Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,178 +1,178 @@
|
|
1 |
-
# Copyright (C) 2020 * Ltd. All rights reserved.
|
2 |
-
# author : Sanghyeon Jo <josanghyeokn@gmail.com>
|
3 |
-
|
4 |
-
import gradio as gr
|
5 |
-
|
6 |
-
import os
|
7 |
-
import sys
|
8 |
-
import copy
|
9 |
-
import shutil
|
10 |
-
import random
|
11 |
-
import argparse
|
12 |
-
import numpy as np
|
13 |
-
|
14 |
-
import imageio
|
15 |
-
|
16 |
-
import torch
|
17 |
-
import torch.nn as nn
|
18 |
-
import torch.nn.functional as F
|
19 |
-
|
20 |
-
from torchvision import transforms
|
21 |
-
from torch.utils.tensorboard import SummaryWriter
|
22 |
-
|
23 |
-
from torch.utils.data import DataLoader
|
24 |
-
|
25 |
-
from core.puzzle_utils import *
|
26 |
-
from core.networks import *
|
27 |
-
from core.datasets import *
|
28 |
-
|
29 |
-
from tools.general.io_utils import *
|
30 |
-
from tools.general.time_utils import *
|
31 |
-
from tools.general.json_utils import *
|
32 |
-
|
33 |
-
from tools.ai.log_utils import *
|
34 |
-
from tools.ai.demo_utils import *
|
35 |
-
from tools.ai.optim_utils import *
|
36 |
-
from tools.ai.torch_utils import *
|
37 |
-
from tools.ai.evaluate_utils import *
|
38 |
-
|
39 |
-
from tools.ai.augment_utils import *
|
40 |
-
from tools.ai.randaugment import *
|
41 |
-
|
42 |
-
parser = argparse.ArgumentParser()
|
43 |
-
|
44 |
-
###############################################################################
|
45 |
-
# Dataset
|
46 |
-
###############################################################################
|
47 |
-
parser.add_argument('--seed', default=2606, type=int)
|
48 |
-
parser.add_argument('--num_workers', default=4, type=int)
|
49 |
-
parser.add_argument('--data_dir', default='../VOCtrainval_11-May-2012/', type=str)
|
50 |
-
|
51 |
-
###############################################################################
|
52 |
-
# Network
|
53 |
-
###############################################################################
|
54 |
-
parser.add_argument('--architecture', default='DeepLabv3+', type=str)
|
55 |
-
parser.add_argument('--backbone', default='resnet50', type=str)
|
56 |
-
parser.add_argument('--mode', default='fix', type=str)
|
57 |
-
parser.add_argument('--use_gn', default=True, type=str2bool)
|
58 |
-
|
59 |
-
###############################################################################
|
60 |
-
# Inference parameters
|
61 |
-
###############################################################################
|
62 |
-
parser.add_argument('--tag', default='', type=str)
|
63 |
-
|
64 |
-
parser.add_argument('--domain', default='val', type=str)
|
65 |
-
|
66 |
-
parser.add_argument('--scales', default='0.5,1.0,1.5,2.0', type=str)
|
67 |
-
parser.add_argument('--iteration', default=10, type=int)
|
68 |
-
|
69 |
-
if __name__ == '__main__':
|
70 |
-
###################################################################################
|
71 |
-
# Arguments
|
72 |
-
###################################################################################
|
73 |
-
args = parser.parse_args()
|
74 |
-
|
75 |
-
model_dir = create_directory('./experiments/models/')
|
76 |
-
model_path = model_dir + f'DeepLabv3+@
|
77 |
-
|
78 |
-
if 'train' in args.domain:
|
79 |
-
args.tag += '@train'
|
80 |
-
else:
|
81 |
-
args.tag += '@' + args.domain
|
82 |
-
|
83 |
-
args.tag += '@scale=%s' % args.scales
|
84 |
-
args.tag += '@iteration=%d' % args.iteration
|
85 |
-
|
86 |
-
set_seed(args.seed)
|
87 |
-
log_func = lambda string='': print(string)
|
88 |
-
|
89 |
-
###################################################################################
|
90 |
-
# Transform, Dataset, DataLoader
|
91 |
-
###################################################################################
|
92 |
-
imagenet_mean = [0.485, 0.456, 0.406]
|
93 |
-
imagenet_std = [0.229, 0.224, 0.225]
|
94 |
-
|
95 |
-
normalize_fn = Normalize(imagenet_mean, imagenet_std)
|
96 |
-
|
97 |
-
# for mIoU
|
98 |
-
meta_dic = read_json('./data/VOC_2012.json')
|
99 |
-
|
100 |
-
###################################################################################
|
101 |
-
# Network
|
102 |
-
###################################################################################
|
103 |
-
if args.architecture == 'DeepLabv3+':
|
104 |
-
model = DeepLabv3_Plus(args.backbone, num_classes=meta_dic['classes'] + 1, mode=args.mode,
|
105 |
-
use_group_norm=args.use_gn)
|
106 |
-
elif args.architecture == 'Seg_Model':
|
107 |
-
model = Seg_Model(args.backbone, num_classes=meta_dic['classes'] + 1)
|
108 |
-
elif args.architecture == 'CSeg_Model':
|
109 |
-
model = CSeg_Model(args.backbone, num_classes=meta_dic['classes'] + 1)
|
110 |
-
|
111 |
-
model = model.cuda()
|
112 |
-
model.eval()
|
113 |
-
|
114 |
-
log_func('[i] Architecture is {}'.format(args.architecture))
|
115 |
-
log_func('[i] Total Params: %.2fM' % (calculate_parameters(model)))
|
116 |
-
log_func()
|
117 |
-
|
118 |
-
load_model(model, model_path, parallel=False)
|
119 |
-
|
120 |
-
#################################################################################################
|
121 |
-
# Evaluation
|
122 |
-
#################################################################################################
|
123 |
-
eval_timer = Timer()
|
124 |
-
scales = [float(scale) for scale in args.scales.split(',')]
|
125 |
-
|
126 |
-
model.eval()
|
127 |
-
eval_timer.tik()
|
128 |
-
|
129 |
-
|
130 |
-
def inference(images, image_size):
|
131 |
-
images = images.cuda()
|
132 |
-
|
133 |
-
logits = model(images)
|
134 |
-
logits = resize_for_tensors(logits, image_size)
|
135 |
-
|
136 |
-
logits = logits[0] + logits[1].flip(-1)
|
137 |
-
logits = get_numpy_from_tensor(logits).transpose((1, 2, 0))
|
138 |
-
return logits
|
139 |
-
|
140 |
-
|
141 |
-
def predict_image(ori_image):
|
142 |
-
with torch.no_grad():
|
143 |
-
ori_w, ori_h = ori_image.size
|
144 |
-
|
145 |
-
cams_list = []
|
146 |
-
|
147 |
-
for scale in scales:
|
148 |
-
image = copy.deepcopy(ori_image)
|
149 |
-
image = image.resize((round(ori_w * scale), round(ori_h * scale)), resample=PIL.Image.BICUBIC)
|
150 |
-
|
151 |
-
image = normalize_fn(image)
|
152 |
-
image = image.transpose((2, 0, 1))
|
153 |
-
|
154 |
-
image = torch.from_numpy(image)
|
155 |
-
flipped_image = image.flip(-1)
|
156 |
-
|
157 |
-
images = torch.stack([image, flipped_image])
|
158 |
-
|
159 |
-
cams = inference(images, (ori_h, ori_w))
|
160 |
-
cams_list.append(cams)
|
161 |
-
|
162 |
-
preds = np.sum(cams_list, axis=0)
|
163 |
-
preds = F.softmax(torch.from_numpy(preds), dim=-1).numpy()
|
164 |
-
|
165 |
-
if args.iteration > 0:
|
166 |
-
preds = crf_inference(np.asarray(ori_image), preds.transpose((2, 0, 1)), t=args.iteration)
|
167 |
-
pred_mask = np.argmax(preds, axis=0)
|
168 |
-
else:
|
169 |
-
pred_mask = np.argmax(preds, axis=-1)
|
170 |
-
|
171 |
-
return pred_mask.astype(np.uint8)
|
172 |
-
|
173 |
-
|
174 |
-
demo = gr.Interface(
|
175 |
-
fn=predict_image,
|
176 |
-
inputs="image",
|
177 |
-
outputs="image"
|
178 |
-
)
|
|
|
1 |
+
# Copyright (C) 2020 * Ltd. All rights reserved.
|
2 |
+
# author : Sanghyeon Jo <josanghyeokn@gmail.com>
|
3 |
+
|
4 |
+
import gradio as gr
|
5 |
+
|
6 |
+
import os
|
7 |
+
import sys
|
8 |
+
import copy
|
9 |
+
import shutil
|
10 |
+
import random
|
11 |
+
import argparse
|
12 |
+
import numpy as np
|
13 |
+
|
14 |
+
import imageio
|
15 |
+
|
16 |
+
import torch
|
17 |
+
import torch.nn as nn
|
18 |
+
import torch.nn.functional as F
|
19 |
+
|
20 |
+
from torchvision import transforms
|
21 |
+
from torch.utils.tensorboard import SummaryWriter
|
22 |
+
|
23 |
+
from torch.utils.data import DataLoader
|
24 |
+
|
25 |
+
from core.puzzle_utils import *
|
26 |
+
from core.networks import *
|
27 |
+
from core.datasets import *
|
28 |
+
|
29 |
+
from tools.general.io_utils import *
|
30 |
+
from tools.general.time_utils import *
|
31 |
+
from tools.general.json_utils import *
|
32 |
+
|
33 |
+
from tools.ai.log_utils import *
|
34 |
+
from tools.ai.demo_utils import *
|
35 |
+
from tools.ai.optim_utils import *
|
36 |
+
from tools.ai.torch_utils import *
|
37 |
+
from tools.ai.evaluate_utils import *
|
38 |
+
|
39 |
+
from tools.ai.augment_utils import *
|
40 |
+
from tools.ai.randaugment import *
|
41 |
+
|
42 |
+
parser = argparse.ArgumentParser()
|
43 |
+
|
44 |
+
###############################################################################
|
45 |
+
# Dataset
|
46 |
+
###############################################################################
|
47 |
+
parser.add_argument('--seed', default=2606, type=int)
|
48 |
+
parser.add_argument('--num_workers', default=4, type=int)
|
49 |
+
parser.add_argument('--data_dir', default='../VOCtrainval_11-May-2012/', type=str)
|
50 |
+
|
51 |
+
###############################################################################
|
52 |
+
# Network
|
53 |
+
###############################################################################
|
54 |
+
parser.add_argument('--architecture', default='DeepLabv3+', type=str)
|
55 |
+
parser.add_argument('--backbone', default='resnet50', type=str)
|
56 |
+
parser.add_argument('--mode', default='fix', type=str)
|
57 |
+
parser.add_argument('--use_gn', default=True, type=str2bool)
|
58 |
+
|
59 |
+
###############################################################################
|
60 |
+
# Inference parameters
|
61 |
+
###############################################################################
|
62 |
+
parser.add_argument('--tag', default='', type=str)
|
63 |
+
|
64 |
+
parser.add_argument('--domain', default='val', type=str)
|
65 |
+
|
66 |
+
parser.add_argument('--scales', default='0.5,1.0,1.5,2.0', type=str)
|
67 |
+
parser.add_argument('--iteration', default=10, type=int)
|
68 |
+
|
69 |
+
if __name__ == '__main__':
|
70 |
+
###################################################################################
|
71 |
+
# Arguments
|
72 |
+
###################################################################################
|
73 |
+
args = parser.parse_args()
|
74 |
+
|
75 |
+
model_dir = create_directory('./experiments/models/')
|
76 |
+
model_path = model_dir + f'DeepLabv3+@ResNet-50@Fix@GN.pth'
|
77 |
+
|
78 |
+
if 'train' in args.domain:
|
79 |
+
args.tag += '@train'
|
80 |
+
else:
|
81 |
+
args.tag += '@' + args.domain
|
82 |
+
|
83 |
+
args.tag += '@scale=%s' % args.scales
|
84 |
+
args.tag += '@iteration=%d' % args.iteration
|
85 |
+
|
86 |
+
set_seed(args.seed)
|
87 |
+
log_func = lambda string='': print(string)
|
88 |
+
|
89 |
+
###################################################################################
|
90 |
+
# Transform, Dataset, DataLoader
|
91 |
+
###################################################################################
|
92 |
+
imagenet_mean = [0.485, 0.456, 0.406]
|
93 |
+
imagenet_std = [0.229, 0.224, 0.225]
|
94 |
+
|
95 |
+
normalize_fn = Normalize(imagenet_mean, imagenet_std)
|
96 |
+
|
97 |
+
# for mIoU
|
98 |
+
meta_dic = read_json('./data/VOC_2012.json')
|
99 |
+
|
100 |
+
###################################################################################
|
101 |
+
# Network
|
102 |
+
###################################################################################
|
103 |
+
if args.architecture == 'DeepLabv3+':
|
104 |
+
model = DeepLabv3_Plus(args.backbone, num_classes=meta_dic['classes'] + 1, mode=args.mode,
|
105 |
+
use_group_norm=args.use_gn)
|
106 |
+
elif args.architecture == 'Seg_Model':
|
107 |
+
model = Seg_Model(args.backbone, num_classes=meta_dic['classes'] + 1)
|
108 |
+
elif args.architecture == 'CSeg_Model':
|
109 |
+
model = CSeg_Model(args.backbone, num_classes=meta_dic['classes'] + 1)
|
110 |
+
|
111 |
+
model = model.cuda()
|
112 |
+
model.eval()
|
113 |
+
|
114 |
+
log_func('[i] Architecture is {}'.format(args.architecture))
|
115 |
+
log_func('[i] Total Params: %.2fM' % (calculate_parameters(model)))
|
116 |
+
log_func()
|
117 |
+
|
118 |
+
load_model(model, model_path, parallel=False)
|
119 |
+
|
120 |
+
#################################################################################################
|
121 |
+
# Evaluation
|
122 |
+
#################################################################################################
|
123 |
+
eval_timer = Timer()
|
124 |
+
scales = [float(scale) for scale in args.scales.split(',')]
|
125 |
+
|
126 |
+
model.eval()
|
127 |
+
eval_timer.tik()
|
128 |
+
|
129 |
+
|
130 |
+
def inference(images, image_size):
|
131 |
+
images = images.cuda()
|
132 |
+
|
133 |
+
logits = model(images)
|
134 |
+
logits = resize_for_tensors(logits, image_size)
|
135 |
+
|
136 |
+
logits = logits[0] + logits[1].flip(-1)
|
137 |
+
logits = get_numpy_from_tensor(logits).transpose((1, 2, 0))
|
138 |
+
return logits
|
139 |
+
|
140 |
+
|
141 |
+
def predict_image(ori_image):
|
142 |
+
with torch.no_grad():
|
143 |
+
ori_w, ori_h = ori_image.size
|
144 |
+
|
145 |
+
cams_list = []
|
146 |
+
|
147 |
+
for scale in scales:
|
148 |
+
image = copy.deepcopy(ori_image)
|
149 |
+
image = image.resize((round(ori_w * scale), round(ori_h * scale)), resample=PIL.Image.BICUBIC)
|
150 |
+
|
151 |
+
image = normalize_fn(image)
|
152 |
+
image = image.transpose((2, 0, 1))
|
153 |
+
|
154 |
+
image = torch.from_numpy(image)
|
155 |
+
flipped_image = image.flip(-1)
|
156 |
+
|
157 |
+
images = torch.stack([image, flipped_image])
|
158 |
+
|
159 |
+
cams = inference(images, (ori_h, ori_w))
|
160 |
+
cams_list.append(cams)
|
161 |
+
|
162 |
+
preds = np.sum(cams_list, axis=0)
|
163 |
+
preds = F.softmax(torch.from_numpy(preds), dim=-1).numpy()
|
164 |
+
|
165 |
+
if args.iteration > 0:
|
166 |
+
preds = crf_inference(np.asarray(ori_image), preds.transpose((2, 0, 1)), t=args.iteration)
|
167 |
+
pred_mask = np.argmax(preds, axis=0)
|
168 |
+
else:
|
169 |
+
pred_mask = np.argmax(preds, axis=-1)
|
170 |
+
|
171 |
+
return pred_mask.astype(np.uint8)
|
172 |
+
|
173 |
+
|
174 |
+
demo = gr.Interface(
|
175 |
+
fn=predict_image,
|
176 |
+
inputs="image",
|
177 |
+
outputs="image"
|
178 |
+
)
|