Support for the Googles' Big Transfer (BiT) model

[coder-magpie]

Checklist

• I have searched related issues but cannot get the expected help.
• I have read related documents and don't know what to do.

Describe the question you meet

Hi mmcls team,

I wanted to know whether we can support the implementation of Googles' Big Transfer (BiT): General Visual Representation Learning model with the current version of mmcls or if you have any plans for the future. I tried both adding this model as a new backbone and using Resnet with pre-trained, but both approaches didn't work. I would appreciate it if someone could provide any suggestions or solutions for implementation.

Google BiT Model - https://github.com/google-research/big_transfer

Thank you,

[Ezra-Yu]

Thanks for your suggestions. We do have a plan to include this model.

It is not feasible to load ResNet parameters into Bit, because their network structures are completely different. You can directly use the weight of the official repo, you need to change the name of the key in their checkpoint. refer to some tools.

It seems that you have tried adding this model as a new backbone. Can you release your code or just initiate a pull and request? we can review it and find the problem.

[coder-magpie]

Thank you for your prompt and detailed response @Ezra-Yu . I have the code on a non-public repository. I will create a branch and push the code, for your review. Closing this issue. Thanks

[ira7bar]

Hi! Was this feature indeed pushed? I didn't find BiT in the repo. thx