alibaba-mmai-research/TAdaConv

I am curious about why the flops and parms of TAdaConvNeXt-T is one half of the ResNet50's, to my knowledge they should be similar.

leexinhao opened this issue · 2 comments

leexinhao commented 
          I am curious about why the flops and parms of TAdaConvNeXt-T is one half of the ResNet50's, to my knowledge they should be similar.

image

Originally posted by @leexinhao in #11 (comment)

Could you give me the implement of TAdaConvNeXt？

huang-ziyuan commented

Sorry for the delayed reply.

The reason for the reduced computation for TAdaConvNeXt is that we used a tublet embedding stem in TAdaConvNeXt model, which reduces the temporal resolution by a factor of 2 in the first downsample layer of ConvNeXt :). In contrast, the TAda2D stem employs only spatial convolutions with no temporal downsampling.

leexinhao commented

Sorry for the delayed reply.

The reason for the reduced computation for TAdaConvNeXt is that we used a tublet embedding stem in TAdaConvNeXt model, which reduces the temporal resolution by a factor of 2 in the first downsample layer of ConvNeXt :). In contrast, the TAda2D stem employs only spatial convolutions with no temporal downsampling.

Thanks for your reply!