-
Notifications
You must be signed in to change notification settings - Fork 5
/
attvis.py
138 lines (121 loc) · 5.86 KB
/
attvis.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
import sys
from PIL import Image
import argparse
import os
import numpy as np
import torch
import cv2
torch.set_printoptions(sci_mode=False, precision=4)
np.set_printoptions(suppress=True, precision=4)
def save_matrix(filename, mat, print_stats=False):
import matplotlib.pyplot as plt
# corr = F.avg_pool2d(corr, 4, stride=4).squeeze(1).squeeze(0)
if print_stats:
print("{}: {}. mean/std: {:.5f}, {:.5f}".format(filename, list(mat.shape),
np.abs(mat).mean(), mat.std()))
plt.imshow(mat)
plt.colorbar()
plt.savefig(filename) # dpi=1200
plt.clf()
print(f"Saved '{filename}'")
def get_boundary(h, w, H, W, radius):
top = max(0, h - radius)
bottom = min(H, h + radius + 1)
left = max(0, w - radius)
right = min(W, w + radius + 1)
return top, bottom, left, right
def vis_attention(model_name, img1_path, img2_path, points, attention5d_path,
radius=16, box_radius=8, img_scale=1, alpha=1, savedir='attvis',
proj_img2=False):
img2_name = os.path.basename(img2_path)
img2_trunk = os.path.splitext(img2_name)[0]
if img1_path is not None:
img1_np = cv2.imread(img1_path)
img1_name = os.path.basename(img1_path)
img1_trunk = os.path.splitext(img1_name)[0]
img1_np = cv2.resize(img1_np, (0,0), fx=img_scale, fy=img_scale)
else:
img1_np = None
img2_np = cv2.imread(img2_path)[:,:,::-1]
img2_np = cv2.resize(img2_np, (0,0), fx=img_scale, fy=img_scale)
H, W = img2_np.shape[:2]
attention5d = torch.load(attention5d_path, map_location='cpu')
if not os.path.exists(savedir):
os.makedirs(savedir, exist_ok=True)
for point in points:
w0, h0 = point
w0 = int(w0 * img_scale)
h0 = int(h0 * img_scale)
h, w = h0 // 8, w0 // 8
if img_scale != 1:
print(f"{point[0]}, {point[1]} => {w0}, {h0} => {w}, {h}")
else:
print(f"{w0}, {h0} => {w}, {h}")
# attention: H//8, W//8
attention = attention5d[0, h, w].numpy()
# Set attention outside the radius to 0.
if radius > 0:
mask = np.zeros_like(attention, dtype=bool)
attn_top, attn_bottom, attn_left, attn_right = get_boundary(h, w, H//8, W//8, radius)
mask[attn_top:attn_bottom, attn_left:attn_right] = True
attention = attention * mask.astype(float)
median = np.median(attention[mask])
else:
median = np.median(attention)
neg_count = np.count_nonzero(attention < 0)
pos_count = np.count_nonzero(attention > 0)
print(f"{point}: median {median}, {pos_count} > 0, {neg_count} < 0")
box_top, box_bottom, box_left, box_right = get_boundary(h0, w0, H, W, radius=box_radius)
if img1_np is not None:
# draw a square around the point
# the side length of the square is 2*radius+1
blank_rect = np.copy(img1_np)
cv2.rectangle(blank_rect, (box_left, box_top), (box_right, box_bottom), (0, 0, 255), 1)
img1_np2 = cv2.addWeighted(img1_np, (1-alpha), blank_rect, alpha, 0)
img1_savename = f"{img1_trunk}-{point[0]},{point[1]}-highlight.png"
img1_savepath = os.path.join(savedir, img1_savename)
cv2.imwrite(img1_savepath, img1_np2)
print(f"Saved '{img1_savepath}'")
attention = cv2.resize(attention, (W, H))
attention -= median
attention[attention < 0] = 0
attention = (255 * attention / attention.max()).astype(np.uint8)
# heatmap: [368, 768, 3]
heatmap = cv2.applyColorMap(attention, cv2.COLORMAP_JET)[:, :, ::-1]
overlaid_img2 = img2_np * 0.6 + heatmap * 0.3
overlaid_img2 = overlaid_img2.astype(np.uint8)
blank_rect = overlaid_img2.copy()
# self attention on Frame-2, draw a red rectangle.
if img1_path == img2_path:
color = (255, 0, 0)
cv2.rectangle(blank_rect, (box_left, box_top), (box_right, box_bottom), color, 1)
# Cross-frame attention. If proj_img2, draw a green rectangle in Frame-2
# at the same location of the query in Frame-1.
elif proj_img2:
color = (0, 255, 0)
cv2.rectangle(blank_rect, (box_left, box_top), (box_right, box_bottom), color, 1)
overlaid_img2 = cv2.addWeighted(overlaid_img2, (1-alpha), blank_rect, alpha, 0)
overlaid_img2_obj = Image.fromarray(overlaid_img2)
img2_savename = f"{img2_trunk}-{point[0]},{point[1]}-{model_name}.png"
img2_savepath = os.path.join(savedir, img2_savename)
overlaid_img2_obj.save(img2_savepath)
print(f"Saved '{img2_savepath}'")
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--model', dest="model_name", type=str)
parser.add_argument('--img1', dest='img1_path', type=str)
parser.add_argument('--img2', dest='img2_path', type=str)
# --points is a list of tuples. specified as: --points 11,22.44,77.33,15
parser.add_argument('--points', type=str)
parser.add_argument('--att', dest='attention5d_path', type=str, required=True)
parser.add_argument('--savedir', type=str, default='attvis')
parser.add_argument('--scale', dest='img_scale', type=float, default=1.0)
parser.add_argument('--radius', dest='radius', type=int, default=16)
parser.add_argument('--box_radius', dest='box_radius', type=int, default=8)
parser.add_argument('--alpha', type=float, default=1)
parser.add_argument('--proj_img2', action='store_true')
args = parser.parse_args()
points = args.points.split(".")
points = [[int(x) for x in p.split(",")] for p in points]
vis_attention(args.model_name, args.img1_path, args.img2_path, points, args.attention5d_path,
args.radius, args.box_radius, args.img_scale, args.alpha, args.savedir, args.proj_img2)