Generative Data Free Model Quantization with Knowledge Matching for Classification

We provide PyTorch implementation for "Generative Data Free Model Quantization with Knowledge Matching for Classification".

Dependencies

Python 3.7
PyTorch 1.7.1
dependencies in requirements.txt

Getting Started

Installation

Clone this repo:

 git clone https://github.com/ZSHsh98/KMDFQ.git
 cd KMDFQ

Install pytorch and other dependencies.
```
 pip install -r requirements.txt
```

Set the paths of datasets for testing

Set the "dataPath" in "cifar100_resnet20.hocon" as the path root of your CIFAR-100 dataset. For example:
```
 dataPath = "/home/datasets/Datasets/cifar"
```
Set the "dataPath" in "imagenet_resnet18.hocon" as the path root of your ImageNet dataset. For example:
```
 dataPath = "/home/datasets/Datasets/imagenet"
```

Training

To quantize the pretrained ResNet-20 on CIFAR-10 to 4-bit:

python main.py 
--conf_path ./cifar10_resnet20.hocon \
--id 1 \
--CE_WEIGHT 1 \
--BNS_WEIGHT 1 \ 
--FEATURE_WEIGHT 1  
--warmup_epochs 20 \ 
--visible_devices 0 \
--qw 4 \
--qa 4

To quantize the pretrained ResNet-20 on CIFAR-100 to 5-bit:

python main.py 
--conf_path ./cifar100_resnet20.hocon \
--id 1 \
--CE_WEIGHT 1 \
--BNS_WEIGHT 1 \ 
--FEATURE_WEIGHT 1  
--warmup_epochs 20 \ 
--visible_devices 0 \
--qw 5 \
--qa 5

To quantize the pretrained ResNet-18 on ImageNet to 6-bit:

python main.py 
--conf_path ./imagenet_resnet18.hocon \
--id 1 \
--CE_WEIGHT 1 \
--BNS_WEIGHT 1  \
--FEATURE_WEIGHT 1 \
--warmup_epochs 50 \
--visible_devices 6 \
--selenet resnet18 \
--qw 6 \
--qa 6

Results

Dataset	Model	Pretrain Top1 Acc(%)	W8A8	W6A6	W5A5	W4A4
CIFAR-10	ResNet-20	93.89	94.05	93.94	93.67	92.24
CIFAR-100	ResNet-20	70.33	70.53	70.35	69.68	67.15
ImageNet	ResNet-18	71.58	70.73	70.61	69.93	64.39
ImageNet	BN-VGG16	74.38	72.33	72.29	71.89	68.79
ImageNet	Inception v3	77.63	76.45	76.43	75.54	71.22
ImageNet	MobileNet v2	73.08	72.58	72.21	71.03	63.48
ImageNet	ShuffleNet	65.16	61.90	60.95	56.73	28.26
ImageNet	ResNet-50	77.76	77.50	77.21	75.67	68.84

Note that we use the pretrained models from pytorchcv.

Citation

If this work is useful for your research, please cite our paper:

@article{xu2023KMDFQ,
title = {Generative Data Free Model Quantization with Knowledge Matching for Classification},
author = {Shoukai Xu, Shuhai Zhang, Jing Liu, Bohan Zhuang, Yaowei Wang , Mingkui Tan},
journal={IEEE Transactions on Circuits and Systems for Video technology},
publisher={IEEE},
year = {2023}
}