Keras dependencies needed for prediction only in AWS Lambda

Question

I'm trying to package my trained Keras classifier models (Theano backend) and their HDF5 weights as AWS Lambda functions that process classification requests coming from an API. Lambda has a 50MB limit for the zip file that contains the function code and dependencies, so I'm wondering whether any of the dependencies needed for training are not needed for prediction so that I can leave them out of the Lambda zip file. Right now I'm just squeaking by at ~45MB with all dependencies (numpy, scipy, etc.), but I'd like to drop any dead weight if possible.

Answer 1

To save some space, strip your binaries

find . | xargs strip