-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdummy.py
147 lines (120 loc) · 5.23 KB
/
dummy.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
"""
Train your model
"""
import os
import shutil
import argparse
import numpy as np
import torch
from torch.optim import *
from opt import *
from data_provider import *
from model import *
from utils import *
from torch.autograd import Variable
from torch.utils.data import DataLoader
def evaluation(model,weight,options, val_dataloader):
val_loss_list = []
val_count = val_dataloader.dataset.__len__()
model.eval()
for i,(input,target,length,mask) in enumerate(val_dataloader):
print('Evaluating batch: #%d'%i)
input=Variable(input,volatile=True).cuda()
target=Variable(target,volatile=True).cuda()
mask=Variable(mask,volatile=True).cuda()
output=model(input,length)+mask
criterion=torch.nn.BCEWithLogitsLoss(weight=weight.expand_as(target)).cuda()
# criterion=torch.nn.BCEWithLogitsLoss().cuda()
loss=criterion(output,target)
val_loss_list.append(loss.data[0]*len(length))
ave_val_loss = sum(val_loss_list) / float(val_count)
return ave_val_loss
def save_checkpoint(filename,state,is_best):
torch.save(state, filename)
if is_best:
shutil.copyfile(filename, os.path.dirname(filename)+'/best_model.pth')
def train(options):
print('Load data ...')
collate_fn=PadCollate()
train_data_provision = DataProvision(options,'train')
train_dataloader=torch.utils.data.DataLoader(train_data_provision,options['batch_size'],True,collate_fn=collate_fn)
val_data_provision = DataProvision(options,'val')
val_dataloader=torch.utils.data.DataLoader(val_data_provision,options['eval_batch_size'],False,collate_fn=collate_fn)
batch_size = options['batch_size']
max_epochs = options['max_epochs']
init_epoch = options['init_epoch']
lr_init = options['lr']
lr = lr_init
n_iters_per_epoch = train_data_provision.__len__() // batch_size
eval_in_iters = int(n_iters_per_epoch / float(options['n_eval_per_epoch']))
# build model
model = ProposalModel(options).cuda()
print('Build model for training stage ...')
weight=train_data_provision._proposal_weight.contiguous()
weight=weight.view(1,1,*weight.shape)
if options['solver'] == 'adam':
optimizer = Adam(model.parameters(),lr)
elif options['solver'] == 'adadelta':
optimizer = Adadelta(model.parameters(),lr,weight_decay=options['reg'])
else:
optimizer = SGD(model.parameters(),lr,weight_decay=options['reg'])
t0 = time.time()
eval_id = 0
total_iter = 0
best_loss=100000.0
for epoch in range(init_epoch, max_epochs):
model.train()
if epoch==options['max_epochs']//2:
lr=lr_init/10
for pg in optimizer.param_groups:
pg['lr']=lr
print('epoch: %d/%d, lr: %.1E (%.1E)'%(epoch, max_epochs, lr, lr_init))
for iter,(input,target,length,mask) in enumerate(train_dataloader):
input_var = Variable(input,requires_grad=True).cuda()
target_var = Variable(target).cuda()
mask=Variable(mask).cuda()
optimizer.zero_grad()
output=model(input_var,length)+mask
criterion=torch.nn.BCEWithLogitsLoss(weight.expand_as(target_var)).cuda()
loss=criterion(output,target_var)
for x in model.parameters():
if x is not None:
loss+= options['reg']*torch.sum(x**2)/2
loss.backward()
torch.nn.utils.clip_grad_norm(model.parameters(),options['clip_gradient_norm'])
optimizer.step()
if iter % options['n_iters_display'] == 0:
print('iter: %d, epoch: %d/%d, \nlr: %.1E, loss: %.4f'%(iter, epoch, max_epochs, lr, loss.data[0]))
if (total_iter+1) % eval_in_iters == 0:
is_best=0
print('Evaluating model ...')
val_loss = evaluation(model,weight,options,val_dataloader)
if val_loss<best_loss:
best_loss=val_loss
is_best=1
print('loss: %.4f'%val_loss)
checkpoint_path = '%sepoch%02d_%.2f_%02d_lr%f.ckpt' % (options['ckpt_prefix'], epoch, val_loss, eval_id, lr)
save_checkpoint(checkpoint_path,
{'state_dict':model.state_dict()},
is_best)
eval_id = eval_id + 1
total_iter += 1
if __name__ == '__main__':
parser = argparse.ArgumentParser()
options = default_options()
for key, value in options.items():
parser.add_argument('--%s'%key, dest=key, type=type(value), default=None)
args = parser.parse_args()
args = vars(args)
for key, value in args.items():
if value is not None:
options[key] = value
options=later_options(options)
work_dir = options['ckpt_prefix']
if not os.path.exists(work_dir) :
os.makedirs(work_dir)
find_idle_gpu(options['gpu'])
options['max_epochs']=1000000
options['train_id']=9999
options['n_eval_per_epoch']=0.0001
train(options)