https://github.com/microsoft/Swin-Transformer

[lucasjinreal]

How about try this tech?

[glenn-jocher]

"The code will be coming soon." they say.

[lucasjinreal]

Oh. yeah

[glenn-jocher]

@jinfagang good find though, their reported mAP is super high, so I'd be interested to see the code if/once it's released.

[zhangming8]

@jinfagang good find though, their reported mAP is super high, so I'd be interested to see the code if/once it's released.

Hi, I find you have pushed "TransformerLayer". What's the result of it ?

[glenn-jocher]

@zhangming8 see https://github.com/ultralytics/yolov5/issues/2329 for experiments with Transformer layers. They seem to show promise, though they require more resources to train, especially CUDA memory.

[lucasjinreal]

@glenn-jocher Looked into the thread seems transformer is also appliceable. Does the speed get faster as well?

[sbbug]

GOOd

[jayyoung23]

Here it comes！ @glenn-jocher https://github.com/microsoft/Swin-Transformer https://github.com/SwinTransformer/Swin-Transformer-Object-Detection

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Bobo-y]

I've used swin transformer as yolov5 backbone in my repo.

[glenn-jocher]

@yl305237731 that's very impressive! Did you see performance improvements with the swin transformers backbone?

[Bobo-y]

@glenn-jocher I just trained on local dataset for person detection, only trained about 10 epochs. and the recall and accuracy were almost the same as using the original yolov5, so I turned it off. 😂