-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathvitprune.py
347 lines (315 loc) · 11.7 KB
/
vitprune.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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
import importlib
import math
import os
import time
import torch
from torch import nn
import torch.backends.cudnn as cudnn
import numpy as np
import argparse
import torchvision
import torchvision.transforms as transforms
from utils import progress_bar, load_partial_weight, _pil_interp
from ptflops import get_model_complexity_info
from models.vit_pw import ViT, channel_selection
from models.vit_slim_pw import ViT_slim
device = 'cuda' if torch.cuda.is_available() else 'cpu'
torch.backends.cudnn.enable =True
torch.backends.cudnn.benchmark = True
# 单独测试每个类精度
def test_one_class(model):
scale_size = int(math.floor(224 / 0.9))
normalize = transforms.Normalize(mean=[0.4914, 0.4822, 0.4465],
std=[0.2023, 0.1994, 0.2010])
transform_test = transforms.Compose([
transforms.Resize(scale_size, _pil_interp('bicubic')),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])
testset = torchvision.datasets.CIFAR10(
root='/home/Datasets',
# root='/home/users/xjs/DataSets',
train=False,
transform=transform_test)
testloader = torch.utils.data.DataLoader(
testset, batch_size=100, shuffle=False,
num_workers=0, pin_memory=True
)
classes = ('plane', 'car', 'bird', 'cat',
'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
N_CLASSES = 10
class_correct = list(0. for _ in range(N_CLASSES))
class_total = list(0. for _ in range(N_CLASSES))
model.eval()
# test
total_correct = 0
total_num = 0
start = time.time()
with torch.no_grad():
for inputs, targets in testloader:
inputs, targets = inputs.to(device), targets.to(device)
outputs = model(inputs)
pred = outputs.argmax(dim=1)
total_correct += torch.eq(pred, targets).float().sum().item()
total_num += inputs.size(0) # 即batch_size
c = (pred == targets).squeeze()
for i in range(len(targets)):
_label = targets[i]
class_correct[_label] += c[i].item()
class_total[_label] += 1
end = time.time()
print('Acc: %.3f%% (%d/%d)' % (100. * total_correct / total_num, total_correct, total_num))
print('Test time:{}s'.format(end - start))
for i in range(N_CLASSES):
print('Accuracy of %5s : %.1f%%' % (
classes[i], 100. * class_correct[i] / class_total[i]))
# 剪枝后的微调
def train(model, epochs, percent):
scale_size = int(math.floor(224 / 0.9))
normalize = transforms.Normalize(mean=[0.4914, 0.4822, 0.4465],
std=[0.2023, 0.1994, 0.2010])
transform_train = transforms.Compose([
transforms.Resize(scale_size, _pil_interp('bicubic')),
transforms.CenterCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
trainset = torchvision.datasets.CIFAR10(
# root='/home/DataSets',
root='/home/users/xjs/DataSets',
train=True,
transform=transform_train)
trainloader = torch.utils.data.DataLoader(
trainset, batch_size=64, shuffle=False,
num_workers=0, pin_memory=True
)
optimizer = torch.optim.SGD(model.parameters(),lr=1e-3,momentum=0.9,weight_decay=1e-4)
torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,50,eta_min=0.0003)
criterion = nn.CrossEntropyLoss().cuda()
for epoch in range(epochs):
print('\nEpoch: %d' % epoch)
model.train()
train_loss = 0
correct = 0
total = 0
for batch_idx, (inputs, targets) in enumerate(trainloader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, targets)
loss.backward()
# sparse_selection()
optimizer.step()
train_loss += loss.item()
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
progress_bar(batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'
% (train_loss / (batch_idx + 1), 100. * correct / total, correct, total))
if epoch % 5 == 0:
torch.save(model.state_dict(),
'checkpoint/cifar10/vtp/pruned_' + str(percent) + '.pth')
print("save checkpoint in epoch:{}".format(epoch))
def getFileSize(filePath):
fsize = os.path.getsize(filePath) # 返回的是字节大小
'''
为了更好地显示，应该时刻保持显示一定整数形式，即单位自适应
'''
if fsize < 1024:
return(round(fsize,2),'Byte')
else:
KBX = fsize/1024
if KBX < 1024:
return str(round(KBX,1))+'K'
else:
MBX = KBX /1024
if MBX < 1024:
return str(round(MBX,1))+'M'
else:
return(round(MBX/1024),'G')
def run(path):
model = importlib.import_module('.vgg', package='{}'.format(path))
return model
if __name__ == '__main__':
path = 'help.models'
m = run(path)
model = m.VGG('VGG19')
print(model)
# 定义模型导入权重文件
model = ViT(
image_size=224,
patch_size=16,
num_classes=10,
dim=768,
depth=12,
heads=12,
mlp_dim=3072,
dropout=0.1,
emb_dropout=0.1,
qkv_bias=True
)
model = model.to(device)
# 导入权重文件
model_path = "checkpoint/cifar10/advance_cs/cifar10_pruned_attn0.0_ffn0.0.pth"
print("1 => loading checkpoint '{}'".format(model_path))
weight = torch.load(model_path)
load_partial_weight(model, weight)
test_one_class(model)
# flops, params = get_model_complexity_info(model, (3, 224, 224), as_strings=True, print_per_layer_stat=True)
# print('flops: ', flops, 'params: ', params)
# total = 0
# for m in model.modules():
# if isinstance(m, channel_selection):
# total += m.indexes.data.shape[0]
#
# bn = torch.zeros(total)
# index = 0
# for m in model.modules():
# if isinstance(m, channel_selection):
# size = m.indexes.data.shape[0]
# bn[index:(index + size)] = m.indexes.data.abs().clone()
# index += size
#
# percent = 0.4
# print(percent)
# y, i = torch.sort(bn)
# thre_index = int(total * percent)
# thre = y[thre_index]
#
#
# # 计算剪枝数量、剩余通道、mask列表
# pruned = 0
# cfg = []
# cfg_mask = []
# for k, m in enumerate(model.modules()):
# if isinstance(m, channel_selection):
# # print(k)
# # print(m)
# if k in [15, 37, 59, 81, 103, 125, 147, 169, 191, 213, 235, 257]:
# weight_copy = m.indexes.data.abs().clone()
# mask = weight_copy.gt(thre).float().cuda()
# thre_ = thre.clone()
# while (torch.sum(mask) % 12 != 0): # heads
# thre_ = thre_ - 0.0001
# mask = weight_copy.gt(thre_).float().cuda()
# else:
# weight_copy = m.indexes.data.abs().clone()
# mask = weight_copy.gt(thre).float().cuda()
# pruned = pruned + mask.shape[0] - torch.sum(mask)
# m.indexes.data.mul_(mask)
# cfg.append(int(torch.sum(mask)))
# cfg_mask.append(mask.clone())
# print('layer index: {:d} \t total channel: {:d} \t remaining channel: {:d}'.
# format(k, mask.shape[0], int(torch.sum(mask))))
#
# pruned_ratio = pruned / total
# print('Pre-processing Successful!')
# print(cfg)
#
# cfg_prune = []
# for i in range(len(cfg)):
# if i % 2 != 0:
# cfg_prune.append([cfg[i - 1], cfg[i]])
#
# 定义一个新的模型结构复制参数
# newmodel = ViT_slim(
# image_size=224,
# patch_size=16,
# num_classes=10,
# dim=768,
# depth=12,
# heads=12,
# mlp_dim=3072,
# dropout=0.1,
# emb_dropout=0.1,
# qkv_bias=True,
# cfg=cfg_prune) # 根据上面得到的剪枝cfg_prune重新定义网络
# newmodel.to(device)
#
# # 参数移植
# newmodel_dict = newmodel.state_dict().copy()
# i = 0
# newdict = {}
# for k, v in model.state_dict().items():
# if 'net.0.weight' in k:
# # print(k)
# # print(v.size())
# # print('----------')
# idx = np.squeeze(np.argwhere(np.asarray(cfg_mask[i].cpu().numpy())))
# newdict[k] = v[idx.tolist()].clone()
# elif 'net.0.bias' in k:
# # print(k)
# # print(v.size())
# # print('----------')
# idx = np.squeeze(np.argwhere(np.asarray(cfg_mask[i].cpu().numpy())))
# newdict[k] = v[idx.tolist()].clone()
# elif 'to_qkv' in k:
# # print(k)
# # print(v.size())
# # print('----------')
# qkv_mask = torch.cat((cfg_mask[i], cfg_mask[i], cfg_mask[i]))
# idx = np.squeeze(np.argwhere(np.asarray(qkv_mask.cpu().numpy())))
# newdict[k] = v[idx.tolist()].clone()
# elif 'net2.0.weight' in k:
# # print(k)
# # print(v.size())
# # print('----------')
# idx = np.squeeze(np.argwhere(np.asarray(cfg_mask[i].cpu().numpy())))
# newdict[k] = v[:, idx.tolist()].clone()
# i = i + 1
# elif 'to_out.0.weight' in k:
# # print(k)
# # print(v.size())
# # print('----------')
# idx = np.squeeze(np.argwhere(np.asarray(cfg_mask[i].cpu().numpy())))
# newdict[k] = v[:, idx.tolist()].clone()
# i = i + 1
#
# elif k in newmodel.state_dict():
# newdict[k] = v
#
# newmodel_dict.update(newdict)
# newmodel.load_state_dict(newmodel_dict)
# print("3 => load new dict successfully!")
#
# # 剪枝后微调
# train(newmodel, 16, percent)
# print("4 => pruned model: ")
# flops, params = get_model_complexity_info(newmodel, (3, 224, 224), as_strings=True, print_per_layer_stat=False)
# print('flops: ', flops, 'params: ', params)
#
# # 微调后测试并保存
# print("5 => after pruning: ")
# test_one_class(newmodel)
# torch.save(newmodel.state_dict(),
# 'checkpoint/cifar10/vtp/pruned_' + str(percent) + '.pth')
# filePath = 'checkpoint/cifar10/vtp/pruned_' + str(percent) + '.pth'
# print('pruned_' + str(percent) + '.pth'+' '+getFileSize(filePath))
# cfg_prune = [[696, 2077], [732, 2047], [732, 2071], [444, 2008], [744, 2066], [756, 2120],
# [504, 2157], [660, 2258], [720, 2420], [552, 1730], [756, 892],[696, 517]]
# # 定义一个新的模型结构复制参数
# model = ViT_slim(
# image_size=224,
# patch_size=16,
# num_classes=10,
# dim=768,
# depth=12,
# heads=12,
# mlp_dim=3072,
# dropout=0.1,
# emb_dropout=0.1,
# qkv_bias=True,
# cfg=cfg_prune) # 根据上面得到的剪枝cfg_prune重新定义网络
# model.to(device)
#
# flops, params = get_model_complexity_info(model, (3, 224, 224), as_strings=True, print_per_layer_stat=False)
# print('flops: ', flops, 'params: ', params)
#
# # 导入权重文件
# model_path = "checkpoint/cifar10/vtp/pruned_0.4.pth"
# print("1 => loading checkpoint '{}'".format(model_path))
# weight = torch.load(model_path)
# load_partial_weight(model, weight)
# test_one_class(model)