Bowen Zhang

Hi, it seems that you are trying to pre-compile the CUDA ops? The CUDA ops in this repo don't need pre-compile and you could try to run the code directly....

What's the CUDA toolkit version you used?

Hi, maybe you could try the solution of [this link](https://discuss.pytorch.org/t/how-to-run-detectron2-using-rtx-3090-and-with-cuda-11-x-on-ubuntu/119448/9) to check whether that helps.

Sure! I have sent my request to join the CVPR org in huggingface, thanks for the reminder.

Hi, although we did not provide the modified model for img2img task, I think this is theoretically possible. A simple idea is just replacing the input to AdaIN with image...

> You supply the path to the checkpoint via `--ckpt`, ie. `--ckpt samples/000001.pt`. Also don't forget to supply original hyperparams, like image size, style dim, if we use SN etc......

Hi, please make sure your sample_path exist without `/samples` added. Thanks.

This augment is used to calculate fid score between the generated image samples and real image samples, which is used in [evaluation](https://github.com/microsoft/StyleSwin/blob/main/train_styleswin.py#L271). If you want to calculate fid score, please...

Hi, just modify the parameters in [torchvision.utils.save_image](https://github.com/microsoft/StyleSwin/blob/main/train_styleswin.py#L286) to generate images in grid view. Thanks.

Hi, we haven't met this in the final results. We have tRGB modules in the model, and please don't remove them. Also please check the training iterations are the same...