Extrapolate Encoder from pre-trained 3D net to classify

[robsver]

Hello everyone,
I would like to know if it is possible to extrapolate from a pre-trained 3D model for segmentation (e.g.: UNETR) its encoder part and use it to perform classification.

How would you do this?

[KumoLiu]

Hi @robsver, you could add a hook in the constructor, a simple demo here.

        self.extra_outputs = {}
        def forward_hook(layer_name):
            def hook(module, input, output):
                self.extra_outputs[layer_name] = output

            return hook

        self.encoder4.register_forward_hook(forward_hook("layer4"))

Hope it helps, thanks!

[robsver]

Hi @KumoLiu, Thanks a lot.