Add mobilenetv4 support and pretrained models?

[LiYufengzz]

🚀 The feature

Google has published the mobilenetv4 model. When will pytorch support it and open the pre-trained model?

Motivation, pitch

I very much hope to use the latest lightweight backbone

Alternatives

No response

Additional context

No response

[ChengpengChen]

Hi, @LiYufengzz

I've implemented MobileNetV4 in pytorch, can refer to this. I am trying to reproduce the results though.
https://github.com/ChengpengChen/RepGhostNetV2

The code also provides a comparable backbone, i.e., RepGhostNetV2. Feel free to reach me if you have any concerns.

Thanks

[LiYufengzz]

Hi, @LiYufengzz

I've implemented MobileNetV4 in pytorch, can refer to this. I am trying to reproduce the results though. https://github.com/ChengpengChen/RepGhostNetV2

The code also provides a comparable backbone, i.e., RepGhostNetV2. Feel free to reach me if you have any concerns.

Thanks

Thanks a lot for the reply，I hope you can reproduce the indicators in the paper.

[NicolasHug]

Thanks for the feature request @LiYufengzz . We're not planning to add mobiletnetv4 to torchvision so we recommend using third-party implementation like that of @LiYufengzz

[rwightman]

timm has an (almost) full set of MobileNetV4 weights now https://github.com/huggingface/pytorch-image-models?tab=readme-ov-file#june-12-2024

Hybrid-large and higher res hybrid-medium still in training.

[LiYufengzz]

timm has an (almost) full set of MobileNetV4 weights now https://github.com/huggingface/pytorch-image-models?tab=readme-ov-file#june-12-2024

Hybrid-large and higher res hybrid-medium still in training.

thanks a lot