Jun-Yan Zhu

icon indicating a website link https://www.cs.cmu.edu/~junyanz/

icon company Carnegie Mellon University: icon location Pittsburgh, PA icon location Understanding and creating pixels.

Results 321 comments of Jun-Yan Zhu

Unet generator for large images

I hypothesize that the ResNet has fewer parameters and fewer downsampling, which are both good for color and style transfer, while U-Net has many more parameters (hard to learn these...

Differentiable data augmentation

Great. Does it help in your experiments? Maybe @zsyzzsoft can help create a PR or a branch for all the models. We are doing some experiments with conditional GANs and...

GPU not using more than 2

For crop_size=512, batch_size=2 will require a significant amount of GPU memory. I am not sure how much memory you have per GPU. I am not sure if you need to...

How to eliminate strange border effect in the CycleGAN output image

Maybe you could use some data augmentation. Currently, both your load_size and crop_size are 512. Maybe use load_size 580 and crop_size 512. Certain artifacts will go away if the model...

Loss D goes to zero when adding a new loss to generator

Maybe your newly added loss is too strong compared to existing CycleGAN losses.

Loss D goes to zero when adding a new loss to generator

Maybe reducing the weight for your new loss.

Loss D goes to zero when adding a new loss to generator

maybe making the discriminator weaker or the generator stronger.

Loss D goes to zero when adding a new loss to generator

You can set different learning rates for G and D. See this [paper](https://arxiv.org/abs/1706.08500) for more details.

How to evaluate the Classification Performance on Cityscapes?

Please see these two discussions [1](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/issues/150) and [2](https://github.com/phillipi/pix2pix/issues/116) @tinghuiz

How to evaluate the Classification Performance on Cityscapes?

I think we just assigned it to its nearest neighbor in color space. @tinghuiz @phillipi