Bad artifacts in output image~~~~

[KindleHe]

1. The below image coms from "out" directory after running CUDA_VISIBLE_DEVICES=1 python test.py --cuda --model=model/model_div2k.pth.
2. The weight model_div2k.pth you give can result the PSNR result as you report in readme.txt, however the output images seems not as good as your PSNR result.

[opteroncx]

To solve this bug, simply edit line147-149

1. remove code like 'np.uint8()'
2. convert to bgr and use opencv for output

[KindleHe]

@opteroncx I am trying it. Thanks for your tips. Could you share more information about your training command?

[opteroncx]

Use div2k, Random rotate or flip the input matrix in dataloader (augmentation) train with 32x32 input for 200 epochs, set lr=1e-4 and decrease 10 times after 100 epochs. then use 48x48 input to finetuning, this will gain about 0.3dB

[opteroncx]

for example: def data_augment(im,num): org_image = im.transpose(1,2,0) if num ==0:
ud_image = np.flipud(org_image) tranform = ud_image elif num ==1:
lr_image = np.fliplr(org_image) tranform = lr_image elif num ==2: lr_image = np.fliplr(org_image) lrud_image = np.flipud(lr_image)
tranform = lrud_image elif num ==3: rotated_image1 = np.rot90(org_image)
tranform = rotated_image1 elif num ==4: rotated_image2 = np.rot90(org_image, -1) tranform = rotated_image2 elif num ==5: rotated_image1 = np.rot90(org_image) ud_image1 = np.flipud(rotated_image1) tranform = ud_image1 elif num ==6:
rotated_image2 = np.rot90(org_image, -1) ud_image2 = np.flipud(rotated_image2) tranform = ud_image2 else: tranform = org_image tranform = tranform.transpose(2,0,1) return tranform

[KindleHe]

@opteroncx thanks for your tips very on image displaying. I try to replace the code in line147-149

img = np.uint8(img)
im = Image.fromarray(img)
im.save(out_path+name[:-4]+'.png')

with the following code:

img_norm = cv2.normalize(img, img, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype= cv2.CV_8U)
img_bgr = cv2.cvtColor(img_norm, cv2.COLOR_RGB2BGR)
cv2.imwrite(out_path+name[:-4]+'.png', img_bgr)

Thus, the picture result will be save correctly without error pixels

[KindleHe]

@opteroncx

Use div2k,
Random rotate or flip the input matrix in dataloader (augmentation)
train with 32x32 input for 200 epochs, set lr=1e-4 and decrease 10 times after 100 epochs.
then use 48x48 input to finetuning, this will gain about 0.3dB

Is that mean the steps I should obey is the following steps ? If something wrong, please point out, thanks!

1. feed DIV2K_train_HR into LapSRN' s generate_train_lap_pry.m script setting its stride = 64 to generate 32x32 input image patches
2. call def data_augment(im,num) function after training_data_loader = DataLoader()
3. use command python main.py --cuda --batchSize=64 --nEpochs=200 --lr=1e-4 --step=100 to execute the first round of traing
4. Then repeat step 1 with stride = 96 to generate 48x48 input image patches

I think your suggestions will work as you say, but what is the finetuning setting of next step after the first round of training?

[KindleHe]

@opteroncx Hi?

[opteroncx]

Decrease the learning rate to 1e-5, and step set to 10