-
Notifications
You must be signed in to change notification settings - Fork 0
/
reconstruct_pruned_new.py
290 lines (260 loc) · 9.29 KB
/
reconstruct_pruned_new.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
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
import os
import re
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
from matplotlib.backends.backend_pdf import PdfPages
from cifar_alex import CifarAlexNet
from prune_utility import *
def imshow(img):
img = img.cpu()
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
plt.show()
def getTargets(img):
torchvision.utils.make_grid(img)
img = img / 2 + 0.5
return img
def show_result(img):
torchvision.utils.make_grid(img)
img = img / 2 + 0.5
imshow(img)
def crit(t):
return (sum(t ** 2) / 2)
class ReconstructNet(nn.Module):
def __init__(self):
super(ReconstructNet, self).__init__()
self.conv1 = nn.Conv2d(128, 128, 3, padding = 1)
self.conv2 = nn.Conv2d(128, 128, 3, padding = 1)
self.conv3 = nn.Conv2d(128, 128, 3, padding = 1)
self.convt1 = nn.ConvTranspose2d(128, 64, 5, padding = 2, output_padding = 1, stride = 2)
self.convt2 = nn.ConvTranspose2d(64, 32, 5, padding = 2, output_padding = 1, stride = 2)
self.convt3 = nn.ConvTranspose2d(32, 3, 5, padding =2 , output_padding = 1, stride = 2)
def forward(self, inputs):
x = inputs
x = F.leaky_relu(self.conv1(x), negative_slope = 0.2)
x = F.leaky_relu(self.conv2(x), negative_slope = 0.2)
x = F.leaky_relu(self.conv3(x), negative_slope = 0.2)
x = F.leaky_relu(self.convt1(x), negative_slope = 0.2)
x = F.leaky_relu(self.convt2(x), negative_slope = 0.2)
x = self.convt3(x)
return x
if __name__ == "__main__":
keepOn = True
# Prepare the dataset
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5),(0.5, 0.5, 0.5))])
trainset = torchvision.datasets.CIFAR10(root = './data', train = True, transform = transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size = 128, shuffle = True, num_workers = 0)
testset = torchvision.datasets.CIFAR10(root='./data', train=False, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size = 128, shuffle=False, num_workers=0)
classes = ('plane', 'car', 'bird', 'cat',
'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
device = torch.device("cuda:0")
# Load the pretrained alexnet
alexnet = torch.load("alex_trained.pkl")
alexnet.eval()
# Init ReconstructNet
net = ReconstructNet()
st = 0
st_retrain = 0
if keepOn:
res = os.listdir("./data/exp5_prune")
for netFile in res:
last = int(re.sub("\D","",netFile))
if last > st:
st = last
if st > 0:
net = torch.load("./data/exp5_prune/reconstruct" + str(st) + ".pkl")
net.to(device)
crit = nn.MSELoss()
alexnet.to(device)
# Defince the detail of training
learningRate = [0.0001 for i in range(50)]
learningRate.extend([0.00005 for i in range(40)])
#ADMM algorithm parameters initialization
Z1 = projection(net.conv1.weight.data, configuration.P1)
U1 = torch.zeros(Z1.size()).cuda()
zeros1 = U1
Z2 = projection(net.conv2.weight.data, configuration.P2)
U2 = torch.zeros(Z2.size()).cuda()
zeros2 = U2
Z3 = projection(net.conv3.weight.data, configuration.P3)
U3 = torch.zeros(Z3.size()).cuda()
zeros3 = U3
Z4 = projection(net.convt1.weight.data, configuration.P4)
U4 = torch.zeros(Z4.size()).cuda()
zeros4 = U4
Z5 = projection(net.convt2.weight.data, configuration.P5)
U5 = torch.zeros(Z5.size()).cuda()
zeros5 = U5
Z6 = projection(net.convt3.weight.data, configuration.P6)
U6 = torch.zeros(Z6.size()).cuda()
zeros6 = U6
# Record performance
retrain_loss = []
test_loss = []
x_axis = []
#train and test
for epoch in range(st, len(learningRate)):
x_axis.append(epoch + 1)
optimizer = optim.Adam(net.parameters(), lr = learningRate[epoch])
# Train
accu_loss = 0
batchNum = 0
for i, data in enumerate(trainloader, 0):
batchNum += 1
# continue
optimizer.zero_grad()
inputs, labels = data
inputs = inputs.to(device)
res, feature = alexnet(inputs)
targets = getTargets(inputs)
outputs = net(feature) #forward step
loss = crit(outputs, targets) + 0.00005*(\
crit(net.conv1.weight.data - Z1 + U1, zeros1) + \
crit(net.conv2.weight.data - Z2 + U2, zeros2) + \
crit(net.conv3.weight.data - Z3 + U3, zeros3) + \
crit(net.convt1.weight.data - Z4 + U4, zeros4) + \
crit(net.convt2.weight.data - Z5 + U5, zeros5) + \
crit(net.convt3.weight.data - Z6 + U6, zeros6))
accu_loss += loss.item()
loss.backward()
optimizer.step()
print('[train] epoch: %d, batch: %d, loss: %.5f' % (epoch + 1, (i + 1), accu_loss / (i+1)))
retrain_loss.append(accu_loss / batchNum)
#ADMM
Z1 = net.conv1.weight.data + U1
Z1 = projection(Z1, percent=configuration.P1)
U1 = U1 + net.conv1.weight.data - Z1
Z2 = net.conv2.weight.data + U2
Z2 = projection(Z2, percent=configuration.P2)
U2 = U2 + net.conv2.weight.data - Z2
Z3 = net.conv3.weight.data + U3
Z3 = projection(Z3, percent=configuration.P3)
U3 = U3 + net.conv3.weight.data - Z3
Z4 = net.convt1.weight.data + U4
Z4 = projection(Z4, percent=configuration.P4)
U4 = U4 + net.convt1.weight.data - Z4
Z5 = net.convt2.weight.data + U5
Z5 = projection(Z5, percent = configuration.P5)
U5 = U5 + net.convt2.weight.data - Z5
Z6 = net.convt3.weight.data + U6
Z6 = projection(Z6, percent = configuration.P6)
U6 = U6 + net.convt3.weight.data - Z6
# Test
batchNum = 0
with torch.no_grad():
accu_loss = 0
for i, data in enumerate(testloader, 0):
batchNum += 1
inputs, labels = data
inputs = inputs.to(device)
res, feature = alexnet(inputs)
targets = getTargets(inputs)
outputs = net(feature)
loss = crit(outputs, targets) + 0.00005 * (\
crit(net.conv1.weight.data - Z1 + U1, zeros1) + \
crit(net.conv2.weight.data - Z2 + U2, zeros2) + \
crit(net.conv3.weight.data - Z3 + U3, zeros3) + \
crit(net.convt1.weight.data - Z4 + U4, zeros4) + \
crit(net.convt2.weight.data - Z5 + U5, zeros5) + \
crit(net.convt3.weight.data - Z6 + U6, zeros6))
accu_loss += loss.item()
# if i == 0:
# imshow(targets[0])
# imshow(outputs[0])
print('[test] epoch: %d, batch: %d, loss: %.5f' % (epoch + 1, (i + 1), accu_loss / (i+1)))
test_loss.append(accu_loss / batchNum)
pdf = PdfPages("reconstruct_v4_prune.pdf")
plt.figure(1)
plt.plot(x_axis, retrain_loss, x_axis, test_loss)
plt.xlabel("epoch")
plt.ylabel("loss")
pdf.savefig()
plt.close()
pdf.close()
# Save the net
net_name = "./data/exp5_prune/reconstruct" + str(epoch+1) + ".pkl"
torch.save(net, net_name)
print("train step complete")
# little test before official work
#dataiter = iter(trainloader)
#images, labels = dataiter.next()
#images = images.to(device)
#res, feature = alexnet(images)
#targets = getTargets(images)
# retrain step
# Record performance
retrain_loss = []
test_loss = []
x_axis = []
for epoch in range(0, 10):
x_axis.append(epoch + 1)
optimizer = optim.Adam(net.parameters(), lr = learningRate[epoch])
# Train
accu_loss = 0
batchNum = 0
for i, data in enumerate(trainloader, 0):
batchNum += 1
optimizer.zero_grad()
inputs, labels = data
inputs = inputs.to(device)
res, feature = alexnet(inputs)
targets = getTargets(inputs)
outputs = net(feature) #forward step
loss = crit(outputs, targets)
accu_loss += loss.item()
loss.backward()
#prune step
print("net.conv1:")
net.conv1.weight.data, net.conv1.weight._grad = apply_prune(net.conv1.weight, configuration.P1)
print("net.conv2:")
net.conv2.weight.data, net.conv2.weight._grad = apply_prune(net.conv2.weight, configuration.P1)
print("net.conv3:")
net.conv3.weight.data, net.conv3.weight._grad = apply_prune(net.conv3.weight, configuration.P1)
print("net.convt1:")
net.convt1.weight.data, net.convt1.weight._grad=apply_prune(net.convt1.weight,configuration.P4)
print("net.convt2:")
net.convt2.weight.data, net.convt2.weight._grad=apply_prune(net.convt2.weight,configuration.P5)
print("net.convt3:")
net.convt3.weight.data, net.convt3.weight._grad=apply_prune(net.convt3.weight,configuration.P6)
optimizer.step()
print('[retrain] epoch: %d, batch: %d, loss: %.5f' % (epoch + 1, (i + 1), accu_loss / (i+1)))
retrain_loss.append(accu_loss / batchNum)
# Test
batchNum = 0
with torch.no_grad():
accu_loss = 0
for i, data in enumerate(testloader, 0):
batchNum += 1
inputs, labels = data
inputs = inputs.to(device)
res, feature = alexnet(inputs)
targets = getTargets(inputs)
outputs = net(feature)
loss = crit(outputs, targets)
accu_loss += loss.item()
# if i == 0:
# imshow(targets[15])
# imshow(outputs[15])
print('[test] epoch: %d, batch: %d, loss: %.5f' % (epoch + 1, (i + 1), accu_loss / (i+1)))
test_loss.append(accu_loss / batchNum)
pdf = PdfPages("reconstruct_v4_prune_retrain.pdf")
plt.figure(1)
plt.plot(x_axis, retrain_loss, x_axis, test_loss)
plt.xlabel("epoch")
plt.ylabel("loss")
pdf.savefig()
plt.close()
pdf.close()
# Save the net
net_name = "./data/exp5_retrain/reconstruct" + str(epoch+1) + ".pkl"
torch.save(net, net_name)
print("retrain step complete")