/
model_gene.py
187 lines (141 loc) · 7.04 KB
/
model_gene.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
import torch
import operator
from queue import PriorityQueue
from torch import nn
from torch.autograd import Variable
from encoder import TypeEncoder
from decoder import CopyNetDecoder, SimpleCopyNetDecoder
from utils import timeit
from utils import to_one_hot, to_np
class TypeEncoderDecoder(nn.Module):
def __init__(self, opid2vec, opcode2idx, word2idx, args):
super(TypeEncoderDecoder, self).__init__()
self.args = args
self.opid2vec = opid2vec
self.opcode2idx = opcode2idx
self.word2idx = word2idx
self.hidden_size = args.hidden_size
self.embedding_dim = args.embedding_dim
self.decoder_type = args.decoder_type
self.vocab_size = len(self.word2idx.keys())
self.encoder = TypeEncoder(opid2vec, opcode2idx, word2idx, self.hidden_size, self.embedding_dim, self.args)
if self.decoder_type == 'simple':
self.decoder = SimpleCopyNetDecoder(self.word2idx, self.args)
elif self.decoder_type == 'copy':
self.decoder = CopyNetDecoder(self.word2idx, self.args)
else:
raise ValueError("decoder_type must be 'attn' or 'copy'")
def forward(self, dfs, values, targets=None, teacher_forcing=0.0):
batch_size = dfs.shape[0]
encoder_outputs = self.encoder(dfs, values)
s = torch.zeros((batch_size, 1)).to(self.args.device)
s[:, 0] = self.word2idx['<SEM>']
values = torch.cat([s, values], dim=1)
hidden = self.encoder.init_hidden(batch_size)
decoder_outputs, sample_idxs = self.decoder(
encoder_outputs, values, hidden, targets=targets, teacher_forcing=teacher_forcing)
return decoder_outputs, sample_idxs
def decode(self, dfs, values, trg, method='beam'):
enc_output = self.encoder(dfs, values)
batch_size = dfs.shape[0]
seq_length = enc_output.data.shape[1]
s = torch.zeros((batch_size, 1)).to(self.args.device)
s[:, 0] = self.word2idx['<SEM>']
inputs = torch.cat([s, values], dim=1)
hidden = Variable(torch.zeros(1, batch_size, self.hidden_size)).to(self.args.device)
if method == 'beam':
return self.beam_decode(hidden, enc_output, inputs, seq_length)
else:
return self.greedy_decode(trg, hidden, enc_output)
@timeit
def beam_decode(self, decoder_hiddens, encoder_outputs, inputs, seq_length):
beam_width = self.args.beam_width
topk = self.args.topk # how many sentence do you want to generate
decoded_batch = []
# decoding goes sentence by sentence
for idx in range(encoder_outputs.size(0)): # batch_size
if isinstance(decoder_hiddens, tuple): # LSTM case
decoder_hidden = (
decoder_hiddens[0][:, idx, :].unsqueeze(0), decoder_hiddens[1][:, idx, :].unsqueeze(0))
else:
decoder_hidden = decoder_hiddens[:, idx, :].unsqueeze(0)
encoder_output = encoder_outputs[idx, :, :].unsqueeze(0)
selective_read = Variable(torch.zeros(1, 1, self.hidden_size))
one_hot_input_seq = to_one_hot(inputs[idx].unsqueeze(0), self.vocab_size + seq_length)
if next(self.parameters()).is_cuda:
selective_read = selective_read.to(self.args.device)
one_hot_input_seq = one_hot_input_seq.to(self.args.device)
# Start with the start of the sentence token
decoder_input_sos = torch.tensor([self.word2idx['<SOS>']], dtype=torch.long).to(self.args.device)
# Number of sentence to generate
endnodes = []
number_required = min((topk + 1), topk - len(endnodes))
# starting node - hidden vector, previous node, word id, logp, length
node = BeamSearchNode(decoder_hidden, None, decoder_input_sos, 0, 1)
nodes = PriorityQueue()
# start the queue
nodes.put((-node.eval(), node))
qsize = 1
while True:
if qsize > 1000:
break
score, n = nodes.get()
decoder_input = n.token_id
decoder_input = decoder_input.view(1, decoder_input.size(0))
decoder_hidden = n.h
if n.token_id.item() == self.word2idx['<EOS>'] and n.prev_node != None:
endnodes.append((score, n))
# if we reached maximum # of sentences required
if len(endnodes) >= number_required:
break
else:
continue
# decode for one step using decoder
_, decoder_output, decoder_hidden, selective_read = self.decoder.step(decoder_input,
decoder_hidden,
encoder_output,
selective_read,
one_hot_input_seq)
# PUT HERE REAL BEAM SEARCH OF TOP
log_prob, indexes = torch.topk(decoder_output, beam_width)
nextnodes = []
for new_k in range(beam_width):
decoded_t = indexes[0][new_k].view(-1)
log_p = log_prob[0][new_k].item()
node = BeamSearchNode(decoder_hidden, n, decoded_t, n.log_prob + log_p, n.length + 1)
score = -node.eval()
nextnodes.append((score, node))
# put them into queue
for i in range(len(nextnodes)):
score, nn = nextnodes[i]
nodes.put((score, nn))
# increase qsize
qsize += len(nextnodes) - 1
# choose nbest paths, back trace them
if len(endnodes) == 0:
endnodes = [nodes.get() for _ in range(topk)]
utterances = []
for score, n in sorted(endnodes, key=operator.itemgetter(0)):
utterance = []
utterance.append(to_np(n.token_id)[0])
while n.prev_node != None:
n = n.prev_node
utterance.append(to_np(n.token_id)[0])
utterance = utterance[::-1][:self.args.max_length_output]
utterances.append(utterance)
decoded_batch.append(utterances)
return decoded_batch
class BeamSearchNode(object):
def __init__(self, hidden, prev_node, token_id, log_prob, length):
self.h = hidden
self.prev_node = prev_node
self.token_id = token_id
self.log_prob = log_prob
self.length = length
def eval(self, alpha=1.0):
reward = 0
return self.log_prob / float(self.length - 1 + 1e-6) + alpha * reward
def __lt__(self, other):
return self.length < other.length
def __gt__(self, other):
return self.length > other.length