forked from boostcampaitech2/final-project-level3-cv-12
-
Notifications
You must be signed in to change notification settings - Fork 0
/
inference.py
154 lines (119 loc) · 5.35 KB
/
inference.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
import os
import json
import random
import pytz
import numpy as np
import torch
import torch.nn as nn
import cv2
from tqdm import tqdm
from datetime import datetime
from module_fold import CEModule, ISModule
from manifold.manifold import KNN, ConstrainedLeastSquareSolver
# ==========================================================
# -- basic function
# ==========================================================
def seed_everything(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(seed)
random.seed(seed)
def get_time():
d = datetime.now(pytz.timezone('Asia/Seoul'))
return f'{d.month:0>2}{d.day:0>2}_{d.hour:0>2}{d.minute:0>2}'
# ==========================================================
# -- helper function _ get models
# ==========================================================
def get_encoder(encoder_path, part, device):
encoder = CEModule.define_part_encoder(part=part)
checkpoint = torch.load(encoder_path)
state_dict = checkpoint.state_dict()
encoder.load_state_dict(state_dict)
return encoder.to(device)
def get_decoder(decoder_path, part, device):
decoder = CEModule.define_part_decoder(part=part)
checkpoint = torch.load(decoder_path)
state_dict = checkpoint.state_dict()
decoder.load_state_dict(state_dict)
return decoder.to(device)
def get_generator(generator_path, device):
generator = ISModule.Generator_U(input_nc=5, output_nc=3, ngf=56,
n_blocks=9, norm_layer=nn.BatchNorm2d, padding_type='reflect')
checkpoint = torch.load(generator_path)
state_dict = checkpoint.state_dict()
generator.load_state_dict(state_dict)
return generator.to(device)
def get_knn(fv_json_path):
fv_json = json.load(open(fv_json_path, 'r'))
fv = [np.array(fv_json[i], dtype=np.float32) for i in fv_json]
return KNN(fv)
# ==========================================================
# -- helper function _ data processing
# ==========================================================
def fv_proj(fv, knn, least_square, device, t, k):
v_in = np.array(fv.detach().cpu().view(-1), dtype=np.float32)
v_k = knn(v_in, k)
ret = least_square(v_in, v_k)
ret = torch.FloatTensor(ret).to(device).unsqueeze(axis=0)
return t * fv + (1 - t) * ret
def save_image(img, img_path):
img = img.squeeze(axis=0) * 255
img = img.detach().cpu().numpy()
img = np.transpose(img, (1, 2, 0))
cv2.imwrite(img_path, img)
# ==========================================================
# -- main function
# ==========================================================
pos = [(256, 256), (302, 256), (244, 326), (244, 186), (385, 256)]
sz = [512, 160, 128, 128, 128]
def inference(img, encoder, decoder, generator, knn, least_square, device, t, k):
img = torch.FloatTensor(img).to(device)
img = img.unsqueeze(axis=0)
img = img.unsqueeze(axis=1)
fv = encoder(img).to(device)
fv = fv_proj(fv, knn, least_square, device, t, k)
img = decoder(fv)
sketch = torch.FloatTensor(np.zeros((1, 5, 512, 512))).to(device)
img_patch = [img[:, :, pos[i][0]-sz[i]//2:pos[i][0]+sz[i]//2, pos[i][1]-sz[i]//2:pos[i][1]+sz[i]//2].clone() for i in range(5)]
for i in range(1, 5): img_patch[0][:, :, pos[i][0]-sz[i]//2:pos[i][0]+sz[i]//2, pos[i][1]-sz[i]//2:pos[i][1]+sz[i]//2] = 0
for i in range(5): sketch[:, i, pos[i][0]-sz[i]//2:pos[i][0]+sz[i]//2, pos[i][1]-sz[i]//2:pos[i][1]+sz[i]//2] = img_patch[i]
output = generator(sketch)
for i in range(5): sketch[:, 0, pos[i][0]-sz[i]//2:pos[i][0]+sz[i]//2, pos[i][1]-sz[i]//2:pos[i][1]+sz[i]//2] = img_patch[i]
return output, sketch[:, 0:1, :, :]
def test(encoder_path, decoder_path, generator_path, fv_json_path, t, k):
# -- setting
seed_everything(42)
use_cuda = torch.cuda.is_available()
device = torch.device('cuda' if use_cuda else 'cpu')
device = torch.device('cpu')
# -- encoder, decoder, generator, knn, least_square
encoder = get_encoder(encoder_path, 'all', device)
decoder = get_decoder(decoder_path, 'all', device)
generator = get_generator(generator_path, device)
knn = get_knn(fv_json_path)
least_square = ConstrainedLeastSquareSolver()
# -- img_path
save_img_path = f'/opt/ml/project/inference_save/{get_time()}_{t}_{k}'
load_img_path = '/opt/ml/project/inference_save/test'
os.makedirs(save_img_path, exist_ok=True)
print(save_img_path)
for img_name in tqdm(sorted(os.listdir(load_img_path))):
img_path = os.path.join(load_img_path, img_name)
img = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE).astype(float)
img = 1 - img / 255
output, sketch = inference(img, encoder, decoder, generator, knn, least_square, device, t, k)
save_image(output, os.path.join(save_img_path, 'output_' + img_name))
save_image(1 - sketch, os.path.join(save_img_path, 'sketch_' + img_name))
if __name__ == '__main__':
# -- file_path: all
encoder_path = '/opt/ml/project/model_save/encoder.pth'
decoder_path = '/opt/ml/project/model_save/decoder.pth'
generator_path = '/opt/ml/project/model_save/generator.pth'
fv_json_path = '/opt/ml/project/data/fv_all_train.json'
# -- parameter
t = 0.4
k = 10
# -- inference
test(encoder_path, decoder_path, generator_path, fv_json_path, t, k)