-
Notifications
You must be signed in to change notification settings - Fork 4
/
inference_3D_LUT_Interpolation.py
241 lines (178 loc) · 10.6 KB
/
inference_3D_LUT_Interpolation.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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
import os
import argparse
import albumentations
from albumentations import Resize
import torch
import torch.backends.cudnn as cudnn
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from model.build_model import build_model
from datasets.build_dataset import dataset_generator
from utils import misc, metrics
import numpy as np
from model.lut_transformation_net import TrilinearInterpolation
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--workers', type=int, default=1,
metavar='N', help='Dataloader threads.')
parser.add_argument('--batch_size', type=int, default=1,
help='You can override model batch size by specify positive number.')
parser.add_argument('--device', type=str, default='cuda',
help="Whether use cuda, 'cuda' or 'cpu'.")
parser.add_argument('--save_path', type=str, default="./logs",
help='Where to save logs and checkpoints.')
parser.add_argument('--dataset_path', type=str, default="/HYouTube/",
help='Dataset path.')
parser.add_argument('--base_size', type=int, default=256,
help='Base size. Resolution of the image input into the Encoder')
parser.add_argument('--input_size', type=int, default=256,
help='Input size. Resolution of the image that want to be generated by the Decoder')
parser.add_argument('--INR_input_size', type=int, default=256,
help='INR input size. Resolution of the image that want to be generated by the Decoder. '
'Should be the same as `input_size`')
parser.add_argument('--INR_MLP_dim', type=int, default=32,
help='Number of channels for INR linear layer.')
parser.add_argument('--LUT_dim', type=int, default=7,
help='Dim of the output LUT. Refer to https://ieeexplore.ieee.org/abstract/document/9206076')
parser.add_argument('--activation', type=str, default='leakyrelu_pe',
help='INR activation layer type: leakyrelu_pe, sine')
parser.add_argument('--pretrained', type=str,
default=r'.\pretrained_models\Video_HYouTube_256.pth',
help='Pretrained weight path')
parser.add_argument('--param_factorize_dim', type=int,
default=10,
help='The intermediate dimensions of the factorization of the predicted MLP parameters. '
'Refer to https://arxiv.org/abs/2011.12026')
parser.add_argument('--embedding_type', type=str,
default="CIPS_embed",
help='Which embedding_type to use.')
parser.add_argument('--optim', type=str,
default='adamw',
help='Which optimizer to use.')
parser.add_argument('--INRDecode', action="store_false",
help='Whether INR decoder. Set it to False if you want to test the baseline '
'(https://github.com/SamsungLabs/image_harmonization)')
parser.add_argument('--isMoreINRInput', action="store_false",
help='Whether to cat RGB and mask. See Section 3.4 in the paper.')
parser.add_argument('--hr_train', action="store_true",
help='Whether use hr_train. See section 3.4 in the paper.')
parser.add_argument('--isFullRes', action="store_true",
help='Whether for original resolution. See section 3.4 in the paper.')
opt = parser.parse_args()
opt.save_path = misc.increment_path(os.path.join(opt.save_path, "test1"))
return opt
def select(checkpoint, curr_step, lut_list):
distance = sum([abs(curr_step - i) for i in checkpoint])
weight = [distance / abs(curr_step - i + 1e-6) for i in checkpoint]
normal_weight = [i / sum(weight) for i in weight]
return sum([i * j for i, j in zip(normal_weight, lut_list)])
"This is the implementation of our 3D LUT interpolation strategy. See Section V-C of our paper for more details."
def inference(val_loader, model, logger, opt):
current_process = 10
model.eval()
# You can change this to your own selected key frames. We will only calculate the 3D LUT of these key frames.
checkpoint = [0, 5, 10, 15, 19]
metric_log = {
'HYouTube': {'Samples': 0, 'MSE': 0, 'fMSE': 0, 'PSNR': 0, 'SSIM': 0},
'All': {'Samples': 0, 'MSE': 0, 'fMSE': 0, 'PSNR': 0, 'SSIM': 0},
}
lut_metric_log = {
'HYouTube': {'Samples': 0, 'MSE': 0, 'fMSE': 0, 'PSNR': 0, 'SSIM': 0},
'All': {'Samples': 0, 'MSE': 0, 'fMSE': 0, 'PSNR': 0, 'SSIM': 0},
}
all_intermediate = {"lut": [], "real": [], "comp": [], "lut_harmon": [], 'mask': [], 'name': []}
for step, batch in enumerate(val_loader):
composite_image = batch['composite_image'].to(opt.device)
real_image = batch['real_image'].to(opt.device)
mask = batch['mask'].to(opt.device)
category = batch['category']
fg_INR_coordinates = batch['fg_INR_coordinates'].to(opt.device)
"The number 20 here denotes the total frame number of one video clip in HYouTube."
if step % 20 in checkpoint:
with torch.no_grad():
_, lut, _ = model(
composite_image,
mask,
fg_INR_coordinates,
)
all_intermediate['lut'].append(lut)
all_intermediate['comp'].append(composite_image)
all_intermediate['real'].append(real_image)
all_intermediate['mask'].append(mask)
all_intermediate['name'].append(batch['file_path'])
if (step + 1) % 20 == 0:
# We do the 3D LUT interpolation here.
temp_lut = [select(checkpoint, x, all_intermediate['lut']) for x in range(20)]
all_intermediate['lut'] = temp_lut
for id in range(len(all_intermediate['lut'])):
lut_transform_image = misc.normalize(torch.stack(
[TrilinearInterpolation(lut, image)[0] for lut, image in
zip(all_intermediate['lut'][id], misc.normalize(all_intermediate['comp'][id], opt, 'inv'))],
dim=0), opt)
mask = all_intermediate['mask'][id]
real_image = all_intermediate['real'][id]
lut_transform_image = lut_transform_image * (mask > 100 / 255.) + real_image * (~(mask > 100 / 255.))
# We here save the outputs as npy files. You can also save them as images by custom changes.
# video, obj, img_number = all_intermediate['name'][id][0].replace('\\', '/').split('/')[-3:]
#
# result_name = os.path.join(
# "./HYouTube_Result",
# video + '_' + obj + '_' + img_number[:-3] + 'npy')
# np.save(result_name, misc.normalize(lut_transform_image, opt, 'inv').detach().cpu().numpy())
lut_mse, lut_fmse, lut_psnr, lut_ssim = metrics.calc_metrics(
misc.normalize(lut_transform_image, opt, 'inv'),
misc.normalize(real_image, opt, 'inv'), mask)
for idx in range(len(category)):
lut_metric_log[category[idx]]['Samples'] += 1
lut_metric_log[category[idx]]['MSE'] += lut_mse[idx]
lut_metric_log[category[idx]]['fMSE'] += lut_fmse[idx]
lut_metric_log[category[idx]]['PSNR'] += lut_psnr[idx]
lut_metric_log[category[idx]]['SSIM'] += lut_ssim[idx]
lut_metric_log['All']['Samples'] += 1
lut_metric_log['All']['MSE'] += lut_mse[idx]
lut_metric_log['All']['fMSE'] += lut_fmse[idx]
lut_metric_log['All']['PSNR'] += lut_psnr[idx]
lut_metric_log['All']['SSIM'] += lut_ssim[idx]
if (step + 1) / len(val_loader) * 100 >= current_process:
logger.info(f'Processing: {current_process}')
current_process += 10
all_intermediate = {"lut": [], "real": [], "comp": [], "lut_harmon": [], 'mask': [], 'name': []}
logger.info('=========================')
for key in metric_log.keys():
if opt.INRDecode:
msg = f"{key}-'LUT_MSE': {lut_metric_log[key]['MSE'] / lut_metric_log[key]['Samples']:.2f}\n" \
f"{key}-'LUT_fMSE': {lut_metric_log[key]['fMSE'] / lut_metric_log[key]['Samples']:.2f}\n" \
f"{key}-'LUT_PSNR': {lut_metric_log[key]['PSNR'] / lut_metric_log[key]['Samples']:.2f}\n" \
f"{key}-'LUT_SSIM': {lut_metric_log[key]['SSIM'] / lut_metric_log[key]['Samples']:.4f}\n"
else:
msg = f"{key}-'LUT_MSE': {lut_metric_log[key]['MSE'] / lut_metric_log[key]['Samples']:.2f}\n" \
f"{key}-'LUT_fMSE': {lut_metric_log[key]['fMSE'] / lut_metric_log[key]['Samples']:.2f}\n" \
f"{key}-'LUT_PSNR': {lut_metric_log[key]['PSNR'] / lut_metric_log[key]['Samples']:.2f}\n" \
f"{key}-'LUT_SSIM': {lut_metric_log[key]['SSIM'] / lut_metric_log[key]['Samples']:.4f}\n"
logger.info(msg)
logger.info('=========================')
def main_process(opt):
logger = misc.create_logger(os.path.join(opt.save_path, "log.txt"))
cudnn.benchmark = True
valset_path = os.path.join(opt.dataset_path, "test_list.txt")
opt.transform_mean = [.5, .5, .5]
opt.transform_var = [.5, .5, .5]
torch_transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize(opt.transform_mean, opt.transform_var)])
valset_alb_transform = albumentations.Compose([Resize(opt.input_size, opt.input_size)],
additional_targets={'real_image': 'image', 'object_mask': 'image'})
valset = dataset_generator(valset_path, valset_alb_transform, torch_transform, opt, mode='Val')
val_loader = DataLoader(valset, opt.batch_size, shuffle=False, drop_last=False, pin_memory=True,
num_workers=opt.workers, persistent_workers=True)
model = build_model(opt).to(opt.device)
logger.info(f"Load pretrained weight from {opt.pretrained}")
load_dict = torch.load(opt.pretrained)['model']
for k in load_dict.keys():
if k not in model.state_dict().keys():
print(f"Skip {k}")
model.load_state_dict(load_dict, strict=False)
inference(val_loader, model, logger, opt)
if __name__ == '__main__':
opt = parse_args()
os.makedirs(opt.save_path, exist_ok=True)
main_process(opt)