Skip to content

Latest commit

 

History

History
32 lines (20 loc) · 999 Bytes

README.md

File metadata and controls

32 lines (20 loc) · 999 Bytes
Raw

Quantization

I have been reading and doing NN model optimization on weight quantization by clusering and weight pruing. I have implemented some basic weight quantization. I will be adding more to this repo.

I have been using libraries like TensorFlow for quantization but I would like to know under the hood of quantization.

I might have made some mistakes. If so, please comment and share so that I can fix it and nonethless we will both get to know something amazing about ongoing research in quantization.

Fixed point weight quantization

  1. float16, float32, float64 -> int16, int8, int4

Clustering based weight quantization

  1. float16, float32, float64 -> int8 [weight indexes] and float32/float16 [cluster centers]

Other weight quantization

  1. Min max weight quantization
  2. Threshold/Probability based weight quantization

Activation quantization

Coming soon

Hoffman encoding based quantization

Coming soon

Weight pruning

Coming soon

LSTM weight quantization