Pytorch Implementation for the Additive Powers of Two Quantization

In this repo, Additive Powers-of-Two (APoT) quantization technique is applied to quantize ResNet models and then evaluated on the MNIST dataset.

Note: The original repo is at https://github.com/yhhhli/APoT_Quantization.git where the APoT quantization technique is evaluated on ImageNet and CIFAR-10 datasets.

To train the ResNet-18 model on 5 bit precision for 25 epochs

python main.py --arch resnet18 --bit 5 --epochs 25

Model	Precision	Hyper-Params	Accuracy
ResNet-18	5-bit	batch128_lr0.01_wd0.0001_25epoch	99.58
ResNet-18	4-bit	batch128_lr0.01_wd0.0001_25epoch	99.49
ResNet-18	3-bit	batch128_lr0.01_wd0.0001_25epoch	99.30
ResNet-18	2-bit	batch128_lr0.01_wd0.0001_25epoch	98.45
ResNet-10	5-bit	batch128_lr0.01_wd0.0001_25epoch	99.78
ResNet-10	4-bit	batch128_lr0.01_wd0.0001_25epoch	99.64
ResNet-10	3-bit	batch128_lr0.01_wd0.0001_25epoch	99.39
ResNet-10	2-bit	batch128_lr0.01_wd0.0001_25epoch	98.72

Compared with Uniform Quantization

Pass the value 0 to the argument power to switch to uniform quantization

python main.py --arch resnet18 --bit 5 --epochs 25 --power 0

Results:

Model	Precision	Hyper-Params	Accuracy	Compared with APoT
ResNet-18	5-bit	batch128_lr0.01_wd0.0001_25epoch	99.40	-0.18
ResNet-18	4-bit	batch128_lr0.01_wd0.0001_25epoch	99.39	-0.10
ResNet-18	3-bit	batch128_lr0.01_wd0.0001_25epoch	99.24	-0.06
ResNet-18	2-bit	batch128_lr0.01_wd0.0001_25epoch	98.31	-0.14
ResNet-10	5-bit	batch128_lr0.01_wd0.0001_25epoch	99.68	-0.10
ResNet-10	4-bit	batch128_lr0.01_wd0.0001_25epoch	99.54	-0.10
ResNet-10	3-bit	batch128_lr0.01_wd0.0001_25epoch	99.25	-0.14
ResNet-10	2-bit	batch128_lr0.01_wd0.0001_25epoch	98.45	-0.27

To evaluate the models, you can run:

python main.py --arch resnet 18 -e --pretrained result/resnet18_4bit/model_best.pth.tar --bit 4