-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
182 lines (149 loc) · 8.06 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
177
178
179
180
181
182
import torch
import torch.nn as nn
import torch.optim as optim
import argparse
import numpy.random as random
import os
import numpy as np
from AWD_LSTM import AWD_LSTM
from utils import get_data,get_vocab,tokenize,batchify,
get_batch,repackage_hidden
def validate(dev_data,model,seq_len=100):
model.eval()
model.drop.training=False
lossFunction=nn.CrossEntropyLoss()
iters=dev_data.shape[0]//seq_len
loss,perplexity=0,0
hidden=[None for _ in range(model.n_layers)]
for i in range(iters):
if i!=0:
hidden=repackage_hidden(hidden)
X,Y=get_batch(dev_data,i*seq_len,seq_len)
logits,hidden,raw_outputs,outputs=model(X,hidden)
logits=logits.view(logits.shape[0]*logits.shape[1],-1)
Y=Y.contiguous().view(-1)
loss_=lossFunction(logits,Y)
loss+=loss_.item()
perplexity+=np.exp(loss_.item())
return loss/iters,perplexity/iters
def train(data_dev,data_train,model,args):
lossFunction=nn.CrossEntropyLoss()
optimizer=optim.SGD(model.parameters(),lr=args.lr)
for epoch in range(args.epochs):
model.train()
model.drop.training=True
hidden=[None for _ in range(model.n_layers)]
i,batch=0,0
max_loss=1000000000
best_dev_loss=[]
while i<data_train.shape[0]-2:
bptt=args.bptt if random.random()<0.95 else args.bptt//2
seq_len=max(5,int(random.normal(bptt,5)))
if data_train.shape[0]-2<i+seq_len-1:
break
lr_=optimizer.param_groups[0]['lr']
optimizer.param_groups[0]['lr']=lr_*seq_len/args.bptt
X,Y=get_batch(data_train,i,seq_len)
if i!=0:
hidden=repackage_hidden(hidden)
optimizer.zero_grad()
logits,hidden,raw_outputs,outputs=model(X,hidden)
logits=logits.view(logits.shape[0]*logits.shape[1],-1)
Y=Y.contiguous().view(-1)
loss=lossFunction(logits,Y)
ppl_loss=loss.item()
loss+=sum(args.alpha*dropped_h.pow(2).mean() \
for dropped_h in outputs[-1])
loss+=sum(args.beta*(h[1:]-h[:-1]).pow(2).mean() \
for h in raw_outputs[-1])
scalar=loss.item()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip)
optimizer.step()
i+=seq_len
batch+=1
if batch%args.log_interval==0:
print('epoch=',epoch+1,'percent done=',i/data_train.shape[0],
'training loss=',scalar,'training perplexity=',np.exp(ppl_loss))
if 't0' in optimizer.param_groups[0]:
temp={}
for param in model.parameters():
temp[param]=param.data.clone()
param.data=optimizer.state[param]['ax'].clone()
dev_loss,dev_perplexity=validate(data_dev,model)
print('validation loss=',dev_loss,
'validation perplexity=',dev_perplexity)
if dev_loss<max_loss:
max_loss=dev_loss
torch.save(model.state_dict(),'model.pth')
for param in model.parameters():
param.data=temp[param].clone()
else:
dev_loss,dev_perplexity=validate(data_dev,model)
print('validation loss=',dev_loss,
'validation perplexity=',dev_perplexity)
if dev_loss<max_loss:
max_loss=dev_loss
torch.save(model.state_dict(),'model.pth')
if 't0' not in optimizer.param_groups[0] and \
len(best_dev_loss)>args.n0 and \
dev_loss>min(best_dev_loss[:-args.n0]):
optimizer=optim.ASGD(modrl.parameters(),t0=0,lr=args.lr)
def main(args):
train_data=get_data(args.path_train)
dev_data=get_data(args.path_dev)
test_data=get_data(args.path_test)
word_to_int,int_to_word,sz1, \
sz2,sz3=get_vocab(train_data,args.t1,args.t2)
train_data=tokenize(train_data,word_to_int)
dev_data=tokenize(dev_data,word_to_int)
test_data=tokenize(test_data,word_to_int)
train_data=batchify(train_data,args.batch_size)
dev_data=batchify(dev_data,args.dev_batch_size)
test_data=batchify(test_data,args.dev_batch_size)
train_data=train_data.cuda()
dev_data=dev_data.cuda()
test_data=test_data.cuda()
vocab_sizes=[sz1,sz2,sz3]
g_list=[int(2**(args.N-i)) for i in range(args.N)]
model=AWD_LSTM(args.embed_dim,args.h_size,args.layers,
vocab_sizes,args.m,args.k,args.N,g_list,args.n_layers,
args.dropouti,args.dropouth,
args.dropout,args.dropout_embed,args.dropoutw).cuda()
train(dev_data,train_data,model,args)
test_loss,test_perplexity=validate(test_data,model)
print('test loss=',test_loss,'test perplexity=',test_perplexity)
def setup():
parser=argparse.ArgumentParser('argument parser')
parser.add_argument('--lr',type=float,default=30)
parser.add_argument('--epochs',type=int,default=40)
parser.add_argument('--k',type=int,default=1024)
parser.add_argument('--embed_dim',type=int,default=128)
parser.add_argument('--h_size',type=int,default=256)
parser.add_argument('--layers',type=int,default=3)
parser.add_argument('--n_layers',type=int,default=3)
parser.add_argument('--path_train',type=str,default=os.getcwd()+'/data/ptb.train.txt')
parser.add_argument('--path_dev',type=str,default=os.getcwd()+'/data/ptb.valid.txt')
parser.add_argument('--path_test',type=str,default=os.getcwd()+'/data/ptb.test.txt')
parser.add_argument('--dropouti',type=float,default=0.65)
parser.add_argument('--dropouth',type=float,default=0.3)
parser.add_argument('--dropout',type=float,default=0.4)
parser.add_argument('--dropout_embed',type=float,default=0.1)
parser.add_argument('--N',type=int,default=7)
parser.add_argument('--t1',type=int,default=1000)
parser.add_argument('--t2',type=int,default=100)
parser.add_argument('--n0',type=int,default=5)
parser.add_argument('--log_interval',type=int,default=10)
parser.add_argument('--batch_size',type=int,default=50)
parser.add_argument('--bptt',type=int,default=70)
parser.add_argument('--alpha',type=float,default=2)
parser.add_argument('--beta',type=float,default=1)
parser.add_argument('--dropoutw',type=float,default=0.5)
parser.add_argument('--dev_batch_size',type=int,default=10)
parser.add_argument('--m',type=int,default=400)
parser.add_argument('--clip',type=float,default=0.25)
args=parser.parse_args()
return args
if __name__=='__main__':
args=setup()
main(args)