How to extract pre-classification-head feature embeddings for RealMLP?

[nearpit]

Question / Feature Request

Is there a supported way to extract feature embeddings from the layer immediately before the classification head?

Context

In the standalone version, this is straightforward using a forward hook:

model_.model_[-2].register_forward_hook(lambda x: ...)

However, I couldn't find an equivalent layer to attach a hook to in the new implementation.

I assume, it should be something like:

clf = clf = RealMLP_TD_Classifier()
layer = clf.alg_interface_.model.model ...
layer.register_forward_hook(lambda x: ...)

What I'm looking for

Either:

• A pointer to the correct layer/module to hook into, or
• A built-in API for extracting intermediate representations

Apologies if this is already documented somewhere and I missed it — happy to be pointed in the right direction.