Skip to content

Latest commit

 

History

History
85 lines (64 loc) · 3.92 KB

README.md

File metadata and controls

85 lines (64 loc) · 3.92 KB
Raw

SQuant ICLR2022

Publication

If you use SQuant in your research, please cite our paper:

@inproceedings{
guo2022squant,
title={{SQ}uant: On-the-Fly Data-Free Quantization via Diagonal Hessian Approximation},
author={Cong Guo and Yuxian Qiu and Jingwen Leng and Xiaotian Gao and Chen Zhang and Yunxin Liu and Fan Yang and Yuhao Zhu and Minyi Guo},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=JXhROKNZzOc}
}

Environment

conda create -n squant python==3.8
conda activate squant
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
pip install pytorchcv==0.0.51
pip install ./squant-kernel

ImageNet Path

Please prepare your dataset with this script and set your dataset path using "--dataset_path /your/imagenet_path".

Run SQuant

# Quantize resnet-18 in 4-bit with SQuant
python -u main.py --model resnet18 --mode squant -wb 4 -ab 4  -s 12

# Test all models in Table 1, 2, and 3 of the ICLR2022 paper.
./run.sh

# Ablation study results in Table 4 and 5 of the ICLR2022 paper.
./run_ablation.sh

Quantization Rules

  1. The weight and activation should be quantized correctly when inference.

    • Weight is per-(output) channel quantization.
      • Affine quantzation, same as ZeroQ and GDFQ.
      • For all results of SQuant, we only set the min/max as the quantization range.
    • Activation (input) must be per-tensor quantization: all elements must have ONE quantization range.
      • The clip method is a BN-based method from DFQ with a wider clip range.
      • All models have the same range setting (alpha) of the BN-based method under the same bit-width (similar with ACIQ (alpha)).
    • The first layer input does not need quantization.
      • Same as other frameworks, such as ZeroQ and GDFQ.
    • The last layer (FC) input uses 8bit quantization.
      • Lower bit-width than other frameworks, such as ZeroQ and GDFQ.
      • Note that other frameworks didn't quantize the FC input activation after average pooling.

  2. The quantization procedure MUST not involve any real dataset information. (data-free)

    • Weight quantization should finish before the inference without training/validation dataset.
    • Activation can only quantize in inference runtime. However, its quantization range should be set before inference without a training/validation dataset.

  3. Test all models

    • All models use the same hyperparameters under the same bit-width.
    • The results should fit the results (with $\pm 0.2$ error) presented in the manuscript.

  4. Ablation study

    • Reproduce the same results as in the manuscript.

SQuant is Accurate.

As shown below, SQuant significantly outperforms all other SOTA DFQ methods, even with synthetic dataset calibrating their networks. The benefit of SQuant becomes more prominent as the bit-width decreases. SQuant outperforms the PTQ methods, i.e., DFQ, ZeroQ, and DSG, by more than 30% on all models with 4-bit quantization. It is noteworthy that SQuant surpasses GDFQ in all cases and even surpasses more than 15% in ResNet50 under 4-bit quantization, although GDFQ is a quantization-aware training method.

SQuant is Fast.

A single layer takes SQuant just 3 milliseconds on average because SQuant does not involve complex algorithms, such as back-propagation and fine-tuning. That means we can implement the SQuant algorithm on inference-only devices such as smartphones and IoT devices and quantize the network on the fly.

SQuant is more than $10,000\times$ faster than GDFQ in quantization time with higher accuracy.