Jun-Yan Zhu

We calculate GAN training loss in this [line](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/cycle_gan_model.py#L169).

I recommend that you read the [DCGAN](https://github.com/pytorch/examples/blob/master/dcgan/main.py) code and the original GAN paper. The objective for generator and discriminator is slightly different, and our code reflects the difference.

Your loss plots look normal. Usually, cycle-consistency loss and identity loss decrease during training, while GAN losses oscillate. To evaluate the quality or detect overfitting/underfitting. you need to apply additional...

This seems to be a cuda-related issue. Are you able to run basic PyTorch command on GPU?

1. I fixed the issue with the latest commit. 2. Yeah, add your images to `train` and `test` directory. I also added a new script `scripts/test_colorization.sh`. 3. I am not...

Loss function can be used to identify failure mode as suggested by [ganhacks](https://github.com/soumith/ganhacks#10-track-failures-early). For example, If D loss is always 0, maybe D is too strong. You can increase the...

1. For your own task, you probably need more images. 25 might not be enough. You probably need thousands or even tens of thousands of training images. You may also...

1. Colorization might be different from artist style transfer. From my experience, 25 images might not be enough. 2. I don't have a good answer. To prevent overfitting, you may...

The model is trained on mostly sedans (certainly not trucks). Training models on larger datasets or augmented data might make it work better for your input.

You need to edit the csv file and run the "test.sh" script. I will add the paper and close the issue.