def LPIPS_cal(dataset_teacher, number_classes, image_dir, output_file, device):
print("Calculate LPIPS value...")
all_pairs_lpips = False
model = perceptual_models.PerceptualLoss(model='net-lin', net='alex', use_gpu=use_gpu, version=0.1)
dists_among_fake_imgs = []
for i in range(number_classes):
files = os.listdir(image_dir + dataset_teacher + '/fake_images/' + str(i))
num = len(files)-1
if num > 10:
num = 10
for (ff, file0) in enumerate(files[0:num]): # [:-1]):
img0 = im2tensor(
load_image(os.path.join(image_dir + dataset_teacher + '/fake_images/' + str(i), file0))).to(device) # RGB image from [-1,1]
if (all_pairs_lpips):
files1 = files[ff + 1:]
else:
files1 = [files[ff + 1], ]
for file1 in files1[0:num]: # :
img1 = im2tensor(
load_image(os.path.join(image_dir + dataset_teacher + '/fake_images/' + str(i), file1))).to(
device) # RGB image from [-1,1]
# Compute distance
dist01 = model.forward(img0, img1)
dists_among_fake_imgs.append(dist01.item())
avg_dist_among_fake_imgs = numpy.mean(numpy.array(dists_among_fake_imgs))
stderr_dist_among_fake_imgs = numpy.std(numpy.array(dists_among_fake_imgs))/numpy.sqrt(len(dists_among_fake_imgs))
print("[LPIPS_mean: %s] [LPIPS_std: %s] " % (
avg_dist_among_fake_imgs, stderr_dist_among_fake_imgs))
with open(output_file, 'a') as f:
f.write(str(avg_dist_among_fake_imgs) + '\n')
f.close()
fn.split('/')[-1]))
# PSNR
img_gt = (val_H * 255).astype(np.uint8)
img_target = ((batch_Out) * 255).astype(np.uint8)
CROP_S = 16
if CROP_S > 0:
img_gt = img_gt[CROP_S:-CROP_S, CROP_S:-CROP_S]
img_target = img_target[CROP_S:-CROP_S, CROP_S:-CROP_S]
psnrs.append(
PSNR(
_rgb2ycbcr(img_gt)[:, :, 0],
_rgb2ycbcr(img_target)[:, :, 0], 0))
# LPIPS
dist = model_LPIPS.forward(
im2tensor(img_target),
im2tensor(img_gt)) # RGB image from [0,255] to [-1,1]
lpips.append(dist)
print('AVG PSNR/LPIPS: Validation: {}/{}'.format(
np.mean(np.asarray(psnrs)), np.mean(np.asarray(lpips))))
# Save best model
if i % I_SAVE == 0:
if np.mean(np.asarray(lpips)) < best_avg_lpips:
best_avg_lpips = np.mean(np.asarray(lpips))
SaveCheckpoint(i, best=True)
testset, out_fn))
# Evaluation
CROP_S = 4
# PSNR
psnrs.append(PSNR(img_gt_y, batch_out_y, CROP_S))
# SSIM
ssims.append(
skimage.metrics.structural_similarity(
img_gt_y,
batch_out_y,
data_range=255,
multichannel=False,
gaussian_weights=True,
sigma=1.5,
use_sample_covariance=False))
# LPIPS
img_gt_t = im2tensor(img_gt)
batch_out_t = im2tensor(batch_out)
if CROP_S > 0:
batch_out_t = batch_out_t[:, :, CROP_S:-CROP_S, CROP_S:-CROP_S]
img_gt_t = img_gt_t[:, :, CROP_S:-CROP_S, CROP_S:-CROP_S]
lpips.append(model_LPIPS.forward(batch_out_t, img_gt_t))
print('{} AVG PSNR/SSIM/LPIPS: {}/{}/{}'.format(
testset, np.mean(np.asarray(psnrs)), np.mean(np.asarray(ssims)),
np.mean(np.asarray(lpips))))
img_gt = np.around(img_H * 255).astype(np.uint8)
img_out = np.around(batch_output * 255).astype(np.uint8)
Image.fromarray(img_out).save('{}/result/{}/{}/{}.png'.format(
EXP_NAME, str(VERSION), valset_name,
fn.split('/')[-1]))
# PSNR
CROP_S = 4
psnrs.append(
PSNR(
_rgb2ycbcr(img_gt)[:, :, 0],
_rgb2ycbcr(img_out)[:, :, 0], CROP_S))
# LPIPS
img_gt = im2tensor(img_gt) # RGB image from [-1,1]
img_out = im2tensor(img_out)
if CROP_S > 0:
img_out = img_out[:, :, CROP_S:-CROP_S, CROP_S:-CROP_S]
img_gt = img_gt[:, :, CROP_S:-CROP_S, CROP_S:-CROP_S]
dist = model_LPIPS.forward(img_out, img_gt)
lpips.append(dist)
mean_psnr = np.mean(np.asarray(psnrs))
mean_lpips = np.mean(np.asarray(lpips))
print('AVG PSNR/LPIPS: {}: {}/{}'.format(valset_name, mean_psnr,
mean_lpips))
writer.add_scalar(valset_name, mean_psnr, i)
writer.add_scalar(valset_name + '_lpips', mean_lpips, i)