Feed network from intermediate layers (and train these layers)


I am using a detection network. I would like to use a model that calculates 3 features vectors and then use them in the model head. I want to split the model in features and head so I can run them separately. Right now, I can extract the features vector using hooks, but I am not sure how can I feed those features in the head to compute the final output. In tensorflow you can just pass a dict with the input variables to the model and compute the prediction… An extra complication is the features are not in consecutive layer, so each goes to different modules in the head.

Thanks a lot,