I recently came across Stochastic Weight Averaging in Pytorch. I found some codes which use it for Image classification and NLP tasks. But are there some example codes which use it for CV tasks involving image generation such as Image Segmentation, Denoising, Object Detection, Super Resolution or deblurring?
I would expect the usage would be the same for the mentioned use cases or are you concerned about a specific issue which would need some adoption in the SWA usage?
Thanks for reply.
Not an adoption request. Just that using SWA I am getting 2dB inferior results, so wanted to go through PyTorch code of someone who has gotten good results with SWA. So far I see SWA used in image classification and NLP tasks only. Just wondering it works well for CV tasks like denoising, segmentation, super resolution, etc.
One last thing, sadly PyTorch does not allow SWA to be used with inbuilt Weight Normalization found in nn.utils. But i am sure this wont be the reason for 2dB lower PSNR.
Ah OK, thanks for clarifying. I haven’t used it in these setup, so let’s wait for some experts who might have experience with it.