Pytorch-cyclegan-and-pix2pix: Image with numbers will be reverse after training

[bill0812]

I used original captcha as our dataset "A" and another style of captcha as dataset "B" to put in model and train, trying to transfer from "A to B" or "B to A".
However, I found there is a problem. I check for intermediate results, and the number in the captcha was reversed. But what I input was an "unreversed" image, besides, there are no problems when I train for other images without numbers, such as zebra2horse or photo2monet ...etc.
Btw, all parameters were set as default.
Wondering what reason casued this problem ?

[davidwessman]

This might be because the framework uses Data augmentation by default, input the option --no_flip.
https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/options/base_options.py#L43

[bill0812]

Thank u so much. I didn’t notice there’s a so important parameter to me ! 😅

[davidwessman]

Thank u so much. I didn’t notice there’s a so important parameter to me ! 😅

We got hit by it too! @bill0812

Btw: Close the issue if it helped!