search

Deci-AI / super-gradients

Easily train or fine-tune SOTA computer vision models with one open source training library. The home of Yolo-NAS.
https://www.supergradients.com
Apache License 2.0
4.58k stars 506 forks source link

There seems to be a problem with the MobilNetV3 structure #365

Closed Daming-TF closed 2 years ago

Daming-TF commented 2 years ago

Hello, the network structure of MobileNetV3 seems to be different from that of the original author. The SE module of V3 is not a module that uses SENet. The author changed nn.Line to nn.Conv

dagshub[bot] commented 2 years ago

Join the discussion on DagsHub!

oferbaratz commented 2 years ago

Hey @Daming-TF Thanks for the issue, we are opening a bug and it will be fixed soon.

lkdci commented 2 years ago

Hi @Daming-TF , there are some Squeeze & Excite versions which uses Conv op insteaf of Linear.

Conv and Linear are equivalent when applied upon an input tensor with only one neuron per channels, which is the output after the Global average pooling operator. (BxCx1x1 or simply BxC).

In both cases the layers num MAC is represented by Cin x Cout.

There are indeed some hardware accelerators which prefer the Convolution instead of the fully connected one, so in some cases the convolution operation might run faster.

Daming-TF commented 2 years ago

Hello@lkdci oh that's true😄, thanks I got it