-
Notifications
You must be signed in to change notification settings - Fork 10
/
demo.py
209 lines (173 loc) · 7.97 KB
/
demo.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
import time
import numpy as np
from PIL import Image
import torch
import torch.nn as nn
import torchvision
from torchvision import transforms
import torchvision.transforms.functional as TF
import timm
assert "0.4.5" <= timm.__version__ <= "0.4.9" # version check
from util.misc import make_grid
import models_mae_cross
device = torch.device('cuda')
"""
python demo.py
"""
class measure_time(object):
def __enter__(self):
self.start = time.perf_counter_ns()
return self
def __exit__(self, typ, value, traceback):
self.duration = (time.perf_counter_ns() - self.start) / 1e9
def load_image():
im_dir = '/GPFS/data/changliu/Dataset/FSC147/images_384_VarV2'
im_id = '222.jpg'
image = Image.open('{}/{}'.format(im_dir, im_id))
image.load()
W, H = image.size
# Resize the image size so that the height is 384
new_H = 384
new_W = 16 * int((W / H * 384) / 16)
scale_factor_H = float(new_H) / H
scale_factor_W = float(new_W) / W
image = transforms.Resize((new_H, new_W))(image)
Normalize = transforms.Compose([transforms.ToTensor()])
image = Normalize(image)
# Coordinates of the exemplar bound boxes
# The left upper corner and the right lower corner
bboxes = [
[[136, 98], [173, 127]],
[[209, 125], [242, 150]],
[[212, 168], [258, 200]]
]
boxes = list()
rects = list()
for bbox in bboxes:
x1 = int(bbox[0][0] * scale_factor_W)
y1 = int(bbox[0][1] * scale_factor_H)
x2 = int(bbox[1][0] * scale_factor_W)
y2 = int(bbox[1][1] * scale_factor_H)
rects.append([y1, x1, y2, x2])
bbox = image[:, y1:y2 + 1, x1:x2 + 1]
bbox = transforms.Resize((64, 64))(bbox)
boxes.append(bbox.numpy())
boxes = np.array(boxes)
boxes = torch.Tensor(boxes)
return image, boxes, rects
def run_one_image(samples, boxes, pos, model):
_, _, h, w = samples.shape
s_cnt = 0
for rect in pos:
if rect[2] - rect[0] < 10 and rect[3] - rect[1] < 10:
s_cnt += 1
if s_cnt >= 1:
r_densities = []
r_images = []
r_images.append(TF.crop(samples[0], 0, 0, int(h / 3), int(w / 3))) # 1
r_images.append(TF.crop(samples[0], 0, int(w / 3), int(h / 3), int(w / 3))) # 3
r_images.append(TF.crop(samples[0], 0, int(w * 2 / 3), int(h / 3), int(w / 3))) # 7
r_images.append(TF.crop(samples[0], int(h / 3), 0, int(h / 3), int(w / 3))) # 2
r_images.append(TF.crop(samples[0], int(h / 3), int(w / 3), int(h / 3), int(w / 3))) # 4
r_images.append(TF.crop(samples[0], int(h / 3), int(w * 2 / 3), int(h / 3), int(w / 3))) # 8
r_images.append(TF.crop(samples[0], int(h * 2 / 3), 0, int(h / 3), int(w / 3))) # 5
r_images.append(TF.crop(samples[0], int(h * 2 / 3), int(w / 3), int(h / 3), int(w / 3))) # 6
r_images.append(TF.crop(samples[0], int(h * 2 / 3), int(w * 2 / 3), int(h / 3), int(w / 3))) # 9
pred_cnt = 0
with measure_time() as et:
for r_image in r_images:
r_image = transforms.Resize((h, w))(r_image).unsqueeze(0)
density_map = torch.zeros([h, w])
density_map = density_map.to(device, non_blocking=True)
start = 0
prev = -1
with torch.no_grad():
while start + 383 < w:
output, = model(r_image[:, :, :, start:start + 384], boxes, 3)
output = output.squeeze(0)
b1 = nn.ZeroPad2d(padding=(start, w - prev - 1, 0, 0))
d1 = b1(output[:, 0:prev - start + 1])
b2 = nn.ZeroPad2d(padding=(prev + 1, w - start - 384, 0, 0))
d2 = b2(output[:, prev - start + 1:384])
b3 = nn.ZeroPad2d(padding=(0, w - start, 0, 0))
density_map_l = b3(density_map[:, 0:start])
density_map_m = b1(density_map[:, start:prev + 1])
b4 = nn.ZeroPad2d(padding=(prev + 1, 0, 0, 0))
density_map_r = b4(density_map[:, prev + 1:w])
density_map = density_map_l + density_map_r + density_map_m / 2 + d1 / 2 + d2
prev = start + 383
start = start + 128
if start + 383 >= w:
if start == w - 384 + 128:
break
else:
start = w - 384
pred_cnt += torch.sum(density_map / 60).item()
r_densities += [density_map]
else:
density_map = torch.zeros([h, w])
density_map = density_map.to(device, non_blocking=True)
start = 0
prev = -1
with measure_time() as et:
with torch.no_grad():
while start + 383 < w:
output, = model(samples[:, :, :, start:start + 384], boxes, 3)
output = output.squeeze(0)
b1 = nn.ZeroPad2d(padding=(start, w - prev - 1, 0, 0))
d1 = b1(output[:, 0:prev - start + 1])
b2 = nn.ZeroPad2d(padding=(prev + 1, w - start - 384, 0, 0))
d2 = b2(output[:, prev - start + 1:384])
b3 = nn.ZeroPad2d(padding=(0, w - start, 0, 0))
density_map_l = b3(density_map[:, 0:start])
density_map_m = b1(density_map[:, start:prev + 1])
b4 = nn.ZeroPad2d(padding=(prev + 1, 0, 0, 0))
density_map_r = b4(density_map[:, prev + 1:w])
density_map = density_map_l + density_map_r + density_map_m / 2 + d1 / 2 + d2
prev = start + 383
start = start + 128
if start + 383 >= w:
if start == w - 384 + 128:
break
else:
start = w - 384
pred_cnt = torch.sum(density_map / 60).item()
e_cnt = 0
for rect in pos:
e_cnt += torch.sum(density_map[rect[0]:rect[2] + 1, rect[1]:rect[3] + 1] / 60).item()
e_cnt = e_cnt / 3
if e_cnt > 1.8:
pred_cnt /= e_cnt
# Visualize the prediction
fig = samples[0]
box_map = torch.zeros([fig.shape[1], fig.shape[2]])
box_map = box_map.to(device, non_blocking=True)
for rect in pos:
for i in range(rect[2] - rect[0]):
box_map[min(rect[0] + i, fig.shape[1] - 1), min(rect[1], fig.shape[2] - 1)] = 10
box_map[min(rect[0] + i, fig.shape[1] - 1), min(rect[3], fig.shape[2] - 1)] = 10
for i in range(rect[3] - rect[1]):
box_map[min(rect[0], fig.shape[1] - 1), min(rect[1] + i, fig.shape[2] - 1)] = 10
box_map[min(rect[2], fig.shape[1] - 1), min(rect[1] + i, fig.shape[2] - 1)] = 10
box_map = box_map.unsqueeze(0).repeat(3, 1, 1)
pred = density_map.unsqueeze(0).repeat(3, 1, 1) if s_cnt < 1 \
else make_grid(r_densities, h, w).unsqueeze(0).repeat(3, 1, 1)
fig = fig + box_map + pred / 2
fig = torch.clamp(fig, 0, 1)
torchvision.utils.save_image(fig, f'./Image/Visualisation.png')
# GT map needs coordinates for all GT dots, which is hard to input and is not a must for the demo. You can provide it yourself.
return pred_cnt, et
# Prepare model
model = models_mae_cross.__dict__['mae_vit_base_patch16'](norm_pix_loss='store_true')
model.to(device)
model_without_ddp = model
checkpoint = torch.load('./output_allnew_dir/checkpoint-400.pth', map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'], strict=False)
print("Resume checkpoint %s" % './output_allnew_dir/checkpoint-400.pth')
model.eval()
# Test on the new image
samples, boxes, pos = load_image()
samples = samples.unsqueeze(0).to(device, non_blocking=True)
boxes = boxes.unsqueeze(0).to(device, non_blocking=True)
result, elapsed_time = run_one_image(samples, boxes, pos, model)
print(result, elapsed_time.duration)