Qiyao Wei

Results 6 comments of Qiyao Wei

Hey Chris! Thanks a lot for the reply! As a follow-up to this question, I have tried directly copying and pasting the code in the last section "Plot Recipe" from...

1. hmmm, is there currently an alternative to BatchNorm? I guess it would be safest to just stick to linear and conv layers + activations, although the accuracy will for...

Oh I thought backpack [doesn't support](https://docs.backpack.pt/en/master/supported-layers.html) GroupNorm BTW I might have figured out the issue, it goes away when I do add an eval like: extend(model).eval(). Not sure why but...

![sp_swin](https://user-images.githubusercontent.com/36057290/211655647-31f00bc9-165b-4ef4-b628-911fd9ff4e3c.png) ![mup_swin](https://user-images.githubusercontent.com/36057290/211655686-da4037bb-2927-4958-b99d-a5af85e7d434.png) @shiyf129 I also think the snippets look reasonable. I have done coord checks on Swin as well, and I attach the plots here. Echoing Edward's suggestion, the widths...

@casper-hansen Firstly, many thanks for your brilliant work! I am very new to this repo, but would love to help out! As far as I can tell, completing this issue...

This has been discussed in [multiple](https://github.com/huggingface/trl/pull/1265) [github](https://github.com/h2oai/h2o-llmstudio/issues/580) [issues](https://github.com/huggingface/trl/issues/1294), and I believe the answer stems from the discussion that Huggingface had with the IPO authors [here](https://huggingface.co/blog/pref-tuning) "After consulting with the authors...