Error while trying to train ESRGAN about mmagic HOT 4 CLOSED

I am also running into this issue. Any help from the devs or people who have successfully trained on their own data sets would be much appreciated!

from mmagic.

@ienko I have found a solution. The issue was in codes/data/LQGT_dataset.py in the random crop step (around line 110). I rewrote the random crop algorithm like this:

        gt_h = random.randint(0, max(0, gH - GT_size))
        gt_w = random.randint(0, max(0, gW - GT_size))
        img_GT = img_GT[gt_h:gt_h + GT_size, gt_w:gt_w + GT_size, :]

        LQ_size = GT_size // scale
        lq_h = gt_h // scale
        lq_w = gt_w // scale
        img_LQ = img_LQ[lq_h:lq_h + LQ_size, lq_w:lq_w + LQ_size, :]

and that fixed the error

from mmagic.

@PSZehnder great, thanks a lot.

I also managed to solve it in a similar way, with writing another code for preprocessing the dataset I had (needed 4:3 with specific resolutions).

    imgScale = H/height
    newX, newY = oriimg.shape[1]*imgScale, oriimg.shape[0]*imgScale
    newimg = cv2.resize(oriimg, (int(newX), int(H)))
    
    newWidth = int(W)
    newHeight = int(H)
    xBegin = (int(newX) - int(newWidth))/2
    xEnd = int(xBegin) + int(newWidth)
    cropped = newimg[0:1920, int(xBegin):int(xEnd)]

For some reason, it still doesn't work for all the batch sizes, but guess that's implied by network architecture.

from mmagic.

The model arch is such that it requires: (128,128,3) as GT (X) and (32,32,3) as LR (Y).

Keeping these two dimensions in mind, I was able to create my own dataset, then create it's corresponding lmdb files using the ./codes/data_scripts/create_lmdb.py and also validate the lmdb files created using the ./code/data_scrpits/test_dataloader.py file

Make sure that you have the correct paths present in the ./option/train/train_model.yaml file before starting the training. I hope if you adhere to these images dimensions you'll be able to fit a decent batch size depending upon your GPU memory available and start the training easily on the custom dataset.

from mmagic.