Efficient net as backbone feature extractor

How to use efficientNet as backbone CNN model for feature extraction, so that embeddings of images can be generated.Most of example on GitHub use 4 layer ConvNet so I can not understand how to use same thing for large CNN model. Like there are implementation of efficient-net for Torch, so what steps I need to use them as feature extractor?
I am using this efficient net code which implemented network in PyTorch -

Did you figure this out? Currently trying to implement this as a feature extractor and was a bit confused on which layer to access and how to access the specific layer for feature output.