forked from XPixelGroup/BasicSR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
video_base_model.py
172 lines (155 loc) · 7.05 KB
/
video_base_model.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
import importlib
import torch
from collections import Counter
from copy import deepcopy
from os import path as osp
from torch import distributed as dist
from tqdm import tqdm
from basicsr.models.sr_model import SRModel
from basicsr.utils import get_root_logger, imwrite, tensor2img
from basicsr.utils.dist_util import get_dist_info
metric_module = importlib.import_module('basicsr.metrics')
class VideoBaseModel(SRModel):
"""Base video SR model."""
def dist_validation(self, dataloader, current_iter, tb_logger, save_img):
dataset = dataloader.dataset
dataset_name = dataset.opt['name']
with_metrics = self.opt['val']['metrics'] is not None
# initialize self.metric_results
# It is a dict: {
# 'folder1': tensor (num_frame x len(metrics)),
# 'folder2': tensor (num_frame x len(metrics))
# }
if with_metrics and not hasattr(self, 'metric_results'):
self.metric_results = {}
num_frame_each_folder = Counter(dataset.data_info['folder'])
for folder, num_frame in num_frame_each_folder.items():
self.metric_results[folder] = torch.zeros(
num_frame,
len(self.opt['val']['metrics']),
dtype=torch.float32,
device='cuda')
rank, world_size = get_dist_info()
if with_metrics:
for _, tensor in self.metric_results.items():
tensor.zero_()
# record all frames (border and center frames)
if rank == 0:
pbar = tqdm(total=len(dataset), unit='frame')
for idx in range(rank, len(dataset), world_size):
val_data = dataset[idx]
val_data['lq'].unsqueeze_(0)
val_data['gt'].unsqueeze_(0)
folder = val_data['folder']
frame_idx, max_idx = val_data['idx'].split('/')
lq_path = val_data['lq_path']
self.feed_data(val_data)
self.test()
visuals = self.get_current_visuals()
result_img = tensor2img([visuals['result']])
if 'gt' in visuals:
gt_img = tensor2img([visuals['gt']])
del self.gt
# tentative for out of GPU memory
del self.lq
del self.output
torch.cuda.empty_cache()
if save_img:
if self.opt['is_train']:
raise NotImplementedError(
'saving image is not supported during training.')
else:
if 'vimeo' in dataset_name.lower(): # vimeo90k dataset
split_result = lq_path.split('/')
img_name = (f'{split_result[-3]}_{split_result[-2]}_'
f'{split_result[-1].split(".")[0]}')
else: # other datasets, e.g., REDS, Vid4
img_name = osp.splitext(osp.basename(lq_path))[0]
if self.opt['val']['suffix']:
save_img_path = osp.join(
self.opt['path']['visualization'], dataset_name,
folder,
f'{img_name}_{self.opt["val"]["suffix"]}.png')
else:
save_img_path = osp.join(
self.opt['path']['visualization'], dataset_name,
folder, f'{img_name}_{self.opt["name"]}.png')
imwrite(result_img, save_img_path)
if with_metrics:
# calculate metrics
opt_metric = deepcopy(self.opt['val']['metrics'])
for metric_idx, opt_ in enumerate(opt_metric.values()):
metric_type = opt_.pop('type')
result = getattr(metric_module,
metric_type)(result_img, gt_img, **opt_)
self.metric_results[folder][int(frame_idx),
metric_idx] += result
# progress bar
if rank == 0:
for _ in range(world_size):
pbar.update(1)
pbar.set_description(
f'Test {folder}:'
f'{int(frame_idx) + world_size}/{max_idx}')
if rank == 0:
pbar.close()
if with_metrics:
if self.opt['dist']:
# collect data among GPUs
for _, tensor in self.metric_results.items():
dist.reduce(tensor, 0)
dist.barrier()
else:
pass # assume use one gpu in non-dist testing
if rank == 0:
self._log_validation_metric_values(current_iter, dataset_name,
tb_logger)
def nondist_validation(self, dataloader, current_iter, tb_logger,
save_img):
logger = get_root_logger()
logger.warning(
'nondist_validation is not implemented. Run dist_validation.')
self.dist_validation(dataloader, current_iter, tb_logger, save_img)
def _log_validation_metric_values(self, current_iter, dataset_name,
tb_logger):
# average all frames for each sub-folder
# metric_results_avg is a dict:{
# 'folder1': tensor (len(metrics)),
# 'folder2': tensor (len(metrics))
# }
metric_results_avg = {
folder: torch.mean(tensor, dim=0).cpu()
for (folder, tensor) in self.metric_results.items()
}
# total_avg_results is a dict: {
# 'metric1': float,
# 'metric2': float
# }
total_avg_results = {
metric: 0
for metric in self.opt['val']['metrics'].keys()
}
for folder, tensor in metric_results_avg.items():
for idx, metric in enumerate(total_avg_results.keys()):
total_avg_results[metric] += metric_results_avg[folder][
idx].item()
# average among folders
for metric in total_avg_results.keys():
total_avg_results[metric] /= len(metric_results_avg)
log_str = f'Validation {dataset_name}\n'
for metric_idx, (metric,
value) in enumerate(total_avg_results.items()):
log_str += f'\t # {metric}: {value:.4f}'
for folder, tensor in metric_results_avg.items():
log_str += f'\t # {folder}: {tensor[metric_idx].item():.4f}'
log_str += '\n'
logger = get_root_logger()
logger.info(log_str)
if tb_logger:
for metric_idx, (metric,
value) in enumerate(total_avg_results.items()):
tb_logger.add_scalar(f'metrics/{metric}', value, current_iter)
for folder, tensor in metric_results_avg.items():
tb_logger.add_scalar(f'metrics/{metric}/{folder}',
tensor[metric_idx].item(),
current_iter)