EfficientNet without Top Layers

[nhorangl]

I tried to remove top layers like follows:

base_model = EfficientNet.from_pretrained('efficientnet-b3')
model = nn.Sequential(*list(base_model.children()[:-3]))

x = torch.randn((1,3,300,300))
model(x)

It throws "NotImplementedError" for forward(). Is there a way to not include top layers?

[Cyril9227]

Overwrite the forward method to yield the features map of the desired layer

[nwschurink]

Yes, or add an option to the initialization of the network that allows dropping these layers (similar to how EfficientNet is implemented in Tensorflow).
I made the adjustments for a EfficientNet3D fork of this repo -> see pull request here shijianjian#4
If you want I can also change the code for this repo and post a pull request here

[23pointsNorth]

@nwschurink 's solution is much better, but in the mean time, an alternative is to substitute them with Identity layers:

model._fc = nn.Identity()
model._swish = nn.Identity()

[lukemelas]

This is now integrated :)

[23pointsNorth]

@lukemelas is it going to be in the next pip version? Or do you know when it's planned for release?