BasicSR

1
import torch
2
from collections import Counter
3
from os import path as osp
4
from torch import distributed as dist
5
from tqdm import tqdm
6

7
from basicsr.metrics import calculate_metric
8
from basicsr.utils import get_root_logger, imwrite, tensor2img
9
from basicsr.utils.dist_util import get_dist_info
10
from basicsr.utils.registry import MODEL_REGISTRY
11
from .sr_model import SRModel
12

13

14
@MODEL_REGISTRY.register()
15
class VideoBaseModel(SRModel):
16
    """Base video SR model."""
17

18
    def dist_validation(self, dataloader, current_iter, tb_logger, save_img):
19
        dataset = dataloader.dataset
20
        dataset_name = dataset.opt['name']
21
        with_metrics = self.opt['val']['metrics'] is not None
22
        # initialize self.metric_results
23
        # It is a dict: {
24
        #    'folder1': tensor (num_frame x len(metrics)),
25
        #    'folder2': tensor (num_frame x len(metrics))
26
        # }
27
        if with_metrics:
28
            if not hasattr(self, 'metric_results'):  # only execute in the first run
29
                self.metric_results = {}
30
                num_frame_each_folder = Counter(dataset.data_info['folder'])
31
                for folder, num_frame in num_frame_each_folder.items():
32
                    self.metric_results[folder] = torch.zeros(
33
                        num_frame, len(self.opt['val']['metrics']), dtype=torch.float32, device='cuda')
34
            # initialize the best metric results
35
            self._initialize_best_metric_results(dataset_name)
36
        # zero self.metric_results
37
        rank, world_size = get_dist_info()
38
        if with_metrics:
39
            for _, tensor in self.metric_results.items():
40
                tensor.zero_()
41

42
        metric_data = dict()
43
        # record all frames (border and center frames)
44
        if rank == 0:
45
            pbar = tqdm(total=len(dataset), unit='frame')
46
        for idx in range(rank, len(dataset), world_size):
47
            val_data = dataset[idx]
48
            val_data['lq'].unsqueeze_(0)
49
            val_data['gt'].unsqueeze_(0)
50
            folder = val_data['folder']
51
            frame_idx, max_idx = val_data['idx'].split('/')
52
            lq_path = val_data['lq_path']
53

54
            self.feed_data(val_data)
55
            self.test()
56
            visuals = self.get_current_visuals()
57
            result_img = tensor2img([visuals['result']])
58
            metric_data['img'] = result_img
59
            if 'gt' in visuals:
60
                gt_img = tensor2img([visuals['gt']])
61
                metric_data['img2'] = gt_img
62
                del self.gt
63

64
            # tentative for out of GPU memory
65
            del self.lq
66
            del self.output
67
            torch.cuda.empty_cache()
68

69
            if save_img:
70
                if self.opt['is_train']:
71
                    raise NotImplementedError('saving image is not supported during training.')
72
                else:
73
                    if 'vimeo' in dataset_name.lower():  # vimeo90k dataset
74
                        split_result = lq_path.split('/')
75
                        img_name = f'{split_result[-3]}_{split_result[-2]}_{split_result[-1].split(".")[0]}'
76
                    else:  # other datasets, e.g., REDS, Vid4
77
                        img_name = osp.splitext(osp.basename(lq_path))[0]
78

79
                    if self.opt['val']['suffix']:
80
                        save_img_path = osp.join(self.opt['path']['visualization'], dataset_name, folder,
81
                                                 f'{img_name}_{self.opt["val"]["suffix"]}.png')
82
                    else:
83
                        save_img_path = osp.join(self.opt['path']['visualization'], dataset_name, folder,
84
                                                 f'{img_name}_{self.opt["name"]}.png')
85
                imwrite(result_img, save_img_path)
86

87
            if with_metrics:
88
                # calculate metrics
89
                for metric_idx, opt_ in enumerate(self.opt['val']['metrics'].values()):
90
                    result = calculate_metric(metric_data, opt_)
91
                    self.metric_results[folder][int(frame_idx), metric_idx] += result
92

93
            # progress bar
94
            if rank == 0:
95
                for _ in range(world_size):
96
                    pbar.update(1)
97
                    pbar.set_description(f'Test {folder}: {int(frame_idx) + world_size}/{max_idx}')
98
        if rank == 0:
99
            pbar.close()
100

101
        if with_metrics:
102
            if self.opt['dist']:
103
                # collect data among GPUs
104
                for _, tensor in self.metric_results.items():
105
                    dist.reduce(tensor, 0)
106
                dist.barrier()
107
            else:
108
                pass  # assume use one gpu in non-dist testing
109

110
            if rank == 0:
111
                self._log_validation_metric_values(current_iter, dataset_name, tb_logger)
112

113
    def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
114
        logger = get_root_logger()
115
        logger.warning('nondist_validation is not implemented. Run dist_validation.')
116
        self.dist_validation(dataloader, current_iter, tb_logger, save_img)
117

118
    def _log_validation_metric_values(self, current_iter, dataset_name, tb_logger):
119
        # ----------------- calculate the average values for each folder, and for each metric  ----------------- #
120
        # average all frames for each sub-folder
121
        # metric_results_avg is a dict:{
122
        #    'folder1': tensor (len(metrics)),
123
        #    'folder2': tensor (len(metrics))
124
        # }
125
        metric_results_avg = {
126
            folder: torch.mean(tensor, dim=0).cpu()
127
            for (folder, tensor) in self.metric_results.items()
128
        }
129
        # total_avg_results is a dict: {
130
        #    'metric1': float,
131
        #    'metric2': float
132
        # }
133
        total_avg_results = {metric: 0 for metric in self.opt['val']['metrics'].keys()}
134
        for folder, tensor in metric_results_avg.items():
135
            for idx, metric in enumerate(total_avg_results.keys()):
136
                total_avg_results[metric] += metric_results_avg[folder][idx].item()
137
        # average among folders
138
        for metric in total_avg_results.keys():
139
            total_avg_results[metric] /= len(metric_results_avg)
140
            # update the best metric result
141
            self._update_best_metric_result(dataset_name, metric, total_avg_results[metric], current_iter)
142

143
        # ------------------------------------------ log the metric ------------------------------------------ #
144
        log_str = f'Validation {dataset_name}\n'
145
        for metric_idx, (metric, value) in enumerate(total_avg_results.items()):
146
            log_str += f'\t # {metric}: {value:.4f}'
147
            for folder, tensor in metric_results_avg.items():
148
                log_str += f'\t # {folder}: {tensor[metric_idx].item():.4f}'
149
            if hasattr(self, 'best_metric_results'):
150
                log_str += (f'\n\t    Best: {self.best_metric_results[dataset_name][metric]["val"]:.4f} @ '
151
                            f'{self.best_metric_results[dataset_name][metric]["iter"]} iter')
152
            log_str += '\n'
153

154
        logger = get_root_logger()
155
        logger.info(log_str)
156
        if tb_logger:
157
            for metric_idx, (metric, value) in enumerate(total_avg_results.items()):
158
                tb_logger.add_scalar(f'metrics/{metric}', value, current_iter)
159
                for folder, tensor in metric_results_avg.items():
160
                    tb_logger.add_scalar(f'metrics/{metric}/{folder}', tensor[metric_idx].item(), current_iter)
161