-
Notifications
You must be signed in to change notification settings - Fork 374
/
TopKDecoder.py
341 lines (288 loc) · 16.4 KB
/
TopKDecoder.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
import torch
import torch.nn.functional as F
from torch.autograd import Variable
def _inflate(tensor, times, dim):
"""
Given a tensor, 'inflates' it along the given dimension by replicating each slice specified number of times (in-place)
Args:
tensor: A :class:`Tensor` to inflate
times: number of repetitions
dim: axis for inflation (default=0)
Returns:
A :class:`Tensor`
Examples::
>> a = torch.LongTensor([[1, 2], [3, 4]])
>> a
1 2
3 4
[torch.LongTensor of size 2x2]
>> b = ._inflate(a, 2, dim=1)
>> b
1 2 1 2
3 4 3 4
[torch.LongTensor of size 2x4]
>> c = _inflate(a, 2, dim=0)
>> c
1 2
3 4
1 2
3 4
[torch.LongTensor of size 4x2]
"""
repeat_dims = [1] * tensor.dim()
repeat_dims[dim] = times
return tensor.repeat(*repeat_dims)
class TopKDecoder(torch.nn.Module):
r"""
Top-K decoding with beam search.
Args:
decoder_rnn (DecoderRNN): An object of DecoderRNN used for decoding.
k (int): Size of the beam.
Inputs: inputs, encoder_hidden, encoder_outputs, function, teacher_forcing_ratio
- **inputs** (seq_len, batch, input_size): list of sequences, whose length is the batch size and within which
each sequence is a list of token IDs. It is used for teacher forcing when provided. (default is `None`)
- **encoder_hidden** (num_layers * num_directions, batch_size, hidden_size): tensor containing the features
in the hidden state `h` of encoder. Used as the initial hidden state of the decoder.
- **encoder_outputs** (batch, seq_len, hidden_size): tensor with containing the outputs of the encoder.
Used for attention mechanism (default is `None`).
- **function** (torch.nn.Module): A function used to generate symbols from RNN hidden state
(default is `torch.nn.functional.log_softmax`).
- **teacher_forcing_ratio** (float): The probability that teacher forcing will be used. A random number is
drawn uniformly from 0-1 for every decoding token, and if the sample is smaller than the given value,
teacher forcing would be used (default is 0).
Outputs: decoder_outputs, decoder_hidden, ret_dict
- **decoder_outputs** (batch): batch-length list of tensors with size (max_length, hidden_size) containing the
outputs of the decoder.
- **decoder_hidden** (num_layers * num_directions, batch, hidden_size): tensor containing the last hidden
state of the decoder.
- **ret_dict**: dictionary containing additional information as follows {*length* : list of integers
representing lengths of output sequences, *topk_length*: list of integers representing lengths of beam search
sequences, *sequence* : list of sequences, where each sequence is a list of predicted token IDs,
*topk_sequence* : list of beam search sequences, each beam is a list of token IDs, *inputs* : target
outputs if provided for decoding}.
"""
def __init__(self, decoder_rnn, k):
super(TopKDecoder, self).__init__()
self.rnn = decoder_rnn
self.k = k
self.hidden_size = self.rnn.hidden_size
self.V = self.rnn.output_size
self.SOS = self.rnn.sos_id
self.EOS = self.rnn.eos_id
def forward(self, inputs=None, encoder_hidden=None, encoder_outputs=None, function=F.log_softmax,
teacher_forcing_ratio=0, retain_output_probs=True):
"""
Forward rnn for MAX_LENGTH steps. Look at :func:`seq2seq.models.DecoderRNN.DecoderRNN.forward_rnn` for details.
"""
inputs, batch_size, max_length = self.rnn._validate_args(inputs, encoder_hidden, encoder_outputs,
function, teacher_forcing_ratio)
self.pos_index = Variable(torch.LongTensor(range(batch_size)) * self.k).view(-1, 1)
# Inflate the initial hidden states to be of size: b*k x h
encoder_hidden = self.rnn._init_state(encoder_hidden)
if encoder_hidden is None:
hidden = None
else:
if isinstance(encoder_hidden, tuple):
hidden = tuple([_inflate(h, self.k, 1) for h in encoder_hidden])
else:
hidden = _inflate(encoder_hidden, self.k, 1)
# ... same idea for encoder_outputs and decoder_outputs
if self.rnn.use_attention:
inflated_encoder_outputs = _inflate(encoder_outputs, self.k, 0)
else:
inflated_encoder_outputs = None
# Initialize the scores; for the first step,
# ignore the inflated copies to avoid duplicate entries in the top k
sequence_scores = torch.Tensor(batch_size * self.k, 1)
sequence_scores.fill_(-float('Inf'))
sequence_scores.index_fill_(0, torch.LongTensor([i * self.k for i in range(0, batch_size)]), 0.0)
sequence_scores = Variable(sequence_scores)
# Initialize the input vector
input_var = Variable(torch.transpose(torch.LongTensor([[self.SOS] * batch_size * self.k]), 0, 1))
# Store decisions for backtracking
stored_outputs = list()
stored_scores = list()
stored_predecessors = list()
stored_emitted_symbols = list()
stored_hidden = list()
for _ in range(0, max_length):
# Run the RNN one step forward
log_softmax_output, hidden, _ = self.rnn.forward_step(input_var, hidden,
inflated_encoder_outputs, function=function)
# If doing local backprop (e.g. supervised training), retain the output layer
if retain_output_probs:
stored_outputs.append(log_softmax_output)
# To get the full sequence scores for the new candidates, add the local scores for t_i to the predecessor scores for t_(i-1)
sequence_scores = _inflate(sequence_scores, self.V, 1)
sequence_scores += log_softmax_output.squeeze(1)
scores, candidates = sequence_scores.view(batch_size, -1).topk(self.k, dim=1)
# Reshape input = (bk, 1) and sequence_scores = (bk, 1)
input_var = (candidates % self.V).view(batch_size * self.k, 1)
sequence_scores = scores.view(batch_size * self.k, 1)
# Update fields for next timestep
predecessors = (candidates / self.V + self.pos_index.expand_as(candidates)).view(batch_size * self.k, 1)
if isinstance(hidden, tuple):
hidden = tuple([h.index_select(1, predecessors.squeeze()) for h in hidden])
else:
hidden = hidden.index_select(1, predecessors.squeeze())
# Update sequence scores and erase scores for end-of-sentence symbol so that they aren't expanded
stored_scores.append(sequence_scores.clone())
eos_indices = input_var.data.eq(self.EOS)
if eos_indices.nonzero().dim() > 0:
sequence_scores.data.masked_fill_(eos_indices, -float('inf'))
# Cache results for backtracking
stored_predecessors.append(predecessors)
stored_emitted_symbols.append(input_var)
stored_hidden.append(hidden)
# Do backtracking to return the optimal values
output, h_t, h_n, s, l, p = self._backtrack(stored_outputs, stored_hidden,
stored_predecessors, stored_emitted_symbols,
stored_scores, batch_size, self.hidden_size)
# Build return objects
decoder_outputs = [step[:, 0, :] for step in output]
if isinstance(h_n, tuple):
decoder_hidden = tuple([h[:, :, 0, :] for h in h_n])
else:
decoder_hidden = h_n[:, :, 0, :]
metadata = {}
metadata['inputs'] = inputs
metadata['output'] = output
metadata['h_t'] = h_t
metadata['score'] = s
metadata['topk_length'] = l
metadata['topk_sequence'] = p
metadata['length'] = [seq_len[0] for seq_len in l]
metadata['sequence'] = [seq[0] for seq in p]
return decoder_outputs, decoder_hidden, metadata
def _backtrack(self, nw_output, nw_hidden, predecessors, symbols, scores, b, hidden_size):
"""Backtracks over batch to generate optimal k-sequences.
Args:
nw_output [(batch*k, vocab_size)] * sequence_length: A Tensor of outputs from network
nw_hidden [(num_layers, batch*k, hidden_size)] * sequence_length: A Tensor of hidden states from network
predecessors [(batch*k)] * sequence_length: A Tensor of predecessors
symbols [(batch*k)] * sequence_length: A Tensor of predicted tokens
scores [(batch*k)] * sequence_length: A Tensor containing sequence scores for every token t = [0, ... , seq_len - 1]
b: Size of the batch
hidden_size: Size of the hidden state
Returns:
output [(batch, k, vocab_size)] * sequence_length: A list of the output probabilities (p_n)
from the last layer of the RNN, for every n = [0, ... , seq_len - 1]
h_t [(batch, k, hidden_size)] * sequence_length: A list containing the output features (h_n)
from the last layer of the RNN, for every n = [0, ... , seq_len - 1]
h_n(batch, k, hidden_size): A Tensor containing the last hidden state for all top-k sequences.
score [batch, k]: A list containing the final scores for all top-k sequences
length [batch, k]: A list specifying the length of each sequence in the top-k candidates
p (batch, k, sequence_len): A Tensor containing predicted sequence
"""
lstm = isinstance(nw_hidden[0], tuple)
# initialize return variables given different types
output = list()
h_t = list()
p = list()
# Placeholder for last hidden state of top-k sequences.
# If a (top-k) sequence ends early in decoding, `h_n` contains
# its hidden state when it sees EOS. Otherwise, `h_n` contains
# the last hidden state of decoding.
if lstm:
state_size = nw_hidden[0][0].size()
h_n = tuple([torch.zeros(state_size), torch.zeros(state_size)])
else:
h_n = torch.zeros(nw_hidden[0].size())
l = [[self.rnn.max_length] * self.k for _ in range(b)] # Placeholder for lengths of top-k sequences
# Similar to `h_n`
# the last step output of the beams are not sorted
# thus they are sorted here
sorted_score, sorted_idx = scores[-1].view(b, self.k).topk(self.k)
# initialize the sequence scores with the sorted last step beam scores
s = sorted_score.clone()
batch_eos_found = [0] * b # the number of EOS found
# in the backward loop below for each batch
t = self.rnn.max_length - 1
# initialize the back pointer with the sorted order of the last step beams.
# add self.pos_index for indexing variable with b*k as the first dimension.
t_predecessors = (sorted_idx + self.pos_index.expand_as(sorted_idx)).view(b * self.k)
while t >= 0:
# Re-order the variables with the back pointer
current_output = nw_output[t].index_select(0, t_predecessors)
if lstm:
current_hidden = tuple([h.index_select(1, t_predecessors) for h in nw_hidden[t]])
else:
current_hidden = nw_hidden[t].index_select(1, t_predecessors)
current_symbol = symbols[t].index_select(0, t_predecessors)
# Re-order the back pointer of the previous step with the back pointer of
# the current step
t_predecessors = predecessors[t].index_select(0, t_predecessors).squeeze()
# This tricky block handles dropped sequences that see EOS earlier.
# The basic idea is summarized below:
#
# Terms:
# Ended sequences = sequences that see EOS early and dropped
# Survived sequences = sequences in the last step of the beams
#
# Although the ended sequences are dropped during decoding,
# their generated symbols and complete backtracking information are still
# in the backtracking variables.
# For each batch, everytime we see an EOS in the backtracking process,
# 1. If there is survived sequences in the return variables, replace
# the one with the lowest survived sequence score with the new ended
# sequences
# 2. Otherwise, replace the ended sequence with the lowest sequence
# score with the new ended sequence
#
eos_indices = symbols[t].data.squeeze(1).eq(self.EOS).nonzero()
if eos_indices.dim() > 0:
for i in range(eos_indices.size(0)-1, -1, -1):
# Indices of the EOS symbol for both variables
# with b*k as the first dimension, and b, k for
# the first two dimensions
idx = eos_indices[i]
b_idx = int(idx[0] / self.k)
# The indices of the replacing position
# according to the replacement strategy noted above
res_k_idx = self.k - (batch_eos_found[b_idx] % self.k) - 1
batch_eos_found[b_idx] += 1
res_idx = b_idx * self.k + res_k_idx
# Replace the old information in return variables
# with the new ended sequence information
t_predecessors[res_idx] = predecessors[t][idx[0]]
current_output[res_idx, :] = nw_output[t][idx[0], :]
if lstm:
current_hidden[0][:, res_idx, :] = nw_hidden[t][0][:, idx[0], :]
current_hidden[1][:, res_idx, :] = nw_hidden[t][1][:, idx[0], :]
h_n[0][:, res_idx, :] = nw_hidden[t][0][:, idx[0], :].data
h_n[1][:, res_idx, :] = nw_hidden[t][1][:, idx[0], :].data
else:
current_hidden[:, res_idx, :] = nw_hidden[t][:, idx[0], :]
h_n[:, res_idx, :] = nw_hidden[t][:, idx[0], :].data
current_symbol[res_idx, :] = symbols[t][idx[0]]
s[b_idx, res_k_idx] = scores[t][idx[0]].data[0]
l[b_idx][res_k_idx] = t + 1
# record the back tracked results
output.append(current_output)
h_t.append(current_hidden)
p.append(current_symbol)
t -= 1
# Sort and re-order again as the added ended sequences may change
# the order (very unlikely)
s, re_sorted_idx = s.topk(self.k)
for b_idx in range(b):
l[b_idx] = [l[b_idx][k_idx.item()] for k_idx in re_sorted_idx[b_idx,:]]
re_sorted_idx = (re_sorted_idx + self.pos_index.expand_as(re_sorted_idx)).view(b * self.k)
# Reverse the sequences and re-order at the same time
# It is reversed because the backtracking happens in reverse time order
output = [step.index_select(0, re_sorted_idx).view(b, self.k, -1) for step in reversed(output)]
p = [step.index_select(0, re_sorted_idx).view(b, self.k, -1) for step in reversed(p)]
if lstm:
h_t = [tuple([h.index_select(1, re_sorted_idx).view(-1, b, self.k, hidden_size) for h in step]) for step in reversed(h_t)]
h_n = tuple([h.index_select(1, re_sorted_idx.data).view(-1, b, self.k, hidden_size) for h in h_n])
else:
h_t = [step.index_select(1, re_sorted_idx).view(-1, b, self.k, hidden_size) for step in reversed(h_t)]
h_n = h_n.index_select(1, re_sorted_idx.data).view(-1, b, self.k, hidden_size)
s = s.data
return output, h_t, h_n, s, l, p
def _mask_symbol_scores(self, score, idx, masking_score=-float('inf')):
score[idx] = masking_score
def _mask(self, tensor, idx, dim=0, masking_score=-float('inf')):
if len(idx.size()) > 0:
indices = idx[:, 0]
tensor.index_fill_(dim, indices, masking_score)