/
seq2seq_base.py
375 lines (321 loc) · 16.7 KB
/
seq2seq_base.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
from typing import Dict, List, Optional
from allennlp.data import Vocabulary
from allennlp.models.encoder_decoders import SimpleSeq2Seq as AllenNlpSimpleSeq2Seq
from allennlp.modules.attention import DotProductAttention
from allennlp.modules.seq2seq_encoders import PytorchSeq2SeqWrapper
from allennlp.modules.text_field_embedders import BasicTextFieldEmbedder
from allennlp.modules.token_embedders import Embedding
from allennlp.nn.util import add_sentence_boundary_token_ids, sequence_cross_entropy_with_logits
from allennlp.training.metrics import Average, SequenceAccuracy, UnigramRecall
import torch
from torch import nn
from torch.nn import functional as F
class Seq2SeqBase(AllenNlpSimpleSeq2Seq):
r"""
A wrapper over AllenNLP's SimpleSeq2Seq class. This serves as a base class for the
:class:`~probnmn.models.program_generator.ProgramGenerator` and
:class:`~probnmn.models.question_reconstructor.QuestionReconstructor`. The key differences
from super class are:
1. This class doesn't use beam search, it performs categorical sampling or greedy decoding
as explicitly passed on :meth:`forward` call.
2. This class records four metrics: perplexity, sequence_accuracy, word error rate and
BLEU score.
3. Has sensible defaults for super class (dot-product attention, embedding etc.).
Parameters
----------
vocabulary: allennlp.data.vocabulary.Vocabulary
AllenNLP's vocabulary. This vocabulary has three namespaces - "questions", "programs" and
"answers", which contain respective token to integer mappings.
source_namespace: str, required
Namespace for source tokens,
"programs" for :class:`~probnmn.models.question_reconstructor.QuestionReconstructor` and
"questions" for :class:`~probnmn.models.program_generator.ProgramGenerator`.
target_namespace: str, required
Namespace for target tokens, "programs" for ProgramGenerator and "questions" for
QuestionReconstructor.
input_size : int, optional (default = 256)
The dimension of the inputs to the LSTM.
hidden_size : int, optional (default = 256)
The dimension of the outputs of the LSTM.
num_layers: int, optional (default = 2)
Number of recurrent layers of the LSTM.
"""
def __init__(
self,
vocabulary: Vocabulary,
source_namespace: str,
target_namespace: str,
input_size: int = 256,
hidden_size: int = 256,
num_layers: int = 2,
dropout: float = 0.0,
max_decoding_steps: int = 30,
):
# @@PADDING@@, @@UNKNOWN@@, @start@, @end@ have same indices in all namespaces.
self._pad_index = vocabulary.get_token_index("@@PADDING@@", namespace=source_namespace)
self._unk_index = vocabulary.get_token_index("@@UNKNOWN@@", namespace=source_namespace)
self._end_index = vocabulary.get_token_index("@end@", namespace=source_namespace)
self._start_index = vocabulary.get_token_index("@start@", namespace=source_namespace)
# Short-hand notations.
__source_vocab_size = vocabulary.get_vocab_size(namespace=source_namespace)
__target_vocab_size = vocabulary.get_vocab_size(namespace=target_namespace)
# Source embedder converts tokenized source sequences to dense embeddings.
__source_embedder = BasicTextFieldEmbedder(
{"tokens": Embedding(__source_vocab_size, input_size, padding_index=self._pad_index)}
)
# Encodes the sequence of source embeddings into a sequence of hidden states.
__encoder = PytorchSeq2SeqWrapper(
nn.LSTM(input_size, hidden_size, num_layers, dropout=dropout, batch_first=True)
)
# Attention mechanism between decoder context and encoder hidden states at each time step.
__attention = DotProductAttention()
super().__init__(
vocabulary,
source_embedder=__source_embedder,
encoder=__encoder,
max_decoding_steps=max_decoding_steps,
attention=__attention,
target_namespace=target_namespace,
use_bleu=True,
)
# Record four metrics - perplexity, sequence accuracy, word error rate and BLEU score.
# super().__init__() already declared "self._bleu",
# perplexity = 2 ** average_val_loss
# word error rate = 1 - unigram recall
self._log2_perplexity = Average()
self._sequence_accuracy = SequenceAccuracy()
self._unigram_recall = UnigramRecall()
def forward(
self,
source_tokens: torch.LongTensor,
target_tokens: Optional[torch.LongTensor] = None,
decoding_strategy: str = "sampling",
) -> Dict[str, torch.Tensor]:
r"""
Override AllenNLP's forward, changing decoder logic. Perform either categorical sampling
or greedy decoding as per specified.
Parameters
----------
source_tokens: torch.LongTensor
Tokenized source sequences padded to maximum length. These are not padded with
@start@ and @end@ sentence boundaries. Shape: (batch_size, max_source_length)
target_tokens: torch.LongTensor, optional (default = None)
Tokenized target sequences padded to maximum length. These are not padded with
@start@ and @end@ sentence boundaries. Shape: (batch_size, max_target_length)
decoding_strategy: str, optional (default = "sampling")
How to perform decoding? One of "sampling" or "greedy".
Returns
-------
Dict[str, torch.Tensor]
"""
# Add "@start@" and "@end@" tokens to source and target sequences.
source_tokens, _ = add_sentence_boundary_token_ids(
source_tokens, (source_tokens != self._pad_index), self._start_index, self._end_index
)
if target_tokens is not None:
target_tokens, _ = add_sentence_boundary_token_ids(
target_tokens,
(target_tokens != self._pad_index),
self._start_index,
self._end_index,
)
# Remove "@start@" from source sequences anyway (it's being encoded).
source_tokens = {"tokens": source_tokens[:, 1:]}
if target_tokens is not None:
target_tokens = {"tokens": target_tokens}
# _encode and _init_decoder_state are super class methods, left untouched.
# keys: {"encoder_outputs", "source_mask"}
state = self._encode(source_tokens)
# keys: {"encoder_outputs", "source_mask", "decoder_hidden", "decoder_context"}
state = self._init_decoder_state(state)
# The `_forward_loop` decodes the input sequence and computes the loss during training
# and validation.
# keys: {"predictions", "loss"}
output_dict = self._forward_loop(state, target_tokens, decoding_strategy)
return output_dict
def _forward_loop(
self,
state: Dict[str, torch.FloatTensor],
target_tokens: Dict[str, torch.LongTensor] = None,
decoding_strategy: str = "sampling",
) -> Dict[str, torch.Tensor]:
# shape: (batch_size, max_input_sequence_length)
source_mask = state["source_mask"]
batch_size = source_mask.size()[0]
if target_tokens:
# shape: (batch_size, max_target_sequence_length)
targets = target_tokens["tokens"]
_, target_sequence_length = targets.size()
# The last input from the question is either padding or the end symbol.
# Either way, we don't have to process it.
num_decoding_steps = target_sequence_length - 1
else:
num_decoding_steps = self._max_decoding_steps
# Initialize question predictions with the start index.
# shape: (batch_size,)
last_predictions = source_mask.new_full((batch_size,), fill_value=self._start_index)
step_logits: List[torch.Tensor] = []
step_logprobs: List[torch.Tensor] = []
step_predictions: List[torch.Tensor] = []
for timestep in range(num_decoding_steps):
if self.training and torch.rand(1).item() < self._scheduled_sampling_ratio:
# Use gold tokens at test time and at a rate of 1 - _scheduled_sampling_ratio
# during training.
# shape: (batch_size,)
input_choices = last_predictions
elif not target_tokens:
# shape: (batch_size,)
input_choices = last_predictions
else:
# shape: (batch_size,)
input_choices = targets[:, timestep]
# shape: (batch_size, num_classes)
output_projections, state = self._prepare_output_projections(input_choices, state)
# shape: (batch_size, num_classes)
class_probabilities = F.softmax(output_projections, dim=-1)
class_logprobs = F.log_softmax(output_projections, dim=-1)
# NOTE -------------------------------------------------------------------------------
# This differs from super()._forward_loop(...)
if decoding_strategy == "greedy":
_, predicted_classes = torch.max(class_probabilities, 1)
elif decoding_strategy == "sampling":
# Perform categorical sampling, don't sample @@PADDING@@, @@UNKNOWN@@, @start@.
class_probabilities[:, self._pad_index] = 0
class_probabilities[:, self._unk_index] = 0
class_probabilities[:, self._start_index] = 0
predicted_classes = torch.multinomial(class_probabilities, 1).squeeze()
# ------------------------------------------------------------------------------------
# shape (predicted_classes): (batch_size,)
last_predictions = predicted_classes
class_logprobs = class_logprobs[torch.arange(batch_size), predicted_classes]
# List of tensors, shape: (batch_size, 1, num_classes)
step_predictions.append(last_predictions.unsqueeze(1))
step_logits.append(output_projections.unsqueeze(1))
step_logprobs.append(class_logprobs.unsqueeze(1))
# shape: (batch_size, num_decoding_steps)
predictions = torch.cat(step_predictions, 1)
# Trim predictions after first "@end@" token.
predictions = self._trim_predictions(predictions)
# Log-probabilities at each time-step (without teacher forcing).
# This will be `loss`, to compute REINFORCE reward in question coding.
# In presence of target tokens, we replace this as cross entropy loss (teacher forcing).
logprobs = torch.cat(step_logprobs, 1)
prediction_mask = (predictions != self._pad_index).float()
# Average the sequence logprob across time-steps. Hence our REINFORCE reward will be
# length-normalized sequence log-probability. This ensures equal importance of all
# sequences irrespective of their lengths.
sequence_logprobs = logprobs.mul(prediction_mask).sum(-1)
prediction_lengths = prediction_mask.sum(-1)
# shape: (batch_size, )
sequence_logprobs /= prediction_lengths + 1e-12
output_dict = {"predictions": predictions, "loss": -sequence_logprobs}
if target_tokens:
# shape: (batch_size, num_decoding_steps, num_classes)
logits = torch.cat(step_logits, 1)
# shape: (batch_size, max_sequence_length)
target_mask = targets != self._pad_index
# shape: (batch_size, )
sequence_cross_entropy = self._get_loss(logits, targets, target_mask)
output_dict["loss"] = sequence_cross_entropy
relevant_targets = targets[:, 1:]
if not self.training:
# Record BLEU, perplexity, word error rate and sequence accuracy during validation.
self._bleu(predictions, targets)
self._log2_perplexity(sequence_cross_entropy.mean().item())
# Sequence accuracy and unigram recall expect a beam dimension.
# Compare generated sequences without "@start@" token.
self._sequence_accuracy(
predictions[:, : relevant_targets.size(-1)].unsqueeze(1),
relevant_targets,
(relevant_targets != self._pad_index).long(),
)
self._unigram_recall(
predictions[:, : relevant_targets.size(-1)].unsqueeze(1),
relevant_targets,
(relevant_targets != self._pad_index).long(),
)
return output_dict
def _trim_predictions(self, predictions: torch.LongTensor):
r"""
Trim output predictions at first "@end@" and pad the rest of sequence.
This includes "@end@" as last token in trimmed sequence.
"""
# shape: (batch_size, num_decoding_steps)
trimmed_predictions = torch.zeros_like(predictions)
for i, prediction in enumerate(predictions):
prediction_indices = list(prediction.detach().cpu().numpy())
if self._end_index in prediction_indices:
end_index = prediction_indices.index(self._end_index)
if end_index > 0:
trimmed_predictions[i][: end_index + 1] = prediction[: end_index + 1]
else:
trimmed_predictions[i] = prediction
return trimmed_predictions
@staticmethod
def _get_loss(
logits: torch.LongTensor, targets: torch.LongTensor, target_mask: torch.LongTensor
):
r"""
Override AllenNLP Seq2Seq model's provided ``_get_loss`` method, which returns sequence
cross entropy averaged over batch by default. Instead, provide sequence cross entropy of
each sequence in a batch separately.
Extended Summary
----------------
From AllenNLP documentation:
Compute loss.
Takes logits (unnormalized outputs from the decoder) of size (batch_size,
num_decoding_steps, num_classes), target indices of size (batch_size, num_decoding_steps+1)
and corresponding masks of size (batch_size, num_decoding_steps+1) steps and computes
cross entropy loss while taking the mask into account.
The length of ``targets`` is expected to be greater than that of ``logits`` because the
decoder does not need to compute the output corresponding to the last timestep of
``targets``. This method aligns the inputs appropriately to compute the loss.
During training, we want the logit corresponding to timestep i to be similar to the target
token from timestep i + 1. That is, the targets should be shifted by one timestep for
appropriate comparison. Consider a single example where the target has 3 words, and
padding is to 7 tokens::
The complete sequence would correspond to <S> w1 w2 w3 <E> <P> <P>
and the mask would be 1 1 1 1 1 0 0
and let the logits be l1 l2 l3 l4 l5 l6
We actually need to compare::
the sequence w1 w2 w3 <E> <P> <P>
with masks 1 1 1 1 0 0
against l1 l2 l3 l4 l5 l6
(where the input was) <S> w1 w2 w3 <E> <P>
"""
# shape: (batch_size, num_decoding_steps)
relevant_targets = targets[:, 1:].contiguous()
# shape: (batch_size, num_decoding_steps)
relevant_mask = target_mask[:, 1:].contiguous()
return sequence_cross_entropy_with_logits(
logits, relevant_targets, relevant_mask, average=None
)
def get_metrics(self, reset: bool = True) -> Dict[str, float]:
r"""
Return recorded metrics - perplexity, sequence accuracy, word error rate, BLEU.
Parameters
----------
reset: bool, optional (default = True)
Whether to reset the accumulated metrics after retrieving them.
Returns
-------
Dict[str, float]
A dictionary with metrics::
{
"perplexity",
"sequence_accuracy",
"word_error_rate",
"BLEU"
}
"""
all_metrics: Dict[str, float] = {}
if not self.training:
all_metrics.update(self._bleu.get_metric(reset=True))
all_metrics.update(
{
"perplexity": 2 ** self._log2_perplexity.get_metric(reset=reset),
"sequence_accuracy": self._sequence_accuracy.get_metric(reset=reset),
"word_error_rate": 1 - self._unigram_recall.get_metric(reset=reset),
}
)
return all_metrics