
if initial_state is None:
mask_seq_len = seq_len
attention_mask = self.__class__ \
.get_attention_mask(src_seq_len=mask_seq_len) \
.to(self.device)
permuted_feature = add_positional_features(feature, max_timescale=seq_len) \
.permute(1, 0, 2)
query = permuted_feature
else:
mask_seq_len = seq_len + 1
# initial_state: (batch, 1, dim)
initial_state = initial_state[0]
feature = torch.cat([feature, initial_state],
dim=1) # previous sequence summary vector
attention_mask = self.__class__ \
.get_attention_mask(src_seq_len=mask_seq_len,
target_seq_len=seq_len) \
.to(self.device)
# (seq, N, dim)
permuted_feature = add_positional_features(feature, max_timescale=mask_seq_len) \
.permute(1, 0, 2)
query = permuted_feature[1:]
# att_output: (seq, batch, dim)
att_output, att_weight = self.encoder(query=query,
key=permuted_feature,
value=permuted_feature,
attn_mask=attention_mask,
need_weights=False)
# (batch, seq, dim)
lstm_out = att_output.permute(1, 0, 2)
# (batch, 1, dim)
summary_vec, _ = lstm_out.max(dim=1, keepdim=True)
state = [summary_vec]
else:
packed_sequence_input = pack_padded_sequence(
feature,
clamped_sequence_lengths.data.tolist(),
enforce_sorted=False,
batch_first=True)
# encoder_out: (batch, seq_len, num_directions * hidden_size):
# h_t: (num_layers * num_directions, batch, hidden_size)
# - this dimension is valid regardless of batch_first=True!!
packed_lstm_out, state = self.encoder(packed_sequence_input,
initial_state)
# lstm_out: (batch, seq, num_directions * hidden_size)
lstm_out, _ = pad_packed_sequence(packed_lstm_out,
batch_first=True)
if self.hparams.get("layer_norm", False):
lstm_out = self.layer_norm(lstm_out)
if self.hparams.get("highway_connection", False):
c = torch.sigmoid(self.highway_C(lstm_out))
h = self.highway_H(lstm_out)
lstm_out = (1 - c) * torch.relu(h) + c * torch.relu(feature)
else:
lstm_out = F.relu(lstm_out)
if self.hparams["output_dropout"] > 0:
lstm_out = self.output_dropout(lstm_out)
y_pred = torch.squeeze(self.hidden2logit(lstm_out),
dim=-1) # (batch, seq)
return y_pred, state
def forward(self,
context_ids: TextFieldTensors,
query_ids: TextFieldTensors,
extend_context_ids: torch.Tensor,
extend_query_ids: torch.Tensor,
context_turn: torch.Tensor,
query_turn: torch.Tensor,
context_len: torch.Tensor,
query_len: torch.Tensor,
oovs_len: torch.Tensor,
rewrite_input_ids: Optional[TextFieldTensors] = None,
rewrite_target_ids: Optional[TextFieldTensors] = None,
extend_rewrite_ids: Optional[torch.Tensor] = None,
rewrite_len: Optional[torch.Tensor] = None,
metadata: Optional[List[Dict[str, Any]]] = None):
"""前向传播的过程"""
context_token_ids = context_ids[self._index_name]["tokens"]
query_token_ids = query_ids[self._index_name]["tokens"]
# get the extended context and query ids
extend_context_ids = context_token_ids + extend_context_ids.to(
dtype=torch.long)
extend_query_ids = query_token_ids + extend_query_ids.to(
dtype=torch.long)
# ---------------- 编码器计算输出 ------------
# 计算context和query的embedding
context_embed = self._get_embeddings(context_ids, context_turn)
query_embed = self._get_embeddings(query_ids, query_turn)
# 计算context和query的长度
max_context_len = context_embed.size(1)
max_query_len = query_embed.size(1)
# 计算mask
context_mask = get_mask_from_sequence_lengths(
context_len, max_length=max_context_len)
query_mask = get_mask_from_sequence_lengths(query_len,
max_length=max_query_len)
# 计算编码器输出
dialogue_embed = torch.cat([context_embed, query_embed], dim=1)
dialogue_mask = torch.cat([context_mask, query_mask], dim=1)
# 如果共享编码器参数，需要提交添加位置编码
if self._share_encoder_params:
dialogue_embed = add_positional_features(dialogue_embed)
# 编码器输出 [B, dialogue_len, encoder_output_dim]
dialogue_output = self.encoder(dialogue_embed, dialogue_mask)
if self._share_encoder_params:
for _ in range(self.encoder_num_layers - 1):
dialogue_output = self.encoder(dialogue_output, dialogue_mask)
# 计算编码结果
# [B, context_len, *]和[B, query_len, *]
context_output, query_output, dec_init_state = self._run_encoder(
dialogue_output, context_mask, query_mask)
output_dict = {"metadata": metadata}
if self.training:
rewrite_input_token_ids = rewrite_input_ids[
self._index_name]["tokens"]
# 计算rewrite的长度
max_rewrite_len = rewrite_input_token_ids.size(1)
rewrite_input_mask = get_mask_from_sequence_lengths(
rewrite_len, max_length=max_rewrite_len)
rewrite_target_ids = rewrite_target_ids[self._index_name]["tokens"]
# 计算rewrite的目标序列索引
rewrite_target_ids = rewrite_target_ids + extend_rewrite_ids.to(
dtype=torch.long)
# 计算embedding输出，[B, rewrite_len, embedding_size]
rewrite_embed = self._get_embeddings(rewrite_input_ids)
# 前向传播计算loss
new_output_dict = self._forward_step(
context_output, query_output, context_mask, query_mask,
rewrite_embed, rewrite_target_ids, rewrite_len,
rewrite_input_mask, extend_context_ids, extend_query_ids,
oovs_len, dec_init_state)
output_dict.update(new_output_dict)
else:
batch_hyps = self._run_inference(context_output,
query_output,
context_mask,
query_mask,
extend_context_ids,
extend_query_ids,
oovs_len,
dec_init_state=dec_init_state)
# get the result of each instance
output_dict['hypothesis'] = batch_hyps
output_dict = self.get_rewrite_string(output_dict)
output_dict["loss"] = torch.tensor(0)
return output_dict
def _eval_decode(
self, state: Dict[str, torch.Tensor],
segments: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
encoder_outputs = state["encoder_outputs"]
source_key_padding_mask = state["source_key_padding_mask"]
source_embedded = state["source_raw_embedded"]
source_token_mask = state["source_token_mask"]
memory_key_padding_mask = (1 - source_token_mask).bool()
# memory_key_padding_mask = source_key_padding_mask
batch_size = source_key_padding_mask.size(0)
encode_length = source_key_padding_mask.size(1)
log_probs_after_end = encoder_outputs.new_full(
(batch_size, self._num_classes + encode_length),
fill_value=float("-inf"))
log_probs_after_end[:, self._end_index] = 0.
start_predictions = state["source_mask"].new_full(
(batch_size, 1), fill_value=self._start_index)
partial_generate_predictions = start_predictions
partial_copy_predictions = state["source_mask"].new_zeros(
(batch_size, 1))
basic_index = torch.arange(batch_size).to(
source_embedded.device).unsqueeze(1).long()
generate_mask = state["source_mask"].new_ones((batch_size, 1)).float()
# shape: (batch_size)
last_prediction = start_predictions.squeeze(1)
for _ in range(self._max_decoding_step):
# shape: (batch_size, partial_len, d_model)
partial_source_embedded_input = source_embedded[
basic_index, partial_copy_predictions]
partial_target_embedded_input = self._target_embedder(
partial_generate_predictions)
partial_embedded_input = partial_target_embedded_input * generate_mask.unsqueeze(-1) \
+ partial_source_embedded_input * (1 - generate_mask).unsqueeze(-1)
partial_embedded_input = util.add_positional_features(
partial_embedded_input)
partial_len = partial_embedded_input.size(1)
partial_embedded_input = partial_embedded_input.permute(1, 0, 2)
mask = (torch.triu(state["source_mask"].new_ones(
partial_len, partial_len)) == 1).transpose(0, 1)
mask = mask.float().masked_fill(mask == 0,
float('-inf')).masked_fill(
mask == 1, float(0.0))
if not self._decode_use_relative_position:
# shape: (partial_len, batch_size, d_model)
outputs = self._decoder(
partial_embedded_input,
memory=encoder_outputs,
tgt_mask=mask,
memory_key_padding_mask=memory_key_padding_mask)
else:
# gnn decoder
edge_mask = get_decode_edge_mask(
partial_embedded_input,
max_decode_clip_range=self._max_decode_clip_range)
tgt_padding_mask = torch.tril(edge_mask.new_ones(
[partial_len, partial_len]),
diagonal=0)
tgt_padding_mask = (1 - tgt_padding_mask.unsqueeze(0).repeat(
batch_size, 1, 1)).float()
# shape: (partial_len, batch_size, d_model)
outputs = self._decoder(
partial_embedded_input,
edge_mask=edge_mask.permute(0, 2, 3, 1),
memory=encoder_outputs,
tgt_padding_mask=tgt_padding_mask,
memory_key_padding_mask=memory_key_padding_mask)
outputs = outputs.permute(1, 0, 2)
# shape: (batch_size, d_model)
curr_outputs = outputs[:, -1, :]
# shape: (batch_size, num_classes)
generate_scores = self.get_generate_scores(curr_outputs)
# shape: (batch_size, encode_length)
copy_scores = self.get_copy_scores(
state, curr_outputs.unsqueeze(1)).squeeze(1)
# Gate
# shape: (batch_size, 1)
generate_gate = F.sigmoid(self.gate_linear(curr_outputs))
copy_gate = 1 - generate_gate
scores = torch.cat(
(generate_scores * generate_gate, copy_scores * copy_gate),
dim=-1)
# scores = torch.cat((generate_scores, copy_scores), dim=-1)
# shape: (batch_size, encode_length)
entity_mask = 1 - (
(segments['tokens'] == self._token_index) |
(segments['tokens'] == self._non_func_symbol_index) |
(segments['tokens'] == self._segment_pad_index)).float()
# shape: (batch_size, num_classes + encode_length)
score_mask = torch.cat((entity_mask.new_ones(
(batch_size, self._num_classes)), entity_mask),
dim=-1)
# shape: (batch_size, num_classes + encode_length)
normalized_scores = util.masked_softmax(scores,
mask=score_mask,
dim=-1)
last_prediction_expanded = last_prediction.unsqueeze(-1).expand(
batch_size, self._num_classes + encode_length)
# shape: (batch_size, num_classes + encode_length)
cleaned_logits = torch.where(
last_prediction_expanded == self._end_index,
log_probs_after_end, normalized_scores)
# shape: (batch_size)
_, predicted = torch.max(input=cleaned_logits,
dim=1,
keepdim=False)
copy_mask = (predicted >= self._num_classes).long()
generate_predicted = predicted * (1 - copy_mask)
copy_predicted = (predicted - self._num_classes) * copy_mask
partial_copy_predictions = torch.cat(
(partial_copy_predictions, copy_predicted.unsqueeze(1)), dim=1)
partial_generate_predictions = torch.cat(
(partial_generate_predictions,
generate_predicted.unsqueeze(1)),
dim=1)
generate_mask = torch.cat(
(generate_mask, (1 - copy_mask).unsqueeze(1).float()), dim=1)
last_prediction = predicted
if (last_prediction == self._end_index).sum() == batch_size:
break
predictions = partial_generate_predictions * generate_mask.long() + \
(1 - generate_mask).long() * (partial_copy_predictions + self._num_classes)
# shape: (batch_size, partial_len)
output_dict = {"predictions": predictions}
return output_dict
def _train_decode(
self, state: Dict[str, torch.Tensor],
target_tokens: [str, torch.Tensor],
generate_targets: torch.Tensor) -> Dict[str, torch.Tensor]:
encoder_outputs = state["encoder_outputs"]
source_key_padding_mask = state["source_key_padding_mask"]
# shape: (batch_size, encode_length, d_model)
source_embedded = state["source_raw_embedded"]
batch_size, _, _ = source_embedded.size()
basic_index = torch.arange(batch_size).to(
source_embedded.device).long()
generate_targets = generate_targets.long()
retrieved_target_embedded_input = source_embedded[
basic_index.unsqueeze(1), generate_targets][:, :-1, :]
target_embedded_input = self._target_embedder(
target_tokens['tokens'])[:, :-1, :]
# shape: (batch_size, max_decode_length)
# where 1 indicates that the target token is generated rather than copied
generate_mask = (generate_targets == 0).float()
target_embedded_input = target_embedded_input * generate_mask[:, :-1].unsqueeze(-1) \
+ retrieved_target_embedded_input * (1 - generate_mask)[:, :-1].unsqueeze(-1)
target_embedded_input = util.add_positional_features(
target_embedded_input)
# shape: (max_target_sequence_length - 1, batch_size, d_model)
target_embedded_input = target_embedded_input.permute(1, 0, 2)
# shape: (batch_size, max_target_sequence_length - 1)
"""
key_padding_mask should be a ByteTensor where True values are positions
that should be masked with float('-inf') and False values will be unchanged.
"""
target_mask = util.get_text_field_mask(target_tokens)[:, 1:]
target_key_padding_mask = (1 - target_mask.byte()).bool()
assert target_key_padding_mask.size(1) == target_embedded_input.size(0) and \
target_embedded_input.size(1) == target_key_padding_mask.size(0)
max_target_seq_length = target_key_padding_mask.size(1)
target_additive_mask = (torch.triu(
target_mask.new_ones(max_target_seq_length,
max_target_seq_length)) == 1).transpose(0, 1)
target_additive_mask = target_additive_mask.float().masked_fill(
target_additive_mask == 0, float('-inf'))
target_additive_mask = target_additive_mask.masked_fill(
target_additive_mask == 1, float(0.0))
assert target_embedded_input.size(1) == encoder_outputs.size(1)
source_token_mask = state["source_token_mask"]
memory_key_padding_mask = (1 - source_token_mask).bool()
# memory_key_padding_mask = source_key_padding_mask
if not self._decode_use_relative_position:
# shape: (max_target_sequence_length, batch_size, d_model)
decoder_outputs = self._decoder(
target_embedded_input,
memory=encoder_outputs,
tgt_mask=target_additive_mask,
tgt_key_padding_mask=None,
memory_key_padding_mask=memory_key_padding_mask)
else:
# gnn decoder
edge_mask = get_decode_edge_mask(
target_embedded_input,
max_decode_clip_range=self._max_decode_clip_range)
batch_size = edge_mask.size(0)
tgt_padding_mask = torch.tril(edge_mask.new_ones(
[max_target_seq_length, max_target_seq_length]),
diagonal=0)
tgt_padding_mask = (1 - (tgt_padding_mask.unsqueeze(0).repeat(
batch_size, 1, 1))).float()
decoder_outputs = self._decoder(
target_embedded_input,
edge_mask=edge_mask.permute(0, 2, 3, 1),
memory=encoder_outputs,
tgt_padding_mask=tgt_padding_mask,
memory_key_padding_mask=memory_key_padding_mask)
# shape: (batch_size, max_target_sequence_length, d_model)
decoder_outputs = decoder_outputs.permute(1, 0, 2)
state.update({
"decoder_outputs": decoder_outputs,
"target_key_padding_mask": target_key_padding_mask,
"target_mask": target_mask,
"generate_mask": generate_mask
})
return state
def forward(
self,
inputs: torch.Tensor,
semantic_views_q: torch.Tensor,
semantic_views_scope_mask: torch.Tensor,
mask: torch.Tensor = None,
return_output_metadata: bool = False
) -> torch.Tensor: # pylint: disable=arguments-differ
if self._use_positional_encoding:
output = add_positional_features(inputs)
else:
output = inputs
total_sublayers = len(self._conv_layers) + 2
sublayer_count = 0
for conv_norm_layer, conv_layer in zip(self._conv_norm_layers,
self._conv_layers):
conv_norm_out = self.dropout(conv_norm_layer(output))
conv_out = self.dropout(
conv_layer(conv_norm_out.transpose_(1, 2)).transpose_(1, 2))
sublayer_count += 1
output = self.residual_with_layer_dropout(output, conv_out,
sublayer_count,
total_sublayers)
attention_norm_out = self.dropout(self.attention_norm_layer(output))
bs, seq_len, _ = inputs.shape
# pad semantic views
if semantic_views_q.shape[1] < self.num_attention_heads:
#logging.info("semantic_views_q.dtype:{0}".format(semantic_views_q.dtype))
# #
# logging.info("semantic_views_q_pad.dtype:{0}".format(semantic_views_q_pad.dtype))
semantic_views_q = torch.cat([
semantic_views_q,
long_fill(
(bs, self.num_attention_heads - semantic_views_q.shape[1],
seq_len), 0, semantic_views_q.is_cuda)
],
dim=1)
#logging.info("semantic_views_q.dtype after concat:{0}".format(semantic_views_q.dtype))
#logging.info("semantic_views_scope_mask.dtype:{0}".format(semantic_views_scope_mask.dtype))
semantic_views_scope_mask = torch.cat([
semantic_views_scope_mask,
long_fill((bs, self.num_attention_heads -
semantic_views_scope_mask.shape[1], seq_len), 0,
semantic_views_scope_mask.is_cuda)
],
dim=1)
if self._replace_zero_semantic_labels_with_per_head_labels:
heads_dim_id = 1
views_shape = list(semantic_views_q.shape)
views_shape_single_head = views_shape[:]
views_shape_single_head[heads_dim_id] = 1
attention_head_mask = torch.cat([
long_fill(views_shape_single_head, x, semantic_views_q.is_cuda)
for x in range(
self.num_semantic_labels -
self.num_attention_heads, self.num_semantic_labels)
],
dim=heads_dim_id).contiguous()
# mask the values
semantic_views_q = semantic_views_q + attention_head_mask * (
semantic_views_q < 1).long()
if len(semantic_views_scope_mask.shape) == len(
semantic_views_q.shape):
semantic_views_scope_mask = semantic_views_scope_mask + attention_head_mask * (
semantic_views_scope_mask < 1).long()
attention_output_meta = None
if is_output_meta_supported(self.attention_layer):
attention_out, attention_output_meta = self.attention_layer(
attention_norm_out,
semantic_views_q,
semantic_views_scope_mask,
mask,
return_output_metadata=return_output_metadata)
attention_out = self.dropout(attention_out)
else:
attention_out = self.dropout(
self.attention_layer(attention_norm_out, semantic_views_q,
semantic_views_scope_mask, mask))
sublayer_count += 1
output = self.residual_with_layer_dropout(output, attention_out,
sublayer_count,
total_sublayers)
feedforward_norm_out = self.dropout(
self.feedforward_norm_layer(output))
feedforward_out = self.dropout(self.feedforward(feedforward_norm_out))
sublayer_count += 1
output = self.residual_with_layer_dropout(output, feedforward_out,
sublayer_count,
total_sublayers)
return output, attention_output_meta
def forward(
self,
inputs: torch.Tensor,
semantic_views_q: torch.Tensor,
semantic_views_k: torch.Tensor,
mask: torch.Tensor = None
) -> torch.Tensor: # pylint: disable=arguments-differ
if self._use_positional_encoding:
output = add_positional_features(inputs)
else:
output = inputs
total_sublayers = len(self._conv_layers) + 2
sublayer_count = 0
for conv_norm_layer, conv_layer in zip(self._conv_norm_layers,
self._conv_layers):
conv_norm_out = self.dropout(conv_norm_layer(output))
conv_out = self.dropout(
conv_layer(conv_norm_out.transpose_(1, 2)).transpose_(1, 2))
sublayer_count += 1
output = self.residual_with_layer_dropout(output, conv_out,
sublayer_count,
total_sublayers)
attention_norm_out = self.dropout(self.attention_norm_layer(output))
if self._replace_zero_semantic_labels_with_per_head_labels:
heads_dim_id = 1
views_shape = list(semantic_views_q.shape)
views_shape_single_head = views_shape[:]
views_shape_single_head[heads_dim_id] = 1
attention_head_mask = torch.cat([
torch.full(views_shape_single_head, x).long() for x in range(
self.num_semantic_labels -
self.num_attention_heads, self.num_semantic_labels)
],
dim=heads_dim_id).contiguous()
if semantic_views_q.is_cuda:
attention_head_mask = attention_head_mask.cuda()
# semantic_views_q_zeros = (semantic_views_q < 1).long()
#
# logging.info("type(semantic_views_q):{0}".format(semantic_views_q.type()))
# logging.info("type(attention_head_mask):{0}".format(attention_head_mask.type()))
# logging.info("type(semantic_views_q_zeros):{0}".format(semantic_views_q_zeros.type()))
# mask the values
semantic_views_q = semantic_views_q + attention_head_mask * (
semantic_views_q < 1).long()
semantic_views_k = semantic_views_k + attention_head_mask * (
semantic_views_k < 1).long()
attention_out = self.dropout(
self.attention_layer(attention_norm_out, semantic_views_q,
semantic_views_k, mask))
sublayer_count += 1
output = self.residual_with_layer_dropout(output, attention_out,
sublayer_count,
total_sublayers)
feedforward_norm_out = self.dropout(
self.feedforward_norm_layer(output))
feedforward_out = self.dropout(self.feedforward(feedforward_norm_out))
sublayer_count += 1
output = self.residual_with_layer_dropout(output, feedforward_out,
sublayer_count,
total_sublayers)
return output
def forward(self,
inputs: torch.Tensor,
source_memory_bank: torch.Tensor,
source_mask: torch.Tensor,
target_mask: torch.Tensor,
is_train: bool = True) -> Dict:
batch_size, source_seq_length, _ = source_memory_bank.size()
__, target_seq_length, __ = inputs.size()
source_padding_mask = None
target_padding_mask = None
if source_mask is not None:
source_padding_mask = ~source_mask.bool()
if target_mask is not None:
target_padding_mask = ~target_mask.bool()
# project to correct dimensionality
outputs = self.input_proj_layer(inputs)
# add pos encoding feats
outputs = add_positional_features(outputs)
# swap to pytorch's batch-second convention
outputs = outputs.permute(1, 0, 2)
source_memory_bank = source_memory_bank.permute(1, 0, 2)
# get a mask
ar_mask = self.make_autoregressive_mask(outputs.shape[0]).to(source_memory_bank.device)
for i in range(len(self.layers)):
outputs , __, __ = self.layers[i](outputs,
source_memory_bank,
tgt_mask=ar_mask,
#memory_mask=None,
tgt_key_padding_mask=target_padding_mask,
memory_key_padding_mask=source_padding_mask
)
# do final norm here
if self.prenorm:
outputs = self.final_norm(outputs)
# switch back from pytorch's absolutely moronic batch-second convention
outputs = outputs.permute(1, 0, 2)
source_memory_bank = source_memory_bank.permute(1, 0, 2)
if not self.use_coverage:
source_attention_output = self.source_attn_layer(outputs,
source_memory_bank,
source_mask,
None)
attentional_tensors = self.dropout(source_attention_output['attentional'])
source_attention_weights = source_attention_output['attention_weights']
coverage_history = None
else:
# need to do step by step because running sum of coverage
source_attention_weights = []
attentional_tensors = []
# init to zeros
coverage = inputs.new_zeros(size=(batch_size, 1, source_seq_length))
coverage_history = []
for timestep in range(outputs.shape[1]):
output = outputs[:,timestep,:].unsqueeze(1)
source_attention_output = self.source_attn_layer(output,
source_memory_bank,
source_mask,
coverage)
attentional_tensors.append(source_attention_output['attentional'])
source_attention_weights.append(source_attention_output['attention_weights'])
coverage = source_attention_output['coverage']
coverage_history.append(coverage)
# [batch_size, tgt_seq_len, hidden_dim]
attentional_tensors = self.dropout(torch.cat(attentional_tensors, dim=1))
# [batch_size, tgt_seq_len, src_seq_len]
source_attention_weights = torch.cat(source_attention_weights, dim=1)
coverage_history = torch.cat(coverage_history, dim=1)
if is_train:
tgt_attn_list = []
for timestep in range(attentional_tensors.shape[1]):
bsz, seq_len, __ = attentional_tensors.shape
attn_mask = torch.ones((bsz, seq_len))
attn_mask[:,timestep:] = 0
attn_mask = attn_mask.to(attentional_tensors.device)
target_attention_output = self.target_attn_layer(attentional_tensors[:,timestep,:].unsqueeze(1),
attentional_tensors,
mask = attn_mask)
if timestep == 0:
# zero out weights at 0, effectively banning target copy since there is nothing to copy
tgt_attn_list.append(torch.zeros_like(target_attention_output["attention_weights"][:,-1,:].unsqueeze(1)))
else:
tgt_attn_list.append(target_attention_output["attention_weights"])
target_attention_weights = torch.cat(tgt_attn_list, dim=1)
else:
target_attention_output = self.target_attn_layer(attentional_tensors,
attentional_tensors,
mask = target_padding_mask)
target_attention_weights = target_attention_output['attention_weights']
return dict(
outputs=outputs,
output=outputs[:,-1,:].unsqueeze(1),
attentional_tensors=attentional_tensors,
attentional_tensor=attentional_tensors[:,-1,:].unsqueeze(1),
target_attention_weights = target_attention_weights,
source_attention_weights = source_attention_weights,
coverage_history = coverage_history,
)
def forward(self,
context: Dict[str, torch.LongTensor],
length: torch.LongTensor = None,
repeat: torch.FloatTensor = None,
label: torch.LongTensor = None) -> Dict[str, torch.Tensor]:
#expected_dim = (self.final_classifier_feedforward.get_input_dim() + 1) / 2
expected_dim = (self.final_classifier_feedforward.get_input_dim() +
2) / 2
dia_len = context['tokens'].size()[1]
if expected_dim - dia_len > 0:
padding = torch.zeros([
context['tokens'].size()[0], (expected_dim - dia_len),
context['tokens'].size()[2]
]).long()
context['tokens'] = torch.cat([context['tokens'], padding], dim=1)
# context: batch_size * dials_len * sentences_len
# embedded_context: batch_size * dials_len * sentences_len * emb_dim
embedded_context = self.text_field_embedder(context)
# utterances_mask: batch_size * dials_len * sentences_len
utterances_mask = get_text_field_mask(context, 1).float()
# encoded_utterances: batch_size * dials_len * emb_dim
encoded_utterances = self.utterances_encoder(embedded_context,
utterances_mask)
encoded_utterances = add_positional_features(encoded_utterances)
#projected_utterances = self.attend_feedforward(encoded_utterances)
# similarity_matrix: batch_size * dials_len * output_dim
mask = get_text_field_mask(context).float()
similarity_matrix = self.matrix_attention(encoded_utterances, mask)
# attended_context: batch * (dials_len - 1) * emb_dim
attended_context = similarity_matrix[:, :-1, :]
# attended_response: batch * (dials_len - 1) * emb_dim
attended_response = similarity_matrix[:, 1:, :]
# embedded_context: batch_size * (dials_len - 1) * emb_dim
embedded_context = encoded_utterances[:, :-1, :]
# embedded_response: batch_size * (dials_len - 1) * emb_dim
embedded_response = encoded_utterances[:, 1:, :]
# context_compare_input: batch * (dials_len - 1) * (emb_dim + emb_dim)
context_compare_input = torch.cat(
[embedded_context, attended_response], dim=-1)
# response_compare_input: batch * (dials_len - 1) * (emb_dim + emb_dim)
response_compare_input = torch.cat(
[embedded_response, attended_context], dim=-1)
# compared_context: batch * (dials_len - 1) * emb_dim
compared_context = self.compare_feedforward(context_compare_input)
compared_context = compared_context * mask[:, :-1].unsqueeze(-1)
# compared_response: batch * (dials_len - 1) * emb_dim
compared_response = self.compare_feedforward(response_compare_input)
compared_response = compared_response * mask[:, 1:].unsqueeze(-1)
# aggregate_input: batch * (dials_len - 1) * (compare_context_dim + compared_response_dim)
aggregate_input = torch.cat([compared_context, compared_response],
dim=-1)
# class_logits & class_probs: batch * (dials_len - 1) * 2
class_logits = self.classifier_feedforward(aggregate_input)
class_probs = F.softmax(class_logits,
dim=-1).reshape(class_logits.size()[0], -1)
length_tensor = torch.FloatTensor(length).reshape(-1, 1)
repeat_tensor = torch.FloatTensor(repeat).reshape(-1, 1)
#class_probs = torch.cat([class_probs, length_tensor, repeat_tensor], dim=1)
#class_probs = torch.cat([class_probs, repeat_tensor], dim=1)
full_logits = self.final_classifier_feedforward(class_probs)
full_probs = F.softmax(full_logits, dim=-1)
output_dict = {
"class_logits": full_logits,
"class_probabilities": full_probs
}
if label is not None:
loss = self.loss(full_logits, label.squeeze(-1))
for metric in self.metrics.values():
metric(full_logits, label.squeeze(-1))
output_dict['loss'] = loss
return output_dict
