how to get features from intermediate layers ?

[Jamesgender]

Thanks for the excellent project.
My question is when I want to use intermediate features for U-net decoder, I dont know how to get them. The issue#18 shows that we can try to change the code and get the outputs from the model. That looks ok.
And another question shows that features_only=True argument can get the features. Yes, but I think that is different. When I want to train a network, the features I want to get are from the model I created before. The features_only seems to create some new features and the weights are fixed because of the pretrained weights. And these features seems to be hard to join the training.
So I can only change the code to get the features, or any other ways?
Thanks for helping me!

[Jamesgender]

Maybe my understanding of the features only is not right.

[hankyul2]

Hi @Jamesgender

You can use features_only=True to extract intermediate features for training the U-NET decoder. As can be seen in _features.py, the features_only=True option is just building a helper class to extract the feature and does not interference with the pre-trained weight. If you want to know about how to use it, please see this docs.

I want this answer could help you. Thank you.

Hankyul.

[Jamesgender]

Thanks for your helpful answer! My understanding was wrong, Features only is actully effective.