Jun-Yan Zhu
Jun-Yan Zhu
Yeah, PatchGAN will produce an output map given a whole image. You do not need to produce patches explicitly.
Not sure. You can also do crop_size 360 or something smaller.
You don't need to use activation functions for least square GANs. The network will learn to regress to 0 and 1 (otherwise it will pay a penalty). You only need...
It is hard to tell without knowing your data and task. Feel free to send an email to Taesung and me.
It's hard to know why. Could you give us more debugging information? @SsnL
We don't have `num_samples` in this repo. Would it be possible to share with us the trace of the program?
You need to set the data path correctly.
We have updated the link to fcn-8s-cityscapes
We did provide an [environment.yml](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/environment.yml), which should work.
It looks like that the generator is hiding information for the reconstruction purpose. See this [study](https://arxiv.org/abs/1712.02950) for more details. It is hard to know what's going on without seeing the...