-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
176 lines (149 loc) · 6.57 KB
/
train.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
from torch.autograd import Variable
import torch.nn as nn
import argparse
import data
import model
import torch
from reinforce_reduced import Reinforce
from tqdm import tqdm
from torch import optim
import time
from utils import Recorder, annealing
def repackage_hidden(h):
"""Wraps hidden states in new Variables, to detach them from their history."""
if type(h) == Variable:
return Variable(h.data)
else:
return tuple(repackage_hidden(v) for v in h)
def get_batch(source, i, cfg):
seq_len = min(cfg['max_len'], len(source) - 1 - i)
data = Variable(source[i:i + seq_len], requires_grad=False).cuda()
target = Variable(source[i + 1: i + 1 + seq_len], requires_grad=False).cuda()
return data, target
def evaluate(data_source, model, cfg):
# Turn on evaluation mode which disables dropout.
model.eval()
total_loss = 0
ntokens = cfg['dict_size']
criterion = nn.CrossEntropyLoss()
for i in tqdm(range(0, data_source.size(0) - 1, cfg['max_len'])):
data, targets = get_batch(data_source, i, cfg)
hidden = model.init_hidden(cfg['batch_size'])
output, _ = model(data, hidden)
output_flat = output.view(-1, ntokens)
total_loss += len(data) * criterion(output_flat, targets.view(-1)).data
# hidden = repackage_hidden(hidden)
model.train()
return total_loss[0] / len(data_source)
def batchify(data, bsz):
# Work out how cleanly we can divide the dataset into bsz parts.
nbatch = data.size(0) // bsz
# Trim off any extra elements that wouldn't cleanly fit (remainders).
data = data.narrow(0, 0, nbatch * bsz)
# Evenly divide the data across the bsz batches.
data = data.view(bsz, -1).t().contiguous()
# data = data.cuda()
return data
def save_model(path, model):
with open(path, 'wb') as f:
torch.save(model, f)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, default='./data/wikitext-2',
help='location of the data corpus')
parser.add_argument('--output', type=str, default='./results/log.txt',
help='location of log file')
parser.add_argument('--lr', type=float, default=0.001,
help='initial learning rate')
parser.add_argument('--sigma', type=float, default=0.01,
help='standard deviation for policy')
parser.add_argument('--gamma', type=float, default=1.0,
help='discount')
parser.add_argument('--alpha', type=float, default=0.1,
help='ratio')
parser.add_argument('--clip', type=float, default=0.25,
help='gradient clipping')
parser.add_argument('--epochs', type=int, default=40,
help='upper epoch limit')
parser.add_argument('--batch_size', type=int, default=20, metavar='N',
help='batch size')
parser.add_argument('--bptt', type=int, default=200,
help='sequence length')
parser.add_argument('--seed', type=int, default=1111,
help='random seed')
parser.add_argument('--gpu', action='store_true',
help='use GPU')
parser.add_argument('--init', type=str,
default='baseline_model.pt', help="The LSTM model")
parser.add_argument('--report', type=int,
default=50, help="The report interval")
args = parser.parse_args()
corpus = data.Corpus(args.data)
cfg = dict()
cfg['dict_size'] = len(corpus.dictionary)
cfg['output_file'] = args.output
cfg['init'] = args.init
cfg['max_len'] = args.bptt
cfg['epochs'] = args.epochs
cfg['GPU'] = args.gpu
cfg['lr'] = args.lr
cfg['sigma'] = args.sigma
cfg['gamma'] = args.gamma
cfg['alpha'] = args.alpha
cfg['batch_size'] = args.batch_size
cfg['saveto'] = './model_200/'
cfg['report_interval'] = args.report
print(cfg)
train_data = batchify(corpus.train, cfg['batch_size'])
val_data = batchify(corpus.valid, cfg['batch_size'])
test_data = batchify(corpus.test, cfg['batch_size'])
with open(cfg['init'], 'rb') as f:
policy = torch.load(f)
print(policy)
reinforce_model = Reinforce(policy=policy, sigma=cfg['sigma'], gamma=cfg['gamma'])
recorder = Recorder(output_path=cfg['output_file'])
valid_loss = []
loss = evaluate(val_data, reinforce_model.policy, cfg)
print('start from valid loss = ', loss)
valid_loss.append(loss)
ntokens = cfg['dict_size']
optimizer = optim.Adam(reinforce_model.parameters(), lr=cfg['lr'])
start_time = time.time()
for epoch in range(cfg['epochs']):
total_loss = 0.0
total_LM_loss = 0.0
for i in range(0, train_data.size(0) - 1, cfg['max_len']):
optimizer.zero_grad()
data, targets = get_batch(train_data, i, cfg)
base_hidden = policy.init_hidden(bsz=cfg['batch_size'])
hidden = policy.init_hidden(bsz=cfg['batch_size'])
loss, LM_loss = reinforce_model(data, targets, hidden, base_hidden, cfg['alpha'])
total_loss += loss.data
total_LM_loss += LM_loss
loss.backward()
optimizer.step()
# base_hidden = repackage_hidden(base_hidden)
# hidden = repackage_hidden(hidden)
nbsz = (i // cfg['max_len'] + 1)
if nbsz % cfg['report_interval'] == 0:
print('batch ', nbsz, ': loss = ', total_loss.cpu().numpy() / cfg['report_interval'])
print('batch ', nbsz, ': LM loss = ', total_LM_loss / cfg['report_interval'])
total_loss = 0.0
total_LM_loss = 0.0
# print('elapse time: ', time.time() - start_time)
# loss = evaluate(val_data, reinforce_model.policy, cfg)
# print('validation loss: ', loss)
print('Epoch: ', epoch, ' elapse:', time.time() - start_time)
loss = evaluate(val_data, reinforce_model.policy, cfg)
if loss > valid_loss[-1]:
annealing(optimizer, decay_rate=2)
cfg['lr'] /= 2
print('learning rate anneals to ', cfg['lr'])
valid_loss.append(loss)
recorder.record(epoch, cfg['alpha'], loss)
save_path = cfg['saveto'] + '_epoch' + str(epoch) + '_loss' + str(loss)
save_model(save_path, reinforce_model.policy)
print('Epoch: ', epoch, ' save to ', save_path)
test_loss = evaluate(test_data, reinforce_model.policy, cfg)
recorder.record('', cfg['alpha'], test_loss)
recorder.close()