Network compression is the process of reducing the size of a trained network, either by removing certain layers or by shrinking layers, while maintaining performance. This notebook implements the a version of network compression using reinforcement learning algorithm similar to the one proposed in [1].
[1] Ashok, Anubhav, Nicholas Rhinehart, Fares Beainy, and Kris M. Kitani. "N2N learning: network to network compression via policy gradient reinforcement learning." arXiv preprint arXiv:1709.06030 (2017).
In this example the network compression notebook uses a Sagemaker docker image containing Ray, tensorflow and OpenAI Gym. The network modification module is
treated as a simulation where the actions produced by reinforcement learning algorithm (remove, shrink, etc.) can be run. The notebook has defined a set of actions for each module. It
demonstrates how one can use the SageMaker Python SDK script mode with a Tensorflow+Ray+Gym container. You can run
rl_network_compression_a3c_ray_tensorflow_NetworkCompressionEnv.ipynb from a SageMaker notebook instance.
This package relies on the SageMaker TensorFlow base RL container that has Ray installed.
The sample notebook demonstrates how to:
- Formulate network comrpession as a gym environment.
- Train the network compression system using Ray in a distributed fasion.
- Produce and download the checkpoints of best compressed models from S3.