Skip to content

deepsemantic/quantized_neural_network

Repository files navigation

=======

The quantization of CNN/LSTM

Ongoing work that tries to quantize the weights, activations as well as gradients to k-bit representation for model compression. Use bit-wise operation to speed up inference procedure in test phase.

Implemented tasks:

  • (1) Add new layer for weights quantization
  • (2) Add new layer for activations quantization
  • (3) CPU and GPU versions of above tasks

The implementation is based on Caffe (https://github.com/BVLC/caffe).

About

The quantization of CNN/LSTM

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published