-
Notifications
You must be signed in to change notification settings - Fork 2
/
NMT.py
138 lines (105 loc) · 5.28 KB
/
NMT.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
#!/usr/bin/env python
#-*- coding:utf-8 -*-
# author:Darksoul
# datetime:11/24/2018 22:02
# software: PyCharm
from network import *
from torch import optim
from torch.nn.utils import clip_grad_norm_
from random import randint
from tqdm import tqdm
# import loss func
import masked_cross_entropy
import numpy as np
def nmt_training(src, tgt, pairs, test_src, test_tgt, test_pairs):
num_batch = len(pairs) // cfg.batch_size
encoder_test = Encoder(src.num, cfg.embed_size, cfg.hidden_size, cfg.n_layers_encoder, dropout=cfg.dropout)
decoder_test = Decoder(cfg.embed_size, cfg.hidden_size, tgt.num, cfg.n_layers_decoder, dropout=cfg.dropout)
net = Seq2Seq(encoder_test,decoder_test).cuda()
net = load_checkpoint(net, cfg)
opt = optim.Adam(net.parameters(), cfg.lr)
total_loss = []
for step in range(1, cfg.iteration-cfg.load_checkpoint):
tmp_loss = 0
for batch_index in range(num_batch):
input_batches, input_lengths, \
target_batches, target_lengths = random_batch(src, tgt, pairs, cfg.batch_size, batch_index)
opt.zero_grad()
output = net(input_batches, input_lengths, target_batches, target_lengths)
# mask loss
if cfg.loss_type == 'mask':
loss = masked_cross_entropy.compute_loss(
output.transpose(0, 1).contiguous(),
target_batches.transpose(0, 1).contiguous(),
target_lengths, ignore_index=cfg.PAD_idx
)
else:
loss = F.nll_loss(output[1:].view(-1, tgt.num),
target_batches[1:].contiguous().view(-1),
ignore_index=cfg.PAD_idx)
tmp_loss += loss.item()
total_loss.append(loss.item())
clip_grad_norm_(net.parameters(), cfg.grad_clip)
loss.backward()
opt.step()
if (step + cfg.load_checkpoint) % cfg.save_iteration == 0:
test_idx = randint(0, cfg.batch_size-1)
with open('./loss_log_train.txt', 'w') as outfile:
for item in total_loss:
outfile.write("%s\n" % item)
print("Epoch: {}, Loss: {}".format(step + cfg.load_checkpoint, tmp_loss/num_batch))
save_checkpoint(net, cfg, step + cfg.load_checkpoint)
_, pred = net.inference(input_batches[:, test_idx].reshape(input_lengths[0].item(), 1),
input_lengths[0].reshape(1))
try:
inp = ' '.join([src.idx2w[t] for t in input_batches[:,test_idx].cpu().numpy() if t != PAD_idx])
pred = ' '.join([tgt.idx2w[t] for t in pred if t != PAD_idx])
gt = ' '.join([tgt.idx2w[t] for t in target_batches[:,test_idx].cpu().numpy() if t != PAD_idx])
print("Input: {}".format(inp))
print("Ground Truth: {}".format(gt))
print("Prediction: {}".format(pred))
except Exception as e:
print(e)
# print("Epoch {} finished".format(str(step)))
random.shuffle(pairs)
def nmt_testing(src, tgt, pairs, test_src, test_tgt, test_pairs):
encoder_test = Encoder(src.num, cfg.embed_size, cfg.hidden_size, cfg.n_layers_encoder, dropout=cfg.dropout)
decoder_test = Decoder(cfg.embed_size, cfg.hidden_size, tgt.num, cfg.n_layers_decoder, dropout=cfg.dropout)
net = Seq2Seq(encoder_test,decoder_test).cuda()
net = BeamSearch(net.encoder, net.decoder, cfg.beam_widths).cuda()
net = load_checkpoint(net, cfg)
# if don't want beam search, set beam width = [1]
for i in cfg.beam_widths:
blue_score = []
for index_sample in tqdm(range(len(test_pairs))):
input_batches, input_lengths, \
target_batches, target_lengths = random_batch(test_src, test_tgt, test_pairs, 1, index_sample)
for test_idx in range(1):
pred = net(input_batches[:, test_idx].reshape(input_lengths[0].item(), 1), input_lengths[0].reshape(1),
i, MAX_LENGTH)
inp = ' '.join([test_src.idx2w[t] for t in input_batches[:, test_idx].cpu().numpy()])
mt = ' '.join([test_tgt.idx2w[t] for t in pred if t != PAD_idx])
idx = mt.find('<eos>')
mt = mt[:idx + 5]
ref = ' '.join([test_tgt.idx2w[t] for t in target_batches[:, test_idx].cpu().numpy() if t != PAD_idx])
blue_score.append(bleu([mt], [[ref]], 4))
if index_sample % 100 == 0:
print(str(index_sample))
# print('INPUT:\n' + inp)
# print('REF:\n' + ref)
# print('PREDICTION:\n' + mt)
# print("------")
print(str(i) + " finished: " + str(np.mean(blue_score)))
if __name__ == '__main__':
if not os.path.exists(cfg.checkpoints_path):
os.mkdir(cfg.checkpoints_path)
src, tgt, pairs = prepareData(cfg.data_path, 'english', 'chinese')
src.trim()
tgt.trim()
test_src, test_tgt, test_pairs = prepareData('data/small_set/test.txt', 'english', 'chinese')
test_src.trim()
test_tgt.trim()
if cfg.is_training:
nmt_training(src, tgt, pairs, test_src, test_tgt, test_pairs)
else:
nmt_testing(src, tgt, pairs, test_src, test_tgt, test_pairs)