Jacky Lee

> 1000 lines are cool, but what is not cool - it's having less than 1000 lines at the cost of not having docstrings and readable code. I agree that...

Thanks, I've added a test case for it (very similar to `test_conv2d`). Interestingly, the parameters are reversed in PyTorch, so it looks something like this: ``` class GroupNorm: def __init__(self,...

> Want to flip it to match PyTorch? Honestly, I think our implementation is more intuitive. Take a look at [this example](https://pytorch.org/docs/stable/generated/torch.nn.GroupNorm.html): ``` >>> input = torch.randn(20, 6, 10, 10)...

Understood. I'll flip the parameters to match PyTorch.

Updated the parameters. I removed the default value for `num_groups` to match PyTorch as well, but had to hard code the parameters in the stable diffusion blocks.

> "elementwise_affine on channels" HACK needs to be fixed too, should match the PyTorch behavior in all cases. Sounds good, I'll try and fix it

Hi George, thanks for the shout out on stream! Looks like PyTorch doesn't support `num_groups=None`. It must be an integer divisible by `num_channels`. Below is the error message I got...

Sounds good. I think I can try and reshape the inputs in stable diffusion. Usually when we call GroupNorm, we expect x to have shape [N, C, H, W]. We...

I'd like to work on vivit

@NielsRogge @nisyad-ms I managed to show the preprocessed image with the correct boxes. Below is the full code. ``` import torch import requests import numpy as np from PIL import...