Piecewise-Quantization

PyTorch implementation of Near-Lossless Post-Training Quantization of Deep Neural Networks via a Piecewise Linear Approximation

Usage

There are 5 main arguments

1. quantize: whether to quantize parameters(per-channel) and activations(per-tensor).
2. imagenet_path: path to folder contains train/val folder of imagenet data
3. model: the type of model, should be one of ['mobilenetv2', 'resnet50', 'inceptionv3'], default to mobilenetv2
4. qtype: the type of quantization for weights, should be one of ['uniform', 'pws', 'pwg', 'pwl'], default to uniform
5. bits_weight: number of bits for weight quantization, default to 8

run the 4-bits quantized pws mobilenetv2 model by:

python main_cls.py --quantize --qtype pws --model mobilenetv2 --bits_Weight 4

Notes

Fake quantization

The quantization in this repo is fake quantization. Inference is NOT pure Int8 arithmetics.

TODO

• [x] Uniform quantization
• [x] PWS quantization
• [ ] update results for classification models
• [ ] PWG quantization
• [ ] PWL quantization
• [ ] detection model
• [ ] segmentation model