-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathvisual_featuremap.py
101 lines (78 loc) · 3.99 KB
/
visual_featuremap.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
import numpy as np
import cv2
import os
import matplotlib.pyplot as plt
def BGR_to_RGB(cvimg):
pilimg = cvimg.copy()
pilimg[:, :, 0] = cvimg[:, :, 2]
pilimg[:, :, 2] = cvimg[:, :, 0]
return pilimg
# heat 为某层的特征图，自己手动获取
def visual(heat, save_path):
# heat:[b, c, h, w]
fig = plt.figure(figsize=(8, 8))
fig.subplots_adjust(left=0.05, right=0.95, bottom=0.05, top=0.95, wspace=0.05, hspace=0.05)
for i in range(len(heat)):
heat1 = heat[i] # [c, h, w]
heat1 = heat1.data.cpu().numpy() # 将tensor格式的feature map转为numpy格式
# print(heat1.shape)
# heat = np.squeeze(heat, 0) # ０维为batch维度，由于是单张图片，所以batch=1，将这一维度删除
for j in range(64): # 64 channels
plt.subplot(8, 8, j + 1)
plt.axis('off')
channel_heat = heat1[j, :] # [h, w]切片获取某几个通道的特征图
heatmap = np.maximum(channel_heat, 0) # heatmap与0比较
# 用cv2加载原始图像
img = cv2.imread('/media/newamax/94d146aa-e21d-4f2d-ae9d-1f5444870820/wyting/projects/SMENet_res_FAM/VOCNWPU/VOC2012/JPEGImages/254.jpg')
img = cv2.resize(img, (400, 400))
heatmap = cv2.resize(heatmap, (img.shape[1], img.shape[0])) # 特征图的大小调整为与原始图像相同
heatmap = np.uint8(255 * heatmap) # 将特征图转换为uin
heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET) # 将特征图转为伪彩色图
heat_img = cv2.addWeighted(img, 0.5, heatmap, 0.5, 0) # 将伪彩色图与原始图片融合
#heat_img = heatmap * 0.5 + img # 也可以用这种方式融合
make_dirs(os.path.join(save_path, "per_channel_heatmap"))
cv2.imwrite(save_path + "/per_channel_heatmap/channel_" + str(j) + ".jpg", heat_img) # 将图像保存
show_fea = BGR_to_RGB(heat_img)
plt.imshow(show_fea)
fig.savefig(save_path + "/channels_heatmap.jpg", dpi=100)
fig.clf()
plt.close()
def make_dirs(path):
if os.path.exists(path) is False:
os.makedirs(path)
def get_feature(outs, save_path):
dst = './Visual_Features_detail'
therd_size = 400
fig = plt.figure(figsize=(8, 8))
fig.subplots_adjust(left=0.05, right=0.95, bottom=0.05, top=0.95, wspace=0.05, hspace=0.05)
for v in range(len(outs)): # 一列表返回字典可遍历的键，值 [键，值]
features = outs[0] # [C, H, W] => v[0,:,:,:]
size = features.shape[1]
for i in range(64):
# for i in range(iter_range):
plt.subplot(8, 8, i + 1)
plt.axis('off')
feature = features.cuda().data.cpu().numpy()
feature_img = feature[i, :, :] # 每一个通道的图 每一个通道显示为一幅独立的图
feature_img = np.maximum(feature_img, 0)
feature_img = np.asarray(feature_img * 255, dtype=np.uint8)
# if size == 19:
# print(str(size) + '_' + str(i))
# print(feature[i,:,:])
# print(feature_img)
dst_path = os.path.join(dst, str(size))
make_dirs(dst_path)
feature_img = cv2.applyColorMap(feature_img, cv2.COLORMAP_JET)
# 对图片形状的控制
if feature_img.shape[0] < therd_size: # 对图片尺寸的控制
tmp_file = os.path.join(dst_path, str(i) + '_' + str(therd_size) + '.png')
tmp_img = feature_img.copy()
tmp_img = cv2.resize(tmp_img, (therd_size, therd_size), interpolation=cv2.INTER_NEAREST)
cv2.imwrite(tmp_file, tmp_img)
dst_file = os.path.join(dst_path, str(i) + '.png')
cv2.imwrite(dst_file, feature_img)
show_fea = BGR_to_RGB(feature_img)
plt.imshow(show_fea)
fig.savefig(save_path, dpi=100)
fig.clf()
plt.close()