-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_crf_glove.py
executable file
·215 lines (184 loc) · 8.91 KB
/
train_crf_glove.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
from __future__ import division
import torch
from data_reader_general import data_reader, data_generator
import pickle
import models
from util import AverageMeter
from util import save_checkpoint as save_best_checkpoint
import os.path as osp
import torch.backends.cudnn as cudnn
import argparse
from torch import optim
from sklearn.metrics import confusion_matrix, f1_score, recall_score, precision_score
torch.manual_seed(222)
# Get model names in the folder
model_names = sorted(name for name in models.__dict__
if name.islower() and not name.startswith("__") and callable(models.__dict__[name]))
def adjust_learning_rate(optimizer, epoch, args):
'''
Descend learning rate
'''
lr = args.lr / (2 ** (epoch // args.adjust_every))
print("Adjust lr to ", lr)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def save_checkpoint(save_model, i_iter, args, is_best=True):
'''
Save the model to local disk
'''
dict_model = save_model.state_dict()
filename = args.snapshot_dir
save_best_checkpoint(dict_model, is_best, i_iter, filename)
def train(model, dg_train, dg_valid, dg_test, optimizer, args):
cls_loss_value = AverageMeter(10)
best_acc = 0
best_f1 = 0
model.train()
is_best = False
loops = int(dg_train.data_len / args.batch_size)
for e_ in range(args.epoch):
print('Epoch numer: ', e_)
dg_train.reset_samples()
if e_ % args.adjust_every == 0:
adjust_learning_rate(optimizer, e_, args)
for idx in range(loops):
sent_vecs, mask_vecs, label_list, sent_lens, _, _, _ = next(dg_train.get_ids_samples())
if args.if_gpu:
sent_vecs, mask_vecs = sent_vecs.cuda(), mask_vecs.cuda()
label_list, sent_lens = label_list.cuda(), sent_lens.cuda()
cls_loss, norm_pen = model(sent_vecs, mask_vecs, label_list, sent_lens)
cls_loss_value.update(cls_loss.item())
total_loss = cls_loss + norm_pen
model.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_norm, norm_type=2)
optimizer.step()
valid_acc, valid_f1 = evaluate_test(dg_valid, model, args)
if valid_f1 > best_f1:
open('output.txt', 'w').close()
open('test_id.txt', 'w').close()
is_best = True
best_f1 = valid_f1
save_checkpoint(model, e_, args, is_best)
output_samples = True
if e_ % 10 == 0:
output_samples = True
test_acc, test_f1 = evaluate_test(dg_test, model, args, output_samples, test=True)
model.train()
is_best = False
return test_acc, test_f1
def evaluate_test(dr_test, model, args, sample_out=False, test=False):
mistake_samples = 'data/mistakes.txt'
result = 'output.txt'
with open(mistake_samples, 'w') as f:
f.write('Test begins...')
dr_test.reset_samples()
model.eval()
all_counter = 0
correct_count = 0
true_labels = []
pred_labels = []
sent_v = []
while dr_test.index < dr_test.data_len:
sent, mask, label, sent_len, texts, targets, _ = next(dr_test.get_ids_samples(test))
sent, mask, sent_len, label = sent.cuda(), mask.cuda(), sent_len.cuda(), label.cuda()
pred_label, best_seq, sent_vs, score = model.predict(sent, mask, sent_len)
# Compute correct predictions
correct_count += sum(pred_label == label).item()
true_labels.extend(label.cpu().numpy())
pred_labels.extend(pred_label.cpu().numpy())
# Output wrong samples, for debugging
indices = torch.nonzero(pred_label != label)
if len(indices) > 0:
indices = indices.squeeze(1)
sent_vss = []
sent_vss.extend(sent_vs.detach().cpu().numpy())
sent_v.extend(sent_vss)
if sample_out:
with open(result, 'a') as f:
for i in range(len(label)):
line = texts[i] + '###' + ' '.join(targets[i]) + '###' + str(label[i]) + '###' + str(pred_label[i]) + '\n'
f.write(line)
with open(mistake_samples, 'a') as f:
for i in indices:
line = texts[i] + '###' + ' '.join(targets[i]) + '###' + str(label[i]) + '###' + str(pred_label[i]) + '\n'
f.write(line)
if test:
pickle.dump(sent_v, open('sent_vs.pkl', 'wb'))
acc = correct_count * 1.0 / dr_test.data_len
f1 = f1_score(true_labels, pred_labels, average='macro')
if test:
print('test result: ', (acc, f1))
return acc, f1
# uncomment the following line for different dataset accordingly
# def main(l_hidden_size,dropout, dropout2, mask_dim , power, batch_size, num_layer):
# 16rest embed_num = 4419 and data path accordingly
# def main(l_hidden_size=2,dropout=4, dropout2=4, mask_dim=2 , power=3, batch_size=2, num_layer=2):
# # #14rest modify embed_num = 5120 and data path accordingly
# def main(l_hidden_size=1,dropout=8, dropout2=5, mask_dim=2, power=1, batch_size=3, num_layer=2):
# #14lap embed_num = 4070 and data path accordingly
def main(l_hidden_size=2, dropout=5, dropout2=6, mask_dim=4, power=1, batch_size=3, num_layer=1):
# 15rest embed_num = 3549 and data path accordingly
# def main(l_hidden_size=1, dropout=3, dropout2=3, mask_dim=3 , power=1, batch_size=2, num_layer=2):
parser = argparse.ArgumentParser()
parser.add_argument('-training', type=bool, default=True)
parser.add_argument('-embed_num', type=int, default=4070, help='The correct vocab size will print to screen, if error appears')
parser.add_argument('-arch', type=str, default='AspectSent')
parser.add_argument('-batch_size', type=int, default=int(32*batch_size))
parser.add_argument('-mask_dim', type=int, default=int(mask_dim*20+10))
parser.add_argument('-l_hidden_size', type=int, default=int(l_hidden_size * 32))
parser.add_argument('-l_num_layers', type=int, default=int(num_layer *2))
parser.add_argument('-l_dropout', type=int, default=0.1)
parser.add_argument('-power', type=int, default=int(power))
parser.add_argument('-dropout', type=int, default=dropout * 0.1)
parser.add_argument('-dropout2', type=int, default=dropout2 * 0.1)
parser.add_argument('-g_num_layer', type=int, default=2)
parser.add_argument('-embed_dim', type=int, default=300)
parser.add_argument('-if_update_embed', type=bool, default=False)
parser.add_argument('-if_reset', type=bool, default=True)
parser.add_argument('-epoch', type=int, default=30)
parser.add_argument('-lr', type=int, default=0.008)
parser.add_argument('-adjust_every', type=int, default=8)
parser.add_argument('-clip_norm', type=int, default=3)
parser.add_argument('-finetune_embed', type=bool, default=False)
parser.add_argument('-if_gpu', type=bool, default=True)
parser.add_argument('-use_gpu', type=bool, default=True)
parser.add_argument('-pretrained_embed_path', type=str, default='data/glove.840B.300d.txt')
parser.add_argument('-exp_name', type=str, default='laptop')
parser.add_argument('-embed_path', type=str, default='data/laptop/vocab/local_emb.pkl')
parser.add_argument('-data_path', type=str, default='data/laptop/')
parser.add_argument('-train_path', type=str, default='data/laptop/train.pkl')
parser.add_argument('-valid_path', type=str, default='data/laptop/valid.pkl')
parser.add_argument('-test_path', type=str, default='data/laptop/test.pkl')
parser.add_argument('-dic_path', type=str, default='data/laptop/vocab/dict.pkl')
parser.add_argument('-bestmodel_path', type=str, default='checkpoints/laptop/bestmodel.pth.tar')
parser.add_argument('-model_path', type=str, default='data/models/')
parser.add_argument('-snapshot_dir', type=str, default='checkpoints/')
args = parser.parse_args()
cudnn.enabled = True
args.snapshot_dir = osp.join(args.snapshot_dir, args.exp_name)
global best_acc
best_acc = 0
# Load datasets
dr = data_reader(args)
train_data = dr.load_data(args.train_path)
valid_data = dr.load_data(args.valid_path)
test_data = dr.load_data(args.test_path)
dg_train = data_generator(args, train_data)
dg_valid = data_generator(args, valid_data, False)
dg_test = data_generator(args, test_data, False)
model = models.__dict__[args.arch](args)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adam(parameters, lr=args.lr)
if args.if_gpu:
model = model.cuda()
if args.training:
# to decide if load saved best model or not
test_f1 = train(model, dg_train, dg_valid, dg_test, optimizer, args)
else:
# modify the best model path if want to use the best model to test
model.load_state_dict(torch.load(args.bestmodel_path))
test_acc, test_f1 = evaluate_test(dg_test, model, args, sample_out=False, test=True)
return test_f1
if __name__ == "__main__":
main()