-
Notifications
You must be signed in to change notification settings - Fork 6
/
visual.py
182 lines (155 loc) · 6.28 KB
/
visual.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
# --------------------------------------
# -*- coding: utf-8 -*-
# @Time : 2022/8/24 15:39
# @Author : wzy
# @File : visual.py
# ---------------------------------------
import math
import cv2
import numpy as np
import torch
from matplotlib import pyplot as plt
from data import data_load, transform, data_to_tensor
from separate_num import separate
import matplotlib
import seaborn as sns
from einops import rearrange
from img_patch import patchify, masking
matplotlib.use('Agg') # 加上这句话plt.show就会报错。作用是控制绘图不显示
def vis_filter(model_weights, layer):
"""
:param model_weights: 传入整个模型的权重
:param layer: 选择可视化哪一层
:return:
"""
print('p====================卷积核可视化====================q')
filter_num = model_weights[layer].shape[0]
filter_channel = model_weights[layer].shape[1]
print(model_weights[layer].shape)
print(f'该层一共有[{filter_num}]个卷积核,每个卷积核的维度为[{filter_channel}]')
row, column = separate(filter_num)
for j in range(filter_channel):
plt.figure(figsize=(5, 5))
for i, filter in enumerate(model_weights[layer]):
plt.subplot(row, column, i + 1)
plt.imshow(filter[j, :, :].detach(), cmap='gray')
plt.axis('off')
plt.savefig(f'./imgs_out/filter_layer{layer + 1}_channel{j}.png')
plt.show()
print('b==================卷积核可视化结束===================d')
def vis_image(image):
h, w = image.shape[0], image.shape[1]
plt.subplots(figsize=(w * 0.01, h * 0.01))
plt.imshow(image, alpha=1)
# plt.axis('off')
# plt.show()
return h, w
def vis_feature(features, num_layers):
"""
:param features: 特征图(即图像经过卷积后的样子)
:param num_layers:所有的特征图层数
:return:
"""
print('p====================特征图可视化====================q')
for num_layer in range(len(features)):
plt.figure(figsize=(5, 5)) # 5*5有点小了，建议改大
layer_vis = features[num_layer][0, :, :, :]
layer_vis = layer_vis.data
print(f'[{num_layer + 1}] feature size :{layer_vis.size()}')
# 得到特征图通道数，也就是后面特征图的数量
feature_num = layer_vis.shape[0]
row, column = separate(feature_num)
for i, filter in enumerate(layer_vis):
plt.subplot(row, column, i + 1)
# plt.imshow(filter, cmap='gray') # 灰度图
plt.imshow(filter)
plt.axis('off')
print(f'Saving layer feature maps : [{num_layer + 1}]/[{num_layers}] ')
plt.savefig(f'./imgs_out/layer_{num_layer}.png')
plt.close()
print('b==================特征图可视化结束===================d')
def vis_attention_matrix(attention_map, index=0, cmap="YlGnBu"):
"""
:param attention_map: 注意力得分矩阵
:param index: map编号,便于多个注意力可视化的存储
:param cmap: 颜色样式
:return:
"""
plt.figure(figsize=(6, 5))
sns.heatmap(
attention_map,
vmin=0.0, vmax=1.0,
cmap=cmap,
# annot=True, # 每个格子上显示数据
square=True)
plt.savefig(f'./imgs_out/attention_matrix_{index}.png')
print(f'[attention_matrix_{index}.png] is generated')
def vis_grid_attention(img_path, attention_map, cmap='jet'):
"""
:param img_path:图像路径
:param attention_map:注意力图
:param cmap: cmap是图像的颜色类型，有很多预设的颜色类型
:return:
"""
# draw the img
img = data_load(img_path)
h, w = vis_image(img)
# draw the attention
map = cv2.resize(attention_map, (w, h))
normed_map = map / map.max()
normed_map = (normed_map * 255).astype('uint8')
plt.imshow(normed_map, alpha=0.4, interpolation='nearest', cmap=cmap) # alpha值决定图像的透明度,0为透明,1不透明
# 去掉图片周边白边
plt.subplots_adjust(top=1, bottom=0, right=1, left=0, hspace=0, wspace=0) # 调整图像与画布的边距(此时填充满)
plt.margins(0, 0)
# 保存图像,以300dpi
img_name = img_path.split('/')[-1].split('.')[0] + "_with_attention.jpg"
plt.savefig(f'./imgs_out/{img_name}', dpi=300)
print(f'[{img_name}] is generated')
# plt.show()
def vis_img_patch(patch_tensor, patch_size, mask=None):
"""
将图片划分成patch块并显示
:param patch_tensor: (B,NUM,patch_size^2*C) 其中NUM是patch块的数量
:param patch_size: patch的边长(默认正方形)
:param mask:masking函数得到的mask矩阵，代表哪些patch被mask
:return:
"""
if mask is not None:
mask = mask.detach()
mask = mask.unsqueeze(-1).repeat(1, 1, 16 ** 2 * 3) # (N, H*W, p*p*3)
patch_tensor = patch_tensor * (1 - mask)
print(patch_tensor.shape)
patch_tensor = patch_tensor.squeeze(0)
n, d = patch_tensor.shape
h = w = int(math.sqrt(n))
plt.figure(figsize=(5, 5))
# way-1 单个patch分离明显(建议查看patch效果时使用)
# divide the img from (NUM_patch,D)->(NUM_patch,H_patch,W_patch,C) C放在后面便于plt处理
# img = rearrange(patch_tensor, 'n (n1 n2 c) -> n n1 n2 c', n1=patch_size, n2=patch_size)
# for i in range(h * w):
# plt.subplot(h, w, i + 1)
# plt.imshow(img[i].data.numpy(), cmap='gray')
# plt.axis('off')
# way-2 patch紧密相连(建议查看mask效果时使用)
img = rearrange(patch_tensor, '(h w) (n1 n2 c) -> (h n1) (w n2) c', h=h, w=w, n1=patch_size, n2=patch_size)
plt.imshow(img)
plt.axis('off')
if mask is None:
plt.savefig(f'./imgs_out/img_patch.png')
else:
plt.savefig(f'./imgs_out/masked_patch.png')
if __name__ == '__main__':
attention_map = np.zeros((20, 20))
attention_map[9][9] = 1
attention_map[10][12] = 1
vis_grid_attention(img_path="./imgs_in/dog_1.jpg", attention_map=attention_map)
attention_map = np.random.normal(size=(10, 10))
vis_attention_matrix(attention_map)
# 进行patch划分及可视化
img = data_to_tensor(img_path="./imgs_in/dog_1.jpg")
img = patchify(img, patch_size=16)
vis_img_patch(img, patch_size=16)
# 进行mask及可视化
x, mask, ids_restore = masking(img, 0.75)
vis_img_patch(img, patch_size=16, mask=mask)