how to train on ADE20K?

[wensihan]

what should I do if I want to train the network on the ADE20K? Do I need to write a ADE20K_train.h5 file like the cityscapes_train.h5? And the rescale script like the resize_cityscapes.sh? Looking forward to your response~

[Justin-Tan]

To train on ADE20K the authors resize the width of each image to 512 px. Then specify the -ds ADE20k flag while training. The reason for this is that the encoder reduces the image width and height by a factor of 16, so the dimensions of the input image need to be cropped to the nearest multiple of 16.

[wensihan]

The code you release didn't provide the part of resizing the width, did you?

[wensihan]

And, if I want to use the conditional gan, the original image and semantic image both need to be resized?

[Justin-Tan]

Yes, but resizing is trivial using imagemagick:

for f in ./**/*.png; do
    convert $f -resize 512x $f
done

[wensihan]

Okay, Thank you very much for your reply~