An experimental compression pipeline for TensorFlow.

This project aims to implement a TensorFlow version of the Deep Compression pipeline outlined by Han et al.

• Initial results:
  • Running compression_test.py can generate uncompressed_model and compressed_model protobuf files from a simple MNIST model for comparison.
  • Note that there is no difference in the protobuf file sizes until you apply gzip or some other compression tool.
  • The provided protobuf files were generated using a LeNet-300-100 model trained for MNIST digit classification.

• Next steps:
  • Improve accuracy preservation and increase compression ratio.
    • Learning both Weights and Connections for Efficient Neural Networks suggests ~12x parameter compression without loss of accuracy can be obtained for this architecture.
  • Refactor compression logic so that it can be more easily applied to other networks.

• Not started.

• Not started.