Daniel Bolya

As stated in the paper:  where [19] is just the pytorch-fid package. For the "5000 class balanced samples", I just took the ImageNet val set, and used the first...

I believe Imagen uses just convnets for its unet, not a transformer like stable diffusion does. So in that respect, it can't be used like I use it for stable...

> and when I set ratio=0.6, the cost time and GPU memory are greater than ratio=0.5 That doesn't seem right. What environment are you in and how are you benchmarking...

I'm not familiar with ONNX. What are the main differences with regular pytorch? If it's easy to do, I can write a patch.

Thank you for the suggestion! I never thought about applying different ratios to the different modules. I wonder though, do you get any speed benefit from doing this? In my...

I see, thanks for all the testing! I'm totally on board with adding finer control, but the issue is the interface. Currently, there are a lot of variables even without...

Hi, do you mean that the content of the image is the same but you get lower quality results, or do you mean that there are noticeable artifacts that ruin...

Seems to be an issue with the requirements of gather, similar to M1 Macs (#4). Are you able to get any more information than just "RuntimeError"? It would be useful...

Ah, nvm I found it: https://learn.microsoft.com/en-us/windows/win32/api/directml/ns-directml-dml_gather_operator_desc It is indeed the same issue: > IndicesTensor > Type: const [DML_TENSOR_DESC](https://learn.microsoft.com/en-us/windows/win32/api/directml/ns-directml-dml_tensor_desc) > A tensor containing the indices. The DimensionCount of this tensor **must...