-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
240 lines (221 loc) · 10.1 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
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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
from __future__ import print_function
import argparse
import torch
import time
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import utils
import os
import sys
import numpy as np
import models
from torch.autograd import Variable
from torchvision import datasets, transforms
import matplotlib.pyplot as plt
import config
import scipy.io as io
import pickle
args = config.get_args()
if args.mode!='evaluate':
if not os.path.exists(args.log_dir):
os.mkdir(args.log_dir)
sys.stdout = utils.Logger(os.path.join(args.log_dir,'print.log'))
print(args)
####################### datasets #################################
train_loader,valid_loader,test_loader=utils.get_dataset(args)
####################### models #################################
model=models.getmodels(args)
print(model)
####################### optimizers #################################
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=0.9, weight_decay=0.0005,nesterov=False)
####################### train_epoch #################################
def train(epoch,Log=False):
utils.toTrain(model)
for batch_idx, (data, target) in enumerate(train_loader):
if args.mode=='reduce' or args.mode=='actquant':
optimizer.param_groups[0]['lr']=utils.lr_scheduler_gran(args.lr,
optimizer.param_groups[0]['lr'],len(train_loader),
args.reduce_arg['epochs'],batch_idx,epoch,'cosine',
max(int(args.reduce_arg['epochs']/10),0))
if args.mode=='actquant' and batch_idx%100==0:
utils.sync_clip(model, args)
if args.cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data), Variable(target)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
if batch_idx % args.step_freq==0:
if args.quantize is not 'Dont':
for p in list(model.parameters()):
if hasattr(p,'fp'):
p.data.copy_(p.fp)
optimizer.step()
for p in list(model.parameters()):
if hasattr(p,'fp'):
p.fp.copy_(p.data)
else:
optimizer.step()
#print loging
if Log and int(100*batch_idx*len(data) / len(train_loader.dataset)) % args.log_interval == 0 :
if int(100*(batch_idx-1)*len(data) / len(train_loader.dataset)) % args.log_interval!= 0 :
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx*len(data) / len(train_loader.dataset), loss.data.item()))
####################### test_epoch #################################
def test(epoch,loader,name='Test'):
utils.toEval(model)
test_loss = 0
correct = 0
correct1 = 0
correct5 = 0
flag=-1
with torch.no_grad():
for data, target in loader:
if args.cuda:
data, target = data.cuda(), target.cuda()
output = model(data)
test_loss += criterion(output, target).data.item() # sum up batch loss
pred = output.data.max(1, keepdim=True)[1] # get the index of the max log-probability
correct += pred.eq(target.data.view_as(pred)).cpu().sum()
correct5 += utils.accuracy(output,target,5)
if flag == 1:
pred_all=torch.cat((pred_all,pred),0)
target_all=torch.cat((target_all,target),0)
else:
pred_all=pred
target_all=target
flag=1
test_loss = test_loss/len(loader.dataset)
print('{} data => Epoch: {} Average loss: {:.6f}, Accuracy: {}/{} ({:.2f}%) Top5 : {:.2f}%'.format(name,
epoch,test_loss, correct, len(loader.dataset),
100. * float(correct) / len(loader.dataset),
100. * float(correct5) / len(loader.dataset)))
## Loging
if args.mode == 'train':
if name=='Test':
log_file_test.write('\n{}\t{:.8f}\t{}\t{}\t{}'.format(epoch,test_loss,correct,correct5,len(loader.dataset)))
elif name=='Train':
log_file_train.write('\n{}\t{:.8f}\t{}\t{}\t{}'.format(epoch,test_loss,correct,correct5,len(loader.dataset)))
return pred_all.cpu().numpy(), target_all.cpu().numpy(), correct
####################### Code Modes #################################
if args.mode == 'actquant':
_,_,correct=test(args.epoch,test_loader,'Test')
reduce_arg=config.reduce_param_act(args)
if not args.actquant_model:
utils.quantize_model(model, args)
utils.sync_clip(model, args,reduce_arg['nbits'])
_,_,correct=test(args.epoch,test_loader,'Test')
correct_old=correct
for i in range(args.iter_restart,200):
print('Iternation: {}'.format(i))
optimizer.param_groups[0]['lr']=reduce_arg['lr_begin']
args.lr=reduce_arg['lr_begin']
print('Learning Rate: {}'.format(args.lr))
for k in range(len(reduce_arg['gransteps'])):
if reduce_arg['gransteps'][k]==i:
gran=reduce_arg['gransize'][k]
t = time.time()
utils.model_size_act(model)
utils.save_checkpoint(model,args,'Reduced_Model{}.pt'.format(i))
utils.get_evalues_act(model,valid_loader,reduce_arg['eval_pct'],reduce_arg['eval_top'])
utils.reduce_precision_act(model,gran,args.quantize)
utils.zero_evalues_act(model)
_,_,correct=test(0,test_loader,'Test')
best_correct=0
for j in range(reduce_arg['epochs']):
elapsed = time.time() - t
print('Time Taken: {}'.format(elapsed))
t = time.time()
train(j, Log=True)
utils.sync_clip(model, args)
_,_,correct=test(j,test_loader,'Test')
utils.save_checkpoint(model,args,'Checkpoint{}.pt'.format(i))
if correct>best_correct:
utils.save_checkpoint(model,args,'Best{}.pt'.format(i))
best_correct=correct
if j in reduce_arg['lr_steps']:
#optimizer.param_groups[0]['lr']=args.lr/10
#args.lr=args.lr/10
args.lr=args.lr
print('Learning Rate: {}'.format(optimizer.param_groups[0]['lr']))
correct_old=correct
elif args.mode == 'reduce':
_,_,correct=test(args.epoch,test_loader,'Test')
reduce_arg=config.reduce_param(args)
for i in range(200):
optimizer.param_groups[0]['lr']=reduce_arg['lr_begin']
args.lr=reduce_arg['lr_begin']
print('Learning Rate: {}'.format(args.lr))
for k in range(len(reduce_arg['gransteps'])):
if reduce_arg['gransteps'][k]==i:
gran=reduce_arg['gransize'][k]
utils.model_size(model)
utils.save_checkpoint(model,args,'Reduced_Model{}.pt'.format(i))
utils.get_evalues(model,valid_loader,reduce_arg['eval_pct'],reduce_arg['eval_top'])
utils.reduce_precision(model,gran,args.quantize)
utils.zero_evalues(model)
_,_,correct=test(0,test_loader,'Test')
best_correct=0
for j in range(reduce_arg['epochs']):
train(j, Log=True)
_,_,correct=test(j,test_loader,'Test')
utils.save_checkpoint(model,args,'Checkpoint{}.pt'.format(i))
if correct>best_correct:
utils.save_checkpoint(model,args,'Best{}.pt'.format(i))
best_correct=correct
if j in reduce_arg['lr_steps']:
#optimizer.param_groups[0]['lr']=args.lr/10
#args.lr=args.lr/10
args.lr=args.lr
print('Learning Rate: {}'.format(optimizer.param_groups[0]['lr']))
_,_,correct=test(i,test_loader,'Test')
elif args.mode == 'evaluate':
folder,modelName=os.path.split(args.checkpoint)
modelName=modelName.replace('.pt','')
pred_test,target_test,correct=test(0,test_loader,'Test')
outdict=utils.plotsizemobilenet(model,'weight')
io.savemat(os.path.join(folder,modelName+'_WeightQuant.mat'),outdict)
pickle.dump(outdict,open(os.path.join(folder,modelName+'_WeightQuant.pkl'),'wb'))
plt.savefig(os.path.join(folder,modelName+'_WeightQuant.pdf'))
print('Plot saved at: '+ os.path.join(folder, modelName+'_WeightQuant.pdf'))
outdict=utils.plotsizemobilenet(model,'activation')
pickle.dump(outdict,open(os.path.join(folder,modelName+'_ActQuant.pkl'),'wb'))
io.savemat(os.path.join(folder,modelName+'_ActQuant.mat'),outdict)
plt.savefig(os.path.join(folder,modelName+'_ActQuant.pdf'))
print('Activation plot saved at: '+ os.path.join(folder, modelName+'_ActQuant.pdf'))
elif args.mode == 'ActquantApt':
utils.quantize_model(model, args)
for i in range(1,16):
print('------------------Act Bits{}------------------'.format(i))
utils.sync_clip(model, args,i)
pred_test,target_test,correct=test(0,test_loader,'Test')
elif args.mode == 'train':
best_correct=0
log_file_test = open(os.path.join(args.log_dir,'LearningCurveTest.txt'), "w")
log_file_train = open(os.path.join(args.log_dir,'LearningCurveTrain.txt'), "w")
log_file_test.write('epoch\ttest-loss\tcorrect\ttest-size')
log_file_train.write('epoch\ttrain-loss\tcorrect\ttrain-size')
while (args.epoch < args.epochs+1):
t = time.time()
optimizer.param_groups[0]['lr']=utils.lr_scheduler(args)
print('Learning rate is {}'.format(optimizer.param_groups[0]['lr']))
train(args.epoch, Log=True)
_,_,correct=test(args.epoch,test_loader,'Test')
log_file_test.flush()
log_file_train.flush()
utils.save_checkpoint(model,args,'Checkpoint.pt')
if correct>best_correct:
best_correct=correct
utils.save_checkpoint(model,args,'Best.pt')
print('New Best Accuracy: {:0.2f}'.format(100. * float(best_correct) / len(test_loader.dataset)))
args.epoch=args.epoch+1;
# do stuff
elapsed = time.time() - t
print('Time Taken: {}'.format(elapsed))
print('Best Accuracy : {:0.2f}%'.format(100. * float(best_correct) / len(test_loader.dataset)))
else:
print('Choose a Mode')