sutd-visual-computing-group/dag-gans

Missing transform_ image function

Closed this issue · 6 comments

SJLeo commented

There's no transform_image function in utils.py when I'm training cyclegan

Hi, thanks for the feedback. I've updated the util.py for cyclegan. If it's ok to you, please close the issue. Thanks.

SJLeo commented

Thank you, but I'm a little puzzled if only rotation or flipping is used on cyclegan, and other data augument is not effective.

Hi, we've only tested with rotation and flipping for cyclegan so far, but I believe that other data augmentations like translation, cropping, ... or their combinations are effective too as we did with other GANs in our paper.

SJLeo commented

It seems that we can't use multiple data augument methods at the same time.

For current implementation for cyclegan, we test only with a single augmentation for the demonstration with image-image translation, will update multiple methods in the near future. If you prefer to use multiple augmentations, may refer our published pytorch module.

SJLeo commented

Thanks!