def compute_lpips(self, source, target):
im1 = lpips.im2tensor(
lpips.load_image(self.base_path + '/train_images/' + source))
im2 = lpips.im2tensor(
lpips.load_image(self.base_path + '/train_images/' + target))
dist = self.loss_fn.forward(im1, im2)
return dist.detach().numpy()
default='./out/',
help='path to save files')
opt = parser.parse_args()
os.makedirs(opt.out, exist_ok=True)
data = []
## Initializing the model
loss_fn = lpips.LPIPS(net=opt.net, version=opt.version)
if (opt.use_gpu):
loss_fn.cuda()
# Load images
img0 = lpips.im2tensor(lpips.load_image(opt.path)) # RGB image from [-1,1]
if (opt.use_gpu):
img0 = img0.cuda()
files = os.listdir(opt.folder)
for (ff, file) in enumerate(files[:-1]):
img1 = lpips.im2tensor(lpips.load_image(os.path.join(
opt.folder, file))) # RGB image from [-1,1]
if (opt.use_gpu):
img1 = img1.cuda()
# Compute distance
dist01 = loss_fn.forward(img0, img1)
loss_fn = lpips.LPIPS(net='alex', spatial=spatial) # Can also set net = 'squeeze' or 'vgg'
# loss_fn = lpips.LPIPS(net='alex', spatial=spatial, lpips=False) # Can also set net = 'squeeze' or 'vgg'
if(use_gpu):
loss_fn.cuda()
## Example usage with dummy tensors
dummy_im0 = torch.zeros(1,3,64,64) # image should be RGB, normalized to [-1,1]
dummy_im1 = torch.zeros(1,3,64,64)
if(use_gpu):
dummy_im0 = dummy_im0.cuda()
dummy_im1 = dummy_im1.cuda()
dist = loss_fn.forward(dummy_im0,dummy_im1)
## Example usage with images
ex_ref = lpips.im2tensor(lpips.load_image('./imgs/ex_ref.png'))
ex_p0 = lpips.im2tensor(lpips.load_image('./imgs/ex_p0.png'))
ex_p1 = lpips.im2tensor(lpips.load_image('./imgs/ex_p1.png'))
if(use_gpu):
ex_ref = ex_ref.cuda()
ex_p0 = ex_p0.cuda()
ex_p1 = ex_p1.cuda()
ex_d0 = loss_fn.forward(ex_ref,ex_p0)
ex_d1 = loss_fn.forward(ex_ref,ex_p1)
if not spatial:
print('Distances: (%.3f, %.3f)'%(ex_d0, ex_d1))
else:
print('Distances: (%.3f, %.3f)'%(ex_d0.mean(), ex_d1.mean())) # The mean distance is approximately the same as the non-spatial distance
list_avg_dist = []
f = open(opt.out, 'w')
for i in range(200):
# crawl directories
img_dir = os.path.join(opt.dir, str(i))
files = os.listdir(img_dir)
if opt.N is not None:
files = files[:opt.N]
F = len(files)
dists = []
for (ff, file) in enumerate(files[:-1]):
img0 = lpips.im2tensor(lpips.load_image(os.path.join(
img_dir, file))) # RGB image from [-1,1]
if (opt.use_gpu):
img0 = img0.cuda()
files1 = files[ff + 1:]
for file1 in files1:
img1 = lpips.im2tensor(
lpips.load_image(os.path.join(img_dir, file1)))
if (opt.use_gpu):
img1 = img1.cuda()
# Compute distance
dist01 = loss_fn.forward(img0, img1)
print('(%s,%s): %.3f' % (file, file1, dist01))
opt = parser.parse_args()
## Initializing the model
loss_fn = lpips.LPIPS(net='alex', version=opt.version)
if (opt.use_gpu):
loss_fn.cuda()
# crawl directories
f = open(opt.out, 'w')
files = os.listdir(opt.dir0)
dists = []
for file in files:
if (os.path.exists(os.path.join(opt.dir1, file))):
# Load images
img0 = lpips.im2tensor(lpips.load_image(os.path.join(
opt.dir0, file))) # RGB image from [-1,1]
img1 = lpips.im2tensor(lpips.load_image(os.path.join(opt.dir1, file)))
if (opt.use_gpu):
img0 = img0.cuda()
img1 = img1.cuda()
# Compute distance
dist01 = loss_fn.forward(img0, img1)
print('%s: %.3f' % (file, dist01))
f.writelines('%s: %.6f\n' % (file, dist01))
dists.append(dist01.cpu().detach().numpy())
mean = np.mean(np.array(dists))
print(mean)
f.close()
loss_fn = lpips.LPIPS(net='alex', version=opt.version)
if (opt.use_gpu):
loss_fn.cuda()
# Crawl directories
f = open(opt.out, 'w')
aligned_images = sorted(os.listdir(opt.dir0))
generated_images = sorted(os.listdir(opt.dir1))
for idx, image in enumerate(aligned_images):
if (idx == 0):
#print("\n")
pass
if (image[:-4] == generated_images[idx][:-4]):
# Load images
img0 = lpips.im2tensor(lpips.load_image(os.path.join(
opt.dir0, image))) # RGB image from [-1,1]
img1 = lpips.im2tensor(
lpips.load_image(os.path.join(opt.dir1, generated_images[idx])))
if (opt.use_gpu):
img0 = img0.cuda()
img1 = img1.cuda()
# Compute distance
dist01 = loss_fn.forward(img0, img1)
#print("LPIPS score (distance) for ", image[:-4], " is:\t", round(dist01, 2), "\n")
#print('LPIPS score (distance) for %s is:\t%.2f\n'%(image[:-4], dist01))
f.writelines('LPIPS score (distance) for %s is:\t%.2f\n' %
(image[:-4], dist01))
f.close()
def main():
# Configurations
# GT - Ground-truth;
# Gen: Generated / Restored / Recovered images
folder_GT = '/data/wcz93762/houselee/OriRealSR(V3)/Canon/Test/HRx4'
folder_Gen = '/home/yzj6850/HouseLee/Test-SR/no_pre_realsr_66000'
f = open('realsr_ours_Result_new.txt', 'a+')
loss_fn = lpips.LPIPS(net='alex')
loss_fn.cuda()
crop_border = 4
suffix = '' # suffix for Gen images
test_Y = False # True: test Y channel only; False: test RGB channels
PSNR_all = []
# PSNR_all_2 = []
SSIM_all = []
LPIPS_all = []
# SSIM_all_2 = []
img_list = sorted(glob.glob(folder_GT + '/*'))
if test_Y:
print('Testing Y channel.')
else:
print('Testing RGB channels.')
for i, img_path in enumerate(img_list):
base_name = os.path.splitext(os.path.basename(img_path))[0]
#Test LPIPS
img0 = lpips.im2tensor(lpips.load_image(img_path))
img1 = lpips.im2tensor(
lpips.load_image(
os.path.join(folder_Gen, base_name + suffix + '.png')))
img0 = img0.cuda()
img1 = img1.cuda()
dist01 = loss_fn.forward(img0, img1)
lp = dist01.cpu().detach().numpy()
LPIPS_all.append(lp)
im_GT = cv2.imread(img_path) / 255.
im_Gen = cv2.imread(
os.path.join(folder_Gen, base_name + suffix + '.png')) / 255.
# im_bic = cv2.imread(os.path.join(folder_bic, base_name + suffix + '.png')) / 255.
if test_Y and im_GT.shape[
2] == 3: # evaluate on Y channel in YCbCr color space
im_GT_in = bgr2ycbcr(im_GT)
im_Gen_in = bgr2ycbcr(im_Gen)
# im_bic_in = bgr2ycbcr(im_bic)
else:
im_GT_in = im_GT
im_Gen_in = im_Gen
# im_bic_in = im_bic
# crop borders
if im_GT_in.ndim == 3:
cropped_GT = im_GT_in[crop_border:-crop_border,
crop_border:-crop_border, :]
cropped_Gen = im_Gen_in[crop_border:-crop_border,
crop_border:-crop_border, :]
# cropped_bic = im_bic_in[crop_border:-crop_border, crop_border:-crop_border, :]
elif im_GT_in.ndim == 2:
cropped_GT = im_GT_in[crop_border:-crop_border,
crop_border:-crop_border]
cropped_Gen = im_Gen_in[crop_border:-crop_border,
crop_border:-crop_border]
# cropped_bic = im_bic_in[crop_border:-crop_border, crop_border:-crop_border]
else:
raise ValueError(
'Wrong image dimension: {}. Should be 2 or 3.'.format(
im_GT_in.ndim))
# calculate PSNR and SSIM
PSNR = calculate_psnr(cropped_GT * 255, cropped_Gen * 255)
# PSNR2 = calculate_psnr(cropped_GT * 255, cropped_bic * 255)
SSIM = calculate_ssim(cropped_GT * 255, cropped_Gen * 255)
# SSIM2 = calculate_ssim(cropped_GT * 255, cropped_bic * 255)
#select pred_better and pred_worse patches
# if PSNR1-PSNR2 >= 0.7:
#if len(glob.glob(folder_pred = '*.png')) <= 200:
# shutil.copy(folder_Gen + '/' + base_name + suffix + '.png', folder_pred)
# if PSNR1-PSNR2 <= 0.3:
#if len(glob.glob(folder_ori = '*.png')) <= 200:
# shutil.copy(folder_Gen + '/' + base_name + suffix + '.png', folder_ori)
print(
'{:3d} - {:25}. \tPSNR: {:.6f} dB, \tSSIM: {:.6f}, \tLPIPS: {:.6f}'
.format(i + 1, base_name, PSNR, SSIM, lp[0][0][0][0]))
f.writelines(
'{:3d} - {:25}. \tPSNR: {:.6f} dB, \tSSIM: {:.6f}, \tLPIPS: {:.6f}\n'
.format(i + 1, base_name, PSNR, SSIM, lp[0][0][0][0]))
PSNR_all.append(PSNR)
SSIM_all.append(SSIM)
avg_lpips = sum(LPIPS_all) / len(LPIPS_all)
print('Average: PSNR: {:.6f} dB, SSIM: {:.6f}, LPIPS: {:.6f}'.format(
sum(PSNR_all) / len(PSNR_all),
sum(SSIM_all) / len(SSIM_all), avg_lpips[0][0][0][0]))
f.writelines(
'Average: PSNR: {:.6f} dB, SSIM: {:.6f}, LPIPS: {:.6f}'.format(
sum(PSNR_all) / len(PSNR_all),
sum(SSIM_all) / len(SSIM_all), avg_lpips[0][0][0][0]))
# print('{:3d} - {:13}. \tPSNR1: {:.6f} dB, \tSSIM1:{:.6f}, \tPSNR2: {:.6f}, \tSSIM2:{:.6f}'.format(
# i + 1, base_name, PSNR1, SSIM1, PSNR2, SSIM2))
# f.writelines('{:3d} - {:13}. \tPSNR1: {:.6f} dB, \tSSIM1:{:.6f}, \tPSNR2: {:.6f}, \tSSIM2:{:.6f}\n'.format(
# i + 1, base_name, PSNR1, SSIM1, PSNR2, SSIM2))
# PSNR_all_1.append(PSNR1)
# PSNR_all_2.append(PSNR2)
# SSIM_all_1.append(SSIM1)
# SSIM_all_2.append(SSIM2)
# SSIM_all.append(SSIM)
# print('Average: PSNR1: {:.6f} dB, SSIM1: {:.6f}, PSNR2: {:.6f}dB, SSIM2: {:.6f}'.format(
# sum(PSNR_all_1) / len(PSNR_all_1),
# sum(SSIM_all_1) / len(SSIM_all_1),
# sum(PSNR_all_2) / len(PSNR_all_2),
# sum(SSIM_all_2) / len(SSIM_all_2)))
# f.writelines('Average: PSNR1: {:.6f} dB, SSIM1: {:.6f}, PSNR2: {:.6f}dB, SSIM2: {:.6f}'.format(
# sum(PSNR_all_1) / len(PSNR_all_1),
# sum(SSIM_all_1) / len(SSIM_all_1),
# sum(PSNR_all_2) / len(PSNR_all_2),
# sum(SSIM_all_2) / len(SSIM_all_2)))
f.close()
loss_fn = lpips.LPIPS(net='vgg').cuda(
) # closer to "traditional" perceptual loss, when used for optimization
total_lpips = 0
widgets = [
'LPIPS :',
Percentage(), ' ',
Bar('#'), ' ',
Timer(), ' ',
ETA(), ' ',
FileTransferSpeed()
]
pbar = ProgressBar(widgets=widgets)
for file in tqdm.tqdm(files):
# Load images
img0 = lpips.im2tensor(lpips.load_image(os.path.join(
input1, file))) # RGB image from [-1,1]
img1 = lpips.im2tensor(lpips.load_image(os.path.join(input2, file)))
img0 = img0.cuda()
img1 = img1.cuda()
# Compute distance
with torch.no_grad():
dist01 = loss_fn(img0, img1)
total_lpips += dist01
lpips = total_lpips / 5000
print("lpips: ", lpips)
score_lpips = pow(1 - 2 * (min(max(lpips, 0.2), 0.7) - 0.2), 0.5)
if isinstance(score_lpips, torch.Tensor):
score_lpips = score_lpips.item()
os.makedirs(opt.out, exist_ok=True)
data = []
## Initializing the model
loss_fn = lpips.LPIPS(net=opt.net, version=opt.version)
if (opt.use_gpu):
loss_fn.cuda()
smallest_file = opt.path
smallest_dist = 3.0
img0 = lpips.im2tensor(
lpips.load_image(os.path.join(opt.folder, smallest_file)))
files = os.listdir(opt.folder)
while (len(files) > 1):
for (ff, file) in enumerate(files[:-1]):
img1 = lpips.im2tensor(lpips.load_image(os.path.join(
opt.folder, file))) # RGB image from [-1,1]
if (opt.use_gpu):
img0 = img0.cuda()
img1 = img1.cuda()
# Compute distance
dist01 = loss_fn.forward(img0, img1)
# print('%s: %.3f'%(file,dist01))
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--ref_path', type=str, default='./imgs/ex_ref.png')
parser.add_argument('--pred_path', type=str, default='./imgs/ex_p1.png')
parser.add_argument('--use_gpu',
action='store_true',
help='turn on flag to use GPU')
opt = parser.parse_args()
loss_fn = lpips.LPIPS(net='vgg')
if (opt.use_gpu):
loss_fn.cuda()
ref = lpips.im2tensor(lpips.load_image(opt.ref_path))
pred = Variable(lpips.im2tensor(lpips.load_image(opt.pred_path)),
requires_grad=True)
if (opt.use_gpu):
with torch.no_grad():
ref = ref.cuda()
pred = pred.cuda()
optimizer = torch.optim.Adam([
pred,
], lr=1e-3, betas=(0.9, 0.999))
plt.ion()
fig = plt.figure(1)
ax = fig.add_subplot(131)
ax.imshow(lpips.tensor2im(ref))
import argparse
import lpips
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-p0','--path0', type=str, default='./imgs/ex_ref.png')
parser.add_argument('-p1','--path1', type=str, default='./imgs/ex_p0.png')
parser.add_argument('-v','--version', type=str, default='0.1')
parser.add_argument('--use_gpu', action='store_true', help='turn on flag to use GPU')
opt = parser.parse_args()
## Initializing the model
loss_fn = lpips.LPIPS(net='alex',version=opt.version)
if(opt.use_gpu):
loss_fn.cuda()
# Load images
img0 = lpips.im2tensor(lpips.load_image(opt.path0)) # RGB image from [-1,1]
img1 = lpips.im2tensor(lpips.load_image(opt.path1))
if(opt.use_gpu):
img0 = img0.cuda()
img1 = img1.cuda()
# Compute distance
dist01 = loss_fn.forward(img0,img1)
print('Distance: %.3f'%dist01)