Why the Gated CNN Blocks are not 24 layers？

[David-Ttao]

I think its necessary to set 24 layers of MambaOut in memory of Kobe Bryant.

[Celestial-Bai]

They can also set 8 layers, but they did not either. Man!

[David-Ttao]

They can also set 8 layers, but they did not either. Man!

its worth discussing and i think its necessary to reproduce the code and change the layers to test result.

[pvbvcv]

It will be a meaningful work!

[ChaohuanDeng123]

就你这个issue显得格格不入。伟大无需多言！

[LightwishWONG]

《changed the title 什么罐头我说？ Why the Gated CNN Blocks are not 24 layers？》hhhhhh

[yuweihao]

Thank you so much for your suggestion. We released MambaOut-Kobe model, a Kobe Memorial version with 24 Gated CNN blocks. MambaOut-Kobe achieves really competitive performance, surpassing ResNet-50 and ViT-S with much fewer parameters and FLOPs. For example, MambaOut-Kobe outperforms ViT-S by 0.2% accuracy with only 41% parameters and 33% FLOPs.

Model	Resolution	Params	MACs	Top1 Acc
ResNet-50 (ResNet strikes back)	224	25.5M	4.1G	79.8
ViT-S	224	22.1M	4.6G	79.8
MambaOut-Kobe	224	9.1M	1.5G	80.0

[Celestial-Bai]

Thank you so much for your suggestion. We released MambaOut-Kobe model, a Kobe Memorial version with 24 Gated CNN blocks. MambaOut-Kobe achieves really competitive performance, surpassing ResNet-50 and ViT-S with much fewer parameters and FLOPs. For example, MambaOut-Kobe outperforms ResNet-50 by 0.2% accuracy with only 36% parameters and MACs.

Model Resolution Params MACs Top1 Acc
ResNet-50 224 25.5M 4.1G 79.8*
ViT-S 224 22.1M 4.6G 79.8
MambaOut-Kobe 224 9.1M 1.5G 80.0

• The result is cited from "ResNet strikes back" paper, a very strong version of ResNet trained for 300 epochs.

Man! Hahahaha

[David-Ttao]

Thank you so much for your suggestion. We released MambaOut-Kobe model, a Kobe Memorial version with 24 Gated CNN blocks. MambaOut-Kobe achieves really competitive performance, surpassing ResNet-50 and ViT-S with much fewer parameters and FLOPs. For example, MambaOut-Kobe outperforms ResNet-50 by 0.2% accuracy with only 36% parameters and MACs.

Model Resolution Params MACs Top1 Acc
ResNet-50 224 25.5M 4.1G 79.8*
ViT-S 224 22.1M 4.6G 79.8
MambaOut-Kobe 224 9.1M 1.5G 80.0

• The result is cited from "ResNet strikes back" paper, a very strong version of ResNet trained for 300 epochs.

its a meaningful work, what can i say?

[WindZh03]

What a great suggestion!