HAWQ: Hessian AWare Quantization

HAWQ is an advanced quantization library written for PyTorch. HAWQ enables low-precision and mixed-precision uniform quantization, with direct hardware implementation through TVM.

For more details please see:

• HAWQ-V3 lightning talk in TVM Conference
• HAWQ-V2 presentation in NeurIPS'20

Installation

• PyTorch version >= 1.4.0
• Python version >= 3.6
• For training new models, you'll also need NVIDIA GPUs and NCCL
• To install HAWQ and develop locally:

git clone https://github.com/Zhen-Dong/HAWQ.git
cd HAWQ
pip install -r requirements.txt

Getting Started

Quantization-Aware Training

An example to run uniform 8-bit quantization for resnet50 on ImageNet.

export CUDA_VISIBLE_DEVICES=0
python quant_train.py -a resnet50 --epochs 1 --lr 0.0001 --batch-size 128 --data /path/to/imagenet/ --pretrained --save-path /path/to/checkpoints/ --act-range-momentum=0.99 --wd 1e-4 --data-percentage 0.0001 --fix-BN --checkpoint-iter -1 --quant-scheme uniform8

The commands for other quantization schemes and for other networks are shown in the model zoo.

Inference Acceleration

• Instructions on Hardware Implementation through TVM

Experimental Results

Table I and Table II in HAWQ-V3: Dyadic Neural Network Quantization

ResNet18 on ImageNet

Model	Quantization	Model Size(MB)	BOPS(G)	Accuracy(%)	Inference Speed (batch=8, ms)	Download
ResNet18	Floating Points	44.6	1858	71.47	9.7 (1.0x)	resnet18_baseline
ResNet18	W8A8	11.1	116	71.56	3.3 (3.0x)	resnet18_uniform8
ResNet18	Mixed Precision	6.7	72	70.22	2.7 (3.6x)	resnet18_bops0.5
ResNet18	W4A4	5.8	34	68.45	2.2 (4.4x)	resnet18_uniform4

ResNet50 on ImageNet

Model	Quantization	Model Size(MB)	BOPS(G)	Accuracy(%)	Inference Speed (batch=8, ms)	Download
ResNet50	Floating Points	97.8	3951	77.72	26.2 (1.0x)	resnet50_baseline
ResNet50	W8A8	24.5	247	77.58	8.5 (3.1x)	resnet50_uniform8
ResNet50	Mixed Precision	18.7	154	75.39	6.9 (3.8x)	resnet50_bops0.5
ResNet50	W4A4	13.1	67	74.24	5.8 (4.5x)	resnet50_uniform4

More results for different quantization schemes and different models (also the corresponding commands and important notes) are available in the model zoo.
To download the quantized models through wget, please refer to a simple command in model zoo.
Checkpoints in model zoo are saved in floating point precision. To shrink the memory size, BitPack can be applied on weight_integer tensors, or directly on quantized_checkpoint.pth.tar file.

Related Works

• HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021)
• HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks (NeurIPS 2020)
• HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision (ICCV 2019)

License

HAWQ is released under the MIT license.