Dynamic-CSD/dynamic_test.py at main · dmcv-ecnu/Dynamic-CSD · GitHub

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
import os
from glob import glob
import datetime
import time

from option import args
from model.edsr import EDSR
from model.carn import CARN
from model.rcan import RCAN
import data
from utils import utility

import random
import torch
import torch.nn as nn
import torch.nn.functional as F

def evaluate(sr_model, test_loader, width_list, args, geneinfo, device, self_ensemble=False, save_results=False):
    scale = args.scale[0]
    sr_model.eval()
    upsampler = nn.Upsample(scale_factor=scale, mode='bicubic')
    width_cnt = {}
    for width in width_list:
        width_cnt[width] = 0
    with torch.no_grad():
        psnr = 0
        niqe_score = 0
        ssim = 0
        total_flops = 0

        for d in test_loader:
            for lr, hr, filename in d:
                lr = lr.to(device)
                hr = hr.to(device)

                llr = F.interpolate(lr, scale_factor=1.0 / scale, mode='bicubic').clamp(min=0, max=args.rgb_range)
                lbic = upsampler(llr)
                lbic = utility.quantize(lbic, args.rgb_range)
                crop_w = lr.shape[-2] - lbic.shape[-2]
                crop_h = lr.shape[-1] - lbic.shape[-1]
                lhr = lr
                if crop_w > 0:
                    lhr = lhr[:, :, int(crop_w / 2):-(crop_w - int(crop_w / 2)), :]
                if crop_h > 0:
                    lhr = lhr[:, :, :, int(crop_h / 2):-(crop_h - int(crop_h / 2))]
                bic_psnr = utility.calc_psnr(lbic, lhr, scale, args.rgb_range, dataset=d)
                # print(bic_psnr)

                width_mult = width_list[-1]
                for i, gene in enumerate(geneinfo):
                    if bic_psnr > gene:
                        width_mult = width_list[i]
                        break
                width_cnt[width_mult] += 1
                print(filename, bic_psnr, width_mult)

                x = [lr]
                for tf in 'v', 'h', 't':
                    x.extend([utility.transform(_x, tf, device) for _x in x])
                op = ['', 'v', 'h', 'hv', 't', 'tv', 'th', 'thv']

                if self_ensemble:
                    res, flops, _ = sr_model(lr, width_mult)
                    total_flops += flops
                    # res, _ = self.model(lr, width_mult)
                    for i in range(1, len(x)):
                        _x = x[i]
                        # _sr, _ = self.model(_x, width_mult)
                        _sr, flops, _ = sr_model(_x, width_mult)
                        total_flops += flops
                        for _op in op[i]:
                            _sr = utility.transform(_sr, _op, device)
                        res = torch.cat((res, _sr), 0)
                    sr = torch.mean(res, 0).unsqueeze(0)
                else:
                    sr, flops, _ = sr_model(lr, width_mult)
                    total_flops += flops
                    # sr, _ = self.model(lr, width_mult)

                sr = utility.quantize(sr, args.rgb_range)
                if save_results:
                    #     if not os.path.exists(f'./output/test/{self.model_str}/{self.model_filename}'):
                    #         self.makedirs = os.makedirs(f'./output/test/{self.model_str}/{self.model_filename}')
                    utility.save_results(str(filename), sr, scale, width_mult,
                                         args.rgb_range, 'SR')
                    utility.save_results(str(filename), lbic, scale, width_mult,
                                         args.rgb_range, 'LBic')

                # print(
                #     f"{filename}\t {width_mult}\t {utility.calc_psnr(sr, hr, self.scale, self.rgb_range, dataset=d)}\t ")
                psnr += utility.calc_psnr(sr, hr, scale, args.rgb_range, dataset=d)

            psnr /= len(d)
            niqe_score /= len(d)
            ssim /= len(d)
            # print(width_mult, d.dataset.name, psnr, niqe_score, ssim)
            # print(f"Total FLOPs: {total_flops}")

            print(width_cnt)
            return psnr, total_flops/len(d)

def GA_test():
    width_list = [0.25, 0.5, 0.75, 1.0]

    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    # sr_model = CARN(args).to(device)
    # sr_model = EDSR(args).to(device)
    sr_model = RCAN(args).to(device)
    sr_model = nn.DataParallel(sr_model, device_ids=range(args.n_GPUs))
    checkpoint_dir = args.pre_train
    print(f"[*] Load model from {checkpoint_dir}")
    if not os.listdir(checkpoint_dir):
        print(f"[!] No checkpoint in {checkpoint_dir}")
        return
    model = glob(os.path.join(checkpoint_dir, f'{args.model_filename}.pth'))

    model_state_dict = torch.load(model[0])
    sr_model.load_state_dict(model_state_dict['model_state_dict'], False)
    loader = data.Data(args)

    threshold = args.threshold
    starttime = time.time()
    psnr, flops = evaluate(sr_model, loader.loader_test, width_list, args,
                           threshold, device)
    endtime = time.time()
    cost = (endtime - starttime)
    print(f"Time cost:{cost}")
    print(f"Evaluate on {args.data_test}: psnr:{psnr}, flops:{flops}")

if __name__ == '__main__':
    GA_test()