tmq 0.1.0

Quantization-aware neural network layers using controllable soft-quantization

TMQ is a framework for training quantization-aware neural networks with an emphasis on ternary quantization. It uses a differentiable transfer functions on the weights of the following layer types:

• Linear
• Conv2d (both transpose and regular)

The transfer function can be parameterized to range from a simple linear mapping to a soft-staircase function, which "forces" the weight values to distinct quantization levels over time.

TMQ-type layers can be used as drop-in replacement into existing models, only requiring minimal change to the training code, resulting in models that can be compressed significantly.