WIPA-Face-Super-Resolution/evaluate.py at main · hamidreza-dastmalchi/WIPA-Face-Super-Resolution · 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
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
from skimage.metrics import peak_signal_noise_ratio, structural_similarity
from Dataset_LH import Dataset_LH
from torch.utils.data import DataLoader
from modules import KFold, eval_acc
from modules import find_best_threshold
from modules import low_generation
import torchvision.transforms as transforms
from fid import calculate_fid
import torch.nn as nn
import numpy as np
import torch
import os
import cv2
import time
from tqdm import tqdm


def wipa_psnr_ssim(test_dataloader, gen_net, args):
    if torch.cuda.is_available() and not args.disable_cuda:
        device = torch.device("cuda")
    else:
        device = torch.device("cpu")
    ssim_mean = 0
    psnr_mean = 0
    counter = 0
    save_folder = os.path.join(args.save_folder, str(args.scale)+"x")
    if not os.path.isdir(save_folder):
        os.makedirs(save_folder)
    print("psnr and ssim is calculated on test dataset, please wait...")
    for count, data in enumerate(tqdm(test_dataloader)):
        high_image, low_image, image_name = data
        low_image = low_image.to(device)
        low_image = 2 * (low_image - 0.5)
        with torch.no_grad():
            sr_image = gen_net(low_image)
            if args.wavelet_integrated:
                sr_image, sr_wavelet = sr_image
        sr_image = sr_image / 2 + 0.5
        sr_image = sr_image.cpu()
        for i in range(high_image.size()[0]):
            sr_image_i = sr_image[i]
            high_image_i = high_image[i]
            sr_image_i = sr_image_i.numpy().transpose(1, 2, 0)
            high_image_i = high_image_i.numpy().transpose(1, 2, 0)
            psnr = peak_signal_noise_ratio(high_image_i, sr_image_i)
            ssim = structural_similarity(high_image_i, sr_image_i, multichannel=True)
            if args.save_flag:
                sr_image_i=(sr_image_i*255).astype(np.uint8)
                sr_image_i=cv2.cvtColor(sr_image_i, cv2.COLOR_RGB2BGR)
                cv2.imwrite(os.path.join(save_folder, image_name[i]), sr_image_i)
            psnr_mean += psnr
            ssim_mean += ssim
            counter += 1
    ssim_mean = ssim_mean / counter
    psnr_mean = psnr_mean / counter
    return psnr_mean, ssim_mean


def fps_estimation(gen_net, args):
    if torch.cuda.is_available() and not args.disable_cuda:
        device = torch.device("cuda")
    else:
        device = torch.device("cpu")
    test_dataset=Dataset_LH(args.test_root, None, args.scale)
    test_dataloader=DataLoader(test_dataset, batch_size=1, shuffle=False)
    str_time=time.time()
    for count, data in enumerate(test_dataloader):
        high_image, low_image, image_name = data
        low_image = low_image.to(device)
        low_image = 2 * (low_image - 0.5)
        with torch.no_grad():
            sr_image = gen_net(low_image)
            if args.wavelet_integrated:
                sr_image, sr_wavelet = sr_image
        sr_image = sr_image / 2 + 0.5
        if count==99:
            break
    end_time=time.time()
    time_dur=end_time-str_time
    fps=100/time_dur
    return fps


def fid_estimation(args):
    orig_file=args.test_root
    sr_file=os.path.join(args.save_folder,str(args.scale)+"x")
    orig_dataset=Dataset_LH(orig_file, transforms.Resize(299), args.scale)
    orig_len=len(orig_dataset)
    orig_dataloader=DataLoader(orig_dataset, batch_size=16, shuffle=False)
    sr_dataset = Dataset_LH(sr_file, transforms.Resize(299), args.scale)
    sr_len=len(sr_dataset)
    sr_dataloader = DataLoader(sr_dataset, batch_size=16, shuffle=False)
    fid_value = calculate_fid(orig_dataloader, sr_dataloader, False, orig_len, sr_len)
    return fid_value


def face_ver(lfw_data_loader, gen_net, fnet,args):
    features1_total = []
    features2_total = []
    labels = []
    with torch.no_grad():
        bs_total = 0
        print("calculating the identities")
        for index, (img1, img2, targets) in enumerate(tqdm(lfw_data_loader)):
            bs = len(targets)
            img2_low = low_generation(img2, args.scale)
            img2_low = 2 * (img2_low - 0.5)
            img2_low = img2_low.cuda()
            img2 = gen_net(img2_low)
            if args.wavelet_integrated:
                img2, _ = img2
            img1 = img1.cuda()
            img2 = img2.cuda()
            img1 = 2*(img1-0.5)
            img1 = img1 * 127.5 / 128
            img2 = img2 * 127.5 / 128
            img1 = img1[:, :, 9:120, 17:112]
            img2 = img2[:, :, 9:120, 17:112]
            features1 = fnet(img1).cpu()
            features2 = fnet(img2).cpu()
            features1_total += [features1]
            features2_total += [features2]
            labels += [targets]
            bs_total += bs
        #the identity features of two sets is calculated
        features1_total = torch.cat(features1_total)
        features2_total = torch.cat(features2_total)
        labels = torch.cat(labels)
        assert bs_total == 6000, print('LFW pairs should be 6,000')
        labels = labels.cpu().numpy()
        scores = nn.CosineSimilarity()(features1_total, features2_total)
        scores = scores.cpu().numpy().reshape(-1, 1)
        accuracy = []
        thd = []
        folds = KFold(n=6000, n_folds=10)
        thresholds = np.linspace(-10000, 10000, 10000 + 1)
        thresholds = thresholds / 10000
        true_positive_rate=[]
        false_positive_rate=[]
        thresholds2 = np.linspace(10000, -10000, 10000 + 1)/10000
        positive_number=labels[labels==1].sum()
        negative_number=labels.shape[0]-positive_number
        print("calculating the verification rate. please wait...")
        #true positive rate and false positive rate is calculated
        for thresh in thresholds2:
            score_mask=scores>thresh
            true_positive_number=labels[score_mask].sum()
            positive_detected_number=labels[score_mask].shape[0]
            false_positive_number=positive_detected_number-true_positive_number
            true_positive_rate.append(true_positive_number/positive_number)
            false_positive_rate.append(false_positive_number/negative_number)
        true_positive_rate_np=np.asarray(true_positive_rate)
        false_positive_rate_np=np.asarray(false_positive_rate)
        #tpr_dict = {"tpr_"+algorithm+str(scale): true_positive_rate_np }
        #fpr_dict = {"fpr_" + algorithm+str(scale): false_positive_rate_np}
        #savemat("tpr_"+algorithm+str(scale), tpr_dict)
        #savemat("fpr_" + algorithm+str(scale), fpr_dict)

    predicts = np.hstack((scores, labels))
    #finding the best threshold and verification rate in each test fold based on the rest of training folds
    for idx, (train, test) in enumerate(folds):
        best_thresh = find_best_threshold(thresholds, predicts[train])
        accuracy.append(eval_acc(best_thresh, predicts[test]))
    thd.append(best_thresh)
    mean_acc, std = np.mean(accuracy), np.std(accuracy)
    #print("mean accuracy of face verification on lfw dataset is : {0:.4f}".format(mean_acc))
    return true_positive_rate_np, false_positive_rate_np, mean_acc