-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
120 lines (94 loc) · 3.34 KB
/
main.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
import sys
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from sentence_classification import SST
from network import Network
parser = argparse.ArgumentParser()
parser.add_argument('--batchsize', type=int, default=32,
help='Batch size for mimi-batch training and evaluating. Default: 32')
parser.add_argument('--epoch', type=int, default=50,
help='Number of training epoch. Default: 50')
parser.add_argument('--mode', type=str, default='GRU',
help='Three modes for the first three tasks: \'GRU\', \'LSTM\', \'Attention\'')
args = parser.parse_args()
def train(dataloader, model, optimizer):
dataloader.restart('train', args.batchsize)
total_loss = []
total_precise = []
while True:
optimizer.zero_grad()
batch = dataloader.get_next_batch('train')
if batch is None:
break
sent = torch.from_numpy(batch['sent']).long().cuda()
sent_length = torch.from_numpy(batch['sent_length']).long().cuda()
label = torch.from_numpy(batch['label']).long().cuda()
logit, loss = model(sent, sent_length, label)
loss.backward()
optimizer.step()
total_precise.append((torch.max(logit, dim=1)[1] == label).float().mean().cpu().data.numpy())
total_loss.append(loss.cpu().data.numpy())
print('[train]loss: %f, precise: %f' % (np.mean(total_loss), np.mean(total_precise)))
def eval(dataloader, model):
'''
evaluate on dev set
'''
dataloader.restart('dev', args.batchsize, shuffle=False)
total_precise = []
while True:
batch = dataloader.get_next_batch('dev')
if batch is None:
break
sent = torch.from_numpy(batch['sent']).long().cuda()
sent_length = torch.from_numpy(batch['sent_length']).long().cuda()
label = torch.from_numpy(batch['label']).long().cuda()
logit = model(sent, sent_length)
total_precise.append((torch.max(logit, dim=1)[1] == label).float().mean().cpu().data.numpy())
precise = np.mean(total_precise)
print('[dev]precise: %f' % (precise))
return precise
def predict(dataloader, model):
'''
predict labels for test set
'''
dataloader.restart('test', args.batchsize, shuffle=False)
with open('prediction', 'w') as w:
while True:
batch = dataloader.get_next_batch('test')
if batch is None:
break
sent = torch.from_numpy(batch['sent']).long().cuda()
sent_length = torch.from_numpy(batch['sent_length']).long().cuda()
logit = model(sent, sent_length)
prediction = torch.max(logit, dim=1)[1]
for x in prediction:
w.write('%d\n' % (x))
def main(dataloader, model, optimizer):
best_precise = {'dev':0., 'epoch':-1}
for e in range(args.epoch):
print('*' * 30)
print('trainging epoch %d...' % (e))
train(dataloader, model, optimizer)
dev_precise = eval(dataloader, model)
if dev_precise > best_precise['dev']:
best_precise['dev'] = dev_precise
best_precise['epoch'] = e
print('make prediction on test set...')
predict(dataloader, model)
print('[best dev performance]epoch: %d, dev: %f' \
% (best_precise['epoch'], best_precise['dev']))
if __name__ == '__main__':
print('load data...')
dataloader = SST()
print('create model...')
if args.mode not in ['GRU', 'LSTM', 'Attention']:
raise ValueError('Please use `--mode [GRU/LSTM/Attention]`')
model = Network(dataloader.emb, mode=args.mode)
model.cuda()
print('create optimizer...')
optimizer = optim.Adam(model.parameters())
print('training')
main(dataloader, model, optimizer)