ResNeXt Much Slower than ResNet in TF?

[xb-chang]

Hi,

I tried to implement ResNeXt using slim structure.

I implement it base on modifying the ResNet code.

I follow the Fig.3(b) structure in ResNeXt paper. Since ResNeXt contains parallel sub-networks in each block, I use for loop to build them.

But the ResNeXt network building and training require much more time than ResNet.
Why?

I am new to TF, but I guess maybe I am not using the right methods to implement such parallel sub-networks structures. And that's why it is slow

Is there any better way to manage such parallel sub-networks in TF?
I mean, for example, the network architecture of Fig.3(a), any efficient way to implement it directly? (not using equvalence Fig.3(b) and Fig.3(c)) (any concrete example to learn from?)

Thanks a lot.