-
Notifications
You must be signed in to change notification settings - Fork 2
/
demo_.py
171 lines (127 loc) · 5.1 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
import os
import cv2
import time
import glob
import torch
import numpy as np
from PIL import Image
from torchvision import transforms as tfs
def demo_image_transforms(demo_image):
transform_demo = tfs.Compose([tfs.ToTensor(),
tfs.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])])
demo_image = transform_demo(demo_image)
demo_image = demo_image.unsqueeze(0) # make batch
return demo_image
@ torch.no_grad()
def demo(opts, device, model):
# 1. make tensors
demo_image_list = glob.glob(os.path.join(opts.demo_root, '*' + '.' + opts.demo_image_type))
total_time = 0
# 2. load .pth
model.eval()
for idx, img_path in enumerate(demo_image_list):
# --------------------- img load ---------------------
demo_image_pil = Image.open(img_path).convert('RGB')
demo_image = demo_image_transforms(demo_image_pil).to(device)
tic = time.time()
pred_boxes, pred_labels, pred_scores = model.module.predict(demo_image, opts)
im_show = visualize_detection_result(demo_image_pil, pred_boxes, pred_labels, pred_scores)
# save_files
demo_result_path = os.path.join(opts.demo_root, 'detection_results')
os.makedirs(demo_result_path, exist_ok=True)
# 0 ~ 1 image -> 0~255 image
im_show = cv2.convertScaleAbs(im_show, alpha=(255.0))
cv2.imwrite(os.path.join(demo_result_path, os.path.basename(img_path)), im_show)
if opts.demo_vis:
cv2.imshow('demo_results', im_show)
cv2.waitKey(0)
toc = time.time()
inference_time = toc - tic
total_time += inference_time
if idx % 100 == 0 or idx == len(demo_image_list) - 1:
# ------------------- check fps -------------------
print('Step: [{}/{}]'.format(idx, len(demo_image_list)))
print("fps : {:.4f}".format((idx + 1) / total_time))
print("complete detection...!")
return
def visualize_detection_result(x, bbox, label, score):
'''
x : pil image range - [0 255], uint8
bbox : np.array, [num_obj, 4], float32
label : np.array, [num_obj] int32
score : np.array, [num_obj] float32
'''
img_width, img_height = x.size
multiplier = np.array([img_width, img_height, img_width, img_height])
bbox *= multiplier
# 2. uint8 -> float32
image_np = np.array(x).astype(np.float32) / 255.
x_img = image_np
im_show = cv2.cvtColor(x_img, cv2.COLOR_RGB2BGR)
for j in range(len(bbox)):
if opts.data_type == 'voc':
from utils import voc_color_array, voc_label_map
label_list = list(voc_label_map.keys())
color_array = voc_color_array
elif opts.data_type == 'coco':
from utils import coco_color_array, coco_label_map
label_list = list(coco_label_map.keys())
color_array = coco_color_array
x_min = int(bbox[j][0])
y_min = int(bbox[j][1])
x_max = int(bbox[j][2])
y_max = int(bbox[j][3])
cv2.rectangle(im_show,
pt1=(x_min, y_min),
pt2=(x_max, y_max),
color=color_array[label[j]],
thickness=2)
# text_size
text_size = cv2.getTextSize(text=label_list[label[j]] + ' {:.2f}'.format(score[j].item()),
fontFace=cv2.FONT_HERSHEY_PLAIN,
fontScale=1,
thickness=1)[0]
# text_rec
cv2.rectangle(im_show,
pt1=(x_min, y_min),
pt2=(x_min + text_size[0] + 3, y_min + text_size[1] + 4),
color=color_array[label[j]],
thickness=-1)
# put text
cv2.putText(im_show,
text=label_list[label[j]] + ' {:.2f}'.format(score[j].item()),
org=(x_min + 10, y_min + 10), # must be int
fontFace=0,
fontScale=0.4,
color=(0, 0, 0))
return im_show
import configargparse
from models.build import build_model
from config import get_args_parser
def demo_worker(rank, opts):
# 1. config
print(opts)
# 2. device
device = torch.device('cuda:{}'.format(int(opts.gpu_ids[opts.rank])))
# 5. model
if opts.data_type == 'voc':
opts.num_classes = 21
if opts.data_type == 'coco':
opts.num_classes = 81
from models.model_ import FRCNN
# model = FRCNN(pretrained=True, num_classes=opts.num_classes)
model = FRCNN(pretrained=True, num_classes=81)
model = torch.nn.DataParallel(module=model, device_ids=[int(id) for id in opts.gpu_ids])
model = model.to(device)
demo(opts=opts,
device=device,
model=model,
)
if __name__ == '__main__':
parser = configargparse.ArgumentParser('faster rcnn demo', parents=[get_args_parser()])
opts = parser.parse_args()
opts.world_size = len(opts.gpu_ids)
opts.num_workers = len(opts.gpu_ids) * 4
print(opts)
demo_worker(0, opts)