-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdeletion_game_tree.py
254 lines (228 loc) · 12.4 KB
/
deletion_game_tree.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
import os.path
import time
from vnbdt_metric import *
from vnbdt import *
import cv2
from nbdt.model import SoftNBDT
import numpy as np
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
import shutil
import gc
if __name__ == "__main__":
cam_methods = ['gradcam', 'gradcam++', 'scorecam', 'efccam']
cam_method = cam_methods[0]
tree_methods = ['induced']
#dataset = 'Imagenet10'
dataset = 'Fashion10'
# method = 'induced'
# arch = 'vgg16'
# base_path = '/home/mist/checkpoints/vgg16_pretrained.pkl'
# nbdt_ft_path = '/home/mist/checkpoints/ckpt-Imagenet10-vgg16-lr0.01-SoftTreeSupLoss.pth'
#arch = 'ResNet50'
arch = 'DFLCNN'
# base_path = '/home/mist/checkpoints/resnet50_pretrained3.pkl'
# nbdt_ft_path = '/home/mist/checkpoints/ckpt-Imagenet10-ResNet50-lr0.01-SoftTreeSupLoss.pth'
base_path = '/home/lzl001/CNN_train/resnet50_pretrained3.pkl'
nbdt_ft_path = '/home/lzl001/NBDT/neural-backed-decision-trees/checkpoint/ckpt-Imagenet10-ResNet50-lr0.01-SoftTreeSupLoss_induced.pth'
pro_pth = '/home/lzl001/NBDT/neural-backed-decision-trees/checkpoint/ckpt-Imagenet10-ResNet50-lr0.01-SoftTreeSupLoss_pro.pth'
random_pth = '/home/lzl001/NBDT/neural-backed-decision-trees/checkpoint/ckpt-Imagenet10-ResNet50-lr0.01-SoftTreeSupLoss_random.pth'
# arch = 'vgg19_bn'
# base_path = '/home/lzl001/CNN_train/model_vgg19bn_pretrained.pkl'
# nbdt_ft_path = '/home/lzl001/NBDT/neural-backed-decision-trees/checkpoint/ckpt-Imagenet10-vgg19_bn-lr0.01-SoftTreeSupLoss.pth'
#tar_path = './pretrain_model/model_best_12.pth.tar'
#exp_img_source_path = '/home/mist/imagenet-10/val'
tar_path ='/home/lzl001/FGVC/DFL-CNN/weight/epoch_0015_top1_68_checkpoint.pth.tar'
exp_img_source_path = '/data/LZL/Fashion10/val'
cls_list = os.listdir(exp_img_source_path)
# cls_list_1 = ['Blouse','Cardigan','Dress','Jacket','Jeans','Romper','Shorts',
# 'Sweater', 'Tee']
# 获取该数据集下所有类别的对应ID
wnids = get_wnids_from_dataset(dataset)
num_cls = len(wnids)
t = 0
issue_img = []
for method in tree_methods:
Colors.bold(method + ' has started')
if method == 'pro':#还可以优化
net = call_pth_model(arch, pro_pth, cls_num=num_cls, device='cuda') #可改
# 生成树并读取
G, path = get_pro_tree(dataset=dataset, arch=arch, method=method)
# 验证树及其节点对应，并返回根节点Node
root = validate_tree(G, path, wnids)
model = SoftNBDT(
pretrained=False,
dataset=dataset,
path_graph=path,
model=net,
classes=wnids
)
elif method == 'random':#还可以优化
net = call_pth_model(arch, random_pth, cls_num=num_cls,device='cuda') #可改
# 生成树并读取
G, path = get_pro_tree(dataset=dataset, arch=arch, method=method)
# 验证树及其节点对应，并返回根节点Node
root = validate_tree(G, path, wnids)
model = SoftNBDT(
pretrained=False,
dataset=dataset,
path_graph=path,
model=net,
classes=wnids
)
elif method == 'base':
net = call_pkl_model(arch, base_path, cls_num=num_cls, device='cuda')
G, path = get_tree(dataset=dataset, arch=arch, model=net, method='induced') #预测与树结构无关，但需要以induced为对比
# 验证树及其节点对应，并返回根节点Node
root = validate_tree(G, path, wnids)
model = SoftNBDT(
pretrained=False,
dataset=dataset,
arch=arch,
model=net,
classes=wnids
)
elif dataset == 'Fashion10' and method == 'induced':
net = calling_DNLCNN_from_tar(tar_path, cls_num=10)
G, path = get_tree(dataset=dataset, arch=arch, model=net, method='induced')
# 验证树及其节点对应，并返回根节点Node
root = validate_tree(G, path, wnids)
model = SoftNBDT(
pretrained=False,
dataset=dataset,
arch=arch,
model=net,
classes=wnids
)
else:
net = call_pth_model(arch, nbdt_ft_path, cls_num=num_cls, device='cuda')
G, path = get_tree(dataset=dataset, arch=arch, model=net, method='induced')
# 验证树及其节点对应，并返回根节点Node
root = validate_tree(G, path, wnids)
model = SoftNBDT(
pretrained=False,
dataset=dataset,
arch=arch,
model=net,
classes=wnids
)
Colors.red(method + ' has been successfully reloaded')
target_layers = None
if arch == "ResNet50":
target_layers = [model.model.layer4[-1]]
elif arch == 'DFLCNN':
target_layers = [model.model.conv5]
elif arch == 'vgg16':
target_layers = [model.model.features[-1]]
elif arch == 'ResNet18':
target_layers = [model.model.layer4[-1]]
elif arch == 'wrn28_10_cifar10':
target_layers = [model.model.features[-3][-1].body.conv2.conv]
output_dir = os.path.join('/data/LZL/Fashion-epoch/complex', method)
t1 = time.time()
# 对每一个数据类别文件夹进行处理
# if method == 'induced':
# l = cls_list_1
# else:
# l = cls_list
# print('all cls:')
# print(l)
for cls in list(reversed(cls_list)):
cls_out_dir = os.path.join(output_dir, cls)
if not os.path.exists(cls_out_dir):
os.makedirs(cls_out_dir)
img_list = os.listdir(os.path.join(exp_img_source_path, cls))[-25:]
for img in img_list:
path_img = os.path.join(os.path.join(exp_img_source_path, cls), img)
img_name = os.path.split(path_img)[1].split('.')[0]
Colors.green("============= " + cls + " : " + img_name + " starts =============")
masked_dir = os.path.join(cls_out_dir, img_name)
if not os.path.exists(masked_dir):
os.makedirs(masked_dir)
# 数据处理,生成遮掩
img1 = Image.open(path_img)
img2 = cv2.imread(path_img, 1)
transform1 = get_transform()
img_tensor = transform1(img1)
x = img_tensor.unsqueeze(0)
# 生成遮掩保存
if method in ['pro','random','induced','base']:
gc.collect()
torch.cuda.empty_cache()
decisions, leaf_to_prob, node_to_prob, predicted = forword_tree(x.cuda(), model.cuda(), wnids,
dataset)
decision_to_wnid = get_decision_wnid(decisions[0])
record_node_prob(node_to_prob, decision_to_wnid,
os.path.join(output_dir, method + '_mask_record.txt'), cls + '+' + img_name)
if cam_method != 'efccam':
cam_dict = get_all_leaf_cam_from_method(x, net, leaf_to_prob,
num_cls, cam_method, target_layers,
aug_smooth=False,
eigen_smooth=False)
else:
cam_dict = get_all_leaf_cam_efc(x.cuda(), net.cuda(), leaf_to_prob, num_cls,
cam_method, target_layers, '', '')
complex_w = compute_complex_weight(cam_dict, predicted)
#生成决策链路所有节点显著图的遮掩图
generate_cam_mask_with_simple_w(decisions[0], img2, img_name, cam_dict, cls_out_dir)
generate_cam_mask_with_complex_w(decisions[0], img2, img_name, cam_dict, cls_out_dir, complex_w, predicted)
del cam_dict
for node_ in range(1, len(decisions[0])):
node_dir = os.path.join(masked_dir, 'node_' + str(node_))
# 根据遮掩保存路径读取图像进行预测，写入预测数据
for masked in os.listdir(node_dir):
img1 = Image.open(os.path.join(node_dir, masked))
#img2 = cv2.imread(os.path.join(node_dir, masked), 1)
transform1 = get_transform()
img_tensor = transform1(img1)
x = img_tensor.unsqueeze(0)
# img2 = cv2.resize(cv2.imread(os.path.join(node_dir, masked), 1), (224, 224))[:, :, ::-1]
# img2 = np.float32(img2) / 255
# x = preprocess_image(img2,
# mean=[0.5, 0.5, 0.5],
# std=[0.5, 0.5, 0.5])
mask_name = os.path.split(masked)[1].split('.')[0]
gc.collect()
torch.cuda.empty_cache()
decisions, leaf_to_prob, n2p, predicted = forword_tree(x.cuda(), model.cuda(), wnids, dataset)
d2w = get_decision_wnid(decisions[0])
record_node_prob(n2p, d2w,
os.path.join(output_dir, method + '_mask_record.txt'),
cls + '+' + mask_name)
# Colors.cyan("== " + mask_name + " finished ==")
Colors.green("============== " + cls + " - " + img_name + " ends ==============")
else:
decisions, leaf_to_prob, node_to_prob, predicted = forword_tree_no(x.cuda(), model.cuda(),
wnids,dataset)
decision_to_wnid = get_decision_wnid(decisions[0])
record_node_prob(node_to_prob, decision_to_wnid,
os.path.join(output_dir, method + '_mask_record.txt'), cls + '+' + img_name)
if cam_method != 'efccam':
cam = get_cam_from_method(x.cuda(), model.model.cuda(), predicted,
cam_method, target_layers,False,False)
else:
cam = get_cam_efc(x.cuda(), model.cuda(),target_layers,'','')
generate_cam_mask_one_sample(img2, img_name, cam, cls_out_dir)
# 根据遮掩保存路径读取图像进行预测，写入预测数据
for masked in os.listdir(masked_dir):
img2 = cv2.imread(os.path.join(masked_dir, masked), 1)
img2 = np.float32(img2) / 255
x = preprocess_image(img2,
mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
mask_name = os.path.split(masked)[1].split('.')[0]
decisions, leaf_to_prob, n2p, predicted = forword_tree_no(x.cuda(), model.cuda(), wnids, dataset)
d2w = get_decision_wnid(decisions[0])
record_node_prob(n2p, d2w,
os.path.join(output_dir, method + '_mask_record.txt'),
cls + '+' + mask_name)
Colors.green("============== " + cls + " - " + img_name + " ends ==============")
shutil.rmtree(masked_dir)
Colors.cyan("============== " + cls + " ends ==============")
t2 = time.time()
t += (t2 - t1)
Colors.red("{} cost time: {:.4f}".format(method, t2 - t1))
for i in issue_img:
print(i)
Colors.bold(
"total time: {}".format(t)
)