-
Notifications
You must be signed in to change notification settings - Fork 7
/
BottleNet++_VGG16.py
188 lines (137 loc) · 7.28 KB
/
BottleNet++_VGG16.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
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torchvision import transforms
from torchvision.utils import save_image
from compression_module import *
import numpy as np
import matplotlib.pyplot as plt
from sklearn.model_selection import StratifiedShuffleSplit
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('-channel', type=str, default='a', help='channel type, using \'a\' as AWGN and \'e\' as BEC ')
parser.add_argument('-noise', type=float, default=0.1, help='channel condition')
parser.add_argument('-hid_dim', type=int, default=32, help='lens of encoded vector')
parser.add_argument('-in_dim', type=int, default=512, help='input dimension')
parser.add_argument('-div_position', type=int, default=13, help='divide the layer')
#parser.add_argument('-sub_div_position', type=int, default=1, help='sub_divide the layer')
parser.add_argument('-spatial', type=int, default=0, help='compress feature map')
parser.add_argument('-epoch', type=int, default=70, help='epoch')
parser.add_argument('-batch', type=int, default=32, help='batch size')
parser.add_argument('-phase', type=int, default=2, help='phase = 1,2,3, means to different training phase')
parser.add_argument('-lr', type=float, default=1e-4, help='leaerning rate')
args = parser.parse_args()
print('splitting point:',str(args.div_position),'input dim:',args.in_dim,'encoded dim',args.hid_dim,'spatial shrink:',args.spatial)
print('channel model:',args.channel,'channel condition:',args.noise)
print('phase:',args.phase,'epoch:',args.epoch,'batch:',args.batch,'learning rate:',args.lr)
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
'''
normalize = transforms.Normalize(mean=[0.5070751592371323, 0.48654887331495095, 0.4409178433670343],
std=[0.2673342858792401, 0.2564384629170883, 0.27615047132568404])
'''
test_set_this = torchvision.datasets.CIFAR100(root='../../CIFAR100', train=False, transform=transforms.Compose([
transforms.ToTensor(),
normalize,
]), download=True)
test_loader_this = torch.utils.data.DataLoader(test_set_this,
batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=True)
class BottleNetPlusPlus_VGG(nn.Module):
def __init__(self,input_channel = args.in_dim, hidden_channel = args.hid_dim, noise = args.noise, channel = args.channel,div_position = args.div_position,spatial =args.spatial):
super(BottleNetPlusPlus_VGG, self).__init__()
self.vgg_model = torch.load('vgg16_74.02.pth')
self.vgg_model.features = self.vgg_model.features.module
self.div_position = args.div_position
for para in self.vgg_model.parameters():
if args.phase == 2:
para.requires_grad = False
elif args.phase ==1 or 3:
para.requires_grad = True
self.compression_module = compression_module(input_channel = input_channel , hidden_channel = hidden_channel,noise = noise,channel = channel, spatial = spatial)
def forward(self, x):
insert_vae_flag = 0
#insert_list=[43,33,23,13,6]
insert_list=[2,6,9,13,16,19,23,26,29,33,36,39,43]
insert_vae_flag = insert_list[args.div_position - 1]
for i in range (len(list(self.vgg_model.features))):
x = list(self.vgg_model.features)[i](x)
if i==insert_vae_flag:
x = self.compression_module(x)
x = x.view(-1,512)
#print('cla',x.size())
x =self.vgg_model.classifier(x)
return x
model = BottleNetPlusPlus_VGG().to(device)
#model = torch.nn.DataParallel(model)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
def accuracy(output, labels):
preds = output.max(1)[1].type_as(labels)
correct = preds.eq(labels).double()
correct = correct.sum()
return correct / len(labels)
# Start training
def train(model=model):
for epoch in range(num_epochs):
if (epoch)%10 == 0:
train_set = torchvision.datasets.CIFAR100(root='../CIFAR100', train=True, transform=transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomCrop(32, 4),
transforms.ToTensor(),
normalize,
]), download=True)
data_loader = torch.utils.data.DataLoader(train_set,batch_size=batch_size, shuffle=True,
num_workers=4, pin_memory=True)
for i, (x, y) in enumerate(data_loader):
if (i+epoch)==0:
print('load model')
x = x.to(device)
y = y.to(device)
model.train()
output = model(x)
# Backprop and optimize
criterion = nn.CrossEntropyLoss()
criterion = criterion.to(device)
loss = criterion(output, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
accuracy_result = accuracy(output,y)
if (i+1) % int(50000/(args.batch*20)) == 0:
print ("Epoch[{}/{}], Step [{}/{}], Reconst Loss: {:.4f}, acc: {:.4f}"
.format(epoch+1, num_epochs, i+1, len(data_loader), loss.item(), accuracy_result.item()))
torch.save(model,'BottleNetPlusPlus_VGG.pkl')
if (epoch)%3 == 0:
test(epoch)
def test(epoch):
with torch.no_grad():
model.eval()
correct = 0
correct_top5 = 0#top5
total = 0
for i, (images, labels) in enumerate(test_loader_this):
images = images.to(device)
labels = labels.to(device)
outputs= model(images)
maxk = max((1,5))
labels_relize = labels.view(-1,1)
_, top5_pred = outputs.topk(maxk, 1, True, True)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct_top5 +=torch.eq(top5_pred, labels_relize).sum().float().item()
correct += (predicted == labels).sum().item()
#print('['+str(total)+'/10000]',(100 * correct / total))
if (100 * correct / total) > 60:
pred_best = (100 * correct / total)
torch.save(model,args.channel+'_div:'+str(args.div_position)+'_spatial:_'+str(args.spatial)+'_hid:'+str(args.hid_dim)+'_noise:'+str(args.noise)+'_acc{:.4f}_top5:{:.4f}_'.format((100 * correct / total),(100 *correct_top5/total))+'epoch:'+str(epoch)+'.pkl')
print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total),'top5: {} %'.format(100* correct_top5/total))
def adjust_learning_rate(optimizer, epoch):
lr = 0.001 * (0.5 ** (epoch // 10))
for param_group in optimizer.param_groups:
param_group['lr'] = lr
if __name__=='__main__':
train()