search

alibaba / lightweight-neural-architecture-search

This is a collection of our zero-cost NAS and efficient vision applications.
Apache License 2.0
383 stars 50 forks source link

the limit of btn #14

Open ct-wei opened 1 year ago

ct-wei commented 1 year ago

The constraint limit the btn in https://github.com/alibaba/lightweight-neural-architecture-search/blob/6bf4d6949ed690b8ef59bcb843e2d36d03ebecd1/tinynas/spaces/mutator/super_res_k1kx_mutator.py#L72 https://github.com/alibaba/lightweight-neural-architecture-search/blob/6bf4d6949ed690b8ef59bcb843e2d36d03ebecd1/tinynas/spaces/mutator/super_res_k1kx_mutator.py#L82 but this https://github.com/tinyvision/DAMO-YOLO/blob/master/damo/base_models/backbones/nas_backbones/tinynas_nano_middle.txt seems to have broken the limit May I ask why? thanks

xiuyu-sxy commented 1 year ago

The constraint limit the btn in

https://github.com/alibaba/lightweight-neural-architecture-search/blob/6bf4d6949ed690b8ef59bcb843e2d36d03ebecd1/tinynas/spaces/mutator/super_res_k1kx_mutator.py#L72

https://github.com/alibaba/lightweight-neural-architecture-search/blob/6bf4d6949ed690b8ef59bcb843e2d36d03ebecd1/tinynas/spaces/mutator/super_res_k1kx_mutator.py#L82

but this https://github.com/tinyvision/DAMO-YOLO/blob/master/damo/base_models/backbones/nas_backbones/tinynas_nano_middle.txt seems to have broken the limit May I ask why? thanks

That's a good question. For small or mobile models, particularly those with depthwise search space, we have observed that the limitation on btn may constrain the final accuracy. The similar observation can be found in mobilenetv3 paper.