Home

Awesome

LSQuantization

The PyTorch implementation of Learned Step size Quantization (LSQ) in ICLR2020 (unofficial)


The related project with training code: https://github.com/hustzxd/EfficientPyTorch (sorry for late.)

The project is working in progress, and experimental results on ImageNet are not as good as shown in the paper.

ImageNet

LSQfp32w4a4w3a3w2a2w8a8(1epoch, quantize data)
AlexNet56.55, 79.0956.96, 79.46 55.31, 78.5951.18, 75.38
ResNet1869.76, 89.0870.26, 89.34 69.45, 88.8569.68 88.92

Hyper-parameter

Hyper-parameterLRLR-schedulerepochsbatch-sizewd
AlexNet-w4a40.01CosineAnnealingLR9020481e-4
ResNet18-w4a40.01CosineAnnealingLR905121e-4

Experimental Results

====VGGsmall + Cifar10=======

VGGsmall
fp3293.34
w4a494.26
w3a393.89
w2a293.42
<img src="alpha_curve.png" width="50%" height="50%">