Do you try larger resolution when evaluation or pretrained on ImageNet-22k to get higher performance?

SHI-Labs / Neighborhood-Attention-Transformer

Neighborhood Attention Transformer, arxiv 2022 / CVPR 2023. Dilated Neighborhood Attention Transformer, arxiv 2022

MIT License

1.04k stars 85 forks source link

Do you try larger resolution when evaluation or pretrained on ImageNet-22k to get higher performance? #39

Closed laisimiao closed 1 year ago

laisimiao commented 2 years ago

I think it would be more convinced about your model capacity if you provide the mentioned results.

alihassanijr commented 2 years ago

As mentioned in the paper, we are currently looking into 22K pretraining. We were facing several issues as Swin did not release their 22k training settings in full until 2 weeks ago, and even now it is still incomplete for reproduction, and training on 22K is time consuming. We plan to release our 22k checkpoints and results when ready, so stay tuned.

laisimiao commented 2 years ago

Look forward. Best regards.

alihassanijr commented 1 year ago

Regarding this issue, we now have new models which we do train on ImageNet-22K and fine-tune on ImageNet-1K at both 224x224 and 384x384 resolutions. I'm closing this issue now, but feel free to reopen it if you have further questions.