-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathPruningResnet18.py
180 lines (155 loc) · 6.84 KB
/
PruningResnet18.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
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
import torch.optim as optim
from ResNet18 import resnet18
from NEUCLSDataLoad import NEUCLASSDATA
import numpy as np
import os
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
BATCH_SIZE = 1
LEARNING_RATE = 0.001
original_list_name = []
weights_l1 = []
list_acc = []
remove_index = {}
base_min_acc = 97.00
original_layer = [64, 64, 64, 128, 128, 256, 256, 512, 512]
model = resnet18(original_layer).to(device)
model.eval()
train_data, test_data = NEUCLASSDATA()._get_data_from_index()
train_data_loader = DataLoader(train_data, batch_size=BATCH_SIZE, shuffle=True)
test_data_loader = DataLoader(test_data, batch_size=BATCH_SIZE, shuffle=True)
def _get_Name():
for name in model.state_dict():
original_list_name.append(name)
def _load_model():
for name in original_list_name:
temp_load_numpy = np.load("./SaveInfo/ResNet18/Para/resnet18/%s.ndim.npy" % (name))
tensor_load = torch.tensor(temp_load_numpy)
model.state_dict()[name].copy_(tensor_load)
def _computer_L1_value():
for module in model.modules():
if isinstance(module, nn.Conv2d):
weights_copy = module.weight.data.abs().clone()
weights_copy = weights_copy.cpu().numpy()
weights_copy_l1 = np.sum(weights_copy, axis=(1, 2, 3))
weights_l1.append(weights_copy_l1.tolist())
#寻找层中最N个最小值的坐标
def _get_minValue_index(value, nums):
value_sort = np.argsort(value)
index = value_sort[:nums]
return index
#获取当前索引层的名称
def _get_layer_name(index):
name_sign = 'feature.' + str(index)
feature_name = []
for name in original_list_name:
if name_sign in name:
feature_name.append(name)
if feature_name != []:
return feature_name
else:
return original_list_name[len(original_list_name)-2:]
def _resizeFeatureMap(featureMap, index, fm_type):
featureMap = featureMap.numpy()
if fm_type == 0:
featureMap = np.delete(featureMap, obj=index, axis=0)
return torch.tensor(featureMap)
else:
featureMap = np.delete(featureMap, obj=index, axis=1)
return torch.tensor(featureMap)
def _pruning_iter():
pruning_layer = [64, 64, 64, 128, 128, 256, 256, 512, 512]
layer_sign = [True]*len(pruning_layer)
while True:
for i in range(len(pruning_layer)):
if layer_sign[i] is False: continue
pruning_layer[i] = pruning_layer[i] - 1
pruning_model = resnet18(pruning_layer).to(device)
print("-------------------------------------------------------------------------------")
print("Current Layer Infor:")
print(pruning_layer)
print("-------------------------------------------------------------------------------")
for j in range(len(weights_l1)):
feature_name = _get_layer_name(j)
sign_name = feature_name[0]
for name in feature_name:
original_weight_shape = model.state_dict()[sign_name].shape
pruning_weight_shape = pruning_model.state_dict()[sign_name].shape
if len(model.state_dict()[name].shape) == 4:
sign_name = name
original_weight_shape = model.state_dict()[sign_name].shape
pruning_weight_shape = pruning_model.state_dict()[sign_name].shape
axis_0_dis = original_weight_shape[0] - pruning_weight_shape[0]
axis_1_dis = original_weight_shape[1] - pruning_weight_shape[1]
axis_1_dis_index, axis_0_dis_index = [], []
if axis_1_dis > 0:
axis_1_dis_index = _get_minValue_index(weights_l1[j-1], axis_1_dis)
if axis_0_dis > 0:
axis_0_dis_index = _get_minValue_index(weights_l1[j], axis_0_dis)
temp_feature = model.state_dict()[name].cpu()
if axis_1_dis > 0 and len(temp_feature.shape) > 1:
temp_feature = _resizeFeatureMap(temp_feature, axis_1_dis_index, 1)
if axis_0_dis > 0 and len(temp_feature.shape) > 0:
temp_feature = _resizeFeatureMap(temp_feature, axis_0_dis_index, 0)
pruning_model.state_dict()[name].copy_(temp_feature)
class_name = _get_layer_name(9)
for name in class_name:
if len(model.state_dict()[name].shape) > 1:
for t in range(model.state_dict()[name].shape[0]):
pruning_model.state_dict()[name][t].copy_(
model.state_dict()[name][t][:pruning_model.state_dict()[name].shape[1]])
else:
pruning_model.state_dict()[name].copy_(model.state_dict()[name])
for o in range(10):
list_acc.append(_test_model(pruning_model))
print("---------------------本轮剪枝精度-------------------------")
print(list_acc)
# pruning_min_acc = min(list_acc)
# list_acc.clear()
# save_sign = True
# print("本轮 min 精度：" + str(pruning_min_acc) + "%")
# if pruning_min_acc < base_min_acc:
# layer_sign[i] = False
# pruning_layer[i] = pruning_layer[i] + 1
# save_sign = False
# if save_sign:
# _save_model(pruning_model)
_save_model(pruning_model)
# print(layer_sign)
break
# if True not in layer_sign:
# print("Goal Layer Numbers: " + str(pruning_layer))
# _saveLayer(pruning_layer, 'pruning_layer_info_iter1')
# break
break
def _save_model(pruning_model):
path = "./SaveInfo/ResNet18/Para/PruningModel/"
if not os.path.exists(path):
os.makedirs(path)
for name in original_list_name:
temp_np = pruning_model.state_dict()[name].cpu().numpy()
np.save(path+"/%s.ndim" % (name), temp_np)
def _saveLayer(layer_info, layer_name):
with open("./SaveResult/PruningResult-Res/"+str(layer_name)+".txt", 'w') as fw:
for layer in layer_info:
fw.write(str(layer)+"-")
fw.close()
def _test_model(pruning_model):
with torch.no_grad():
total = 0
correct = 0
for data in test_data_loader:
inputs, labels = data
inputs, labels = inputs.to(device), labels.to(device)
outputs = pruning_model(inputs)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
return round((correct / total) * 100, 2)
if __name__ == "__main__":
_get_Name()
_load_model()
_computer_L1_value()
_pruning_iter()