About the method

[mingliangzhang2018]

Thank you for sharing your interesting work. Would you mind clarify that what is the method of "CE(Balanced)" ?

[YyzHarry]

Hi, thanks for your interest! "CE(Balanced)" means CE with class-balanced sampling.

[mingliangzhang2018]

Hi, thanks for your interest! "CE(Balanced)" means CE with class-balanced sampling.

Thanks!

[mingliangzhang2018]

I still have some questions. At first, why not apply models of ResNeXt same as the paper "DECOUPLING"? Secondly, I find that you use small batch size (100 or 128) in the models training with SSP for the datasets of Imagenet-LT and iNat2018. Do you test the performance of all baseline methods without SSP using same batch size? In other words, whether or not the performance improvement mainly comes from the adjustment of the batch size but less from SSP?

[YyzHarry]

• ResNeXt architecture: Simply because we have limited computation resources. As we are comparing to many methods rather than only decoupling method, we just choose 1-2 representative architectures. I believe similar results should hold for different architectures.
• Batch size: For ImageNet-LT I'm using 128 for all models (and 100 for iNat), as these are the maximum size my GPU can hold. Note that both the baselines and models with SSP are using the same training setups, despite the latter one loading a pre-trained model.

[mingliangzhang2018]

• ResNeXt architecture: Simply because we have limited computation resources. As we are comparing to many methods rather than only decoupling method, we just choose 1-2 representative architectures. I believe similar results should hold for different architectures.
• Batch size: For ImageNet-LT I'm using 128 for all models (and 100 for iNat), as these are the maximum size my GPU can hold. Note that both the baselines and models with SSP are using the same training setups, despite the latter one loading a pre-trained model.

thank you very much!