Knowing nothing about assement and tensor flow: Could the trained model be downloaded via s3 at init instead of being included inside the function zip? I believe the memory limits are more lax than the actual function limits. And the /tmp limit is 512MB.