This repo contains PyTorch implementation for paper HAQ: Hardware-Aware Automated Quantization with Mixed Precision (CVPR2019, oral)
@inproceedings{haq,
author = {Wang, Kuan and Liu, Zhijian and Lin, Yujun and Lin, Ji and Han, Song},
title = {HAQ: Hardware-Aware Automated Quantization With Mixed Precision},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2019}
}
Other papers related to automated model design:
AMC: AutoML for Model Compression and Acceleration on Mobile Devices (ECCV 2018)
ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware (ICLR 2019)
We evaluate this code with Pytorch 1.1 (cuda10) and torchvision 0.3.0, you can install pytorch with conda:
# install pytorch
conda install -y pytorch torchvision cudatoolkit=10.0 -c pytorch
And you can use the following command to set up the environment:
# install packages and download the pretrained model
bash run/setup.sh
(If the server is down, you can download the pretrained model from google drive: mobilenetv2-150.pth.tar)
Current code base is tested under following environment:
If you already have the ImageNet dataset for pytorch, you could create a link to data folder and use it:
# prepare dataset, change the path to your own
ln -s /path/to/imagenet/ data/
If you do not have the ImageNet yet, you can download the ImageNet dataset and move validation images to labeled subfolders. To do this, you can use the following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
We use a subset of ImageNet in the linear quantizaiton search phase to save the training time, to create the link of the subset, you can use the following tool:
# prepare imagenet100 dataset
python lib/utils/make_data.py
# K-Means quantization, for model size
bash run/run_kmeans_quantize_search.sh
# Linear quantization, for latency/energy
bash run/run_linear_quantize_search.sh
python rl_quantize.py --help
# preserve ratio 10%
strategy = [6, 6, 5, 5, 5, 5, 4, 5, 5, 4, 5, 5, 5, 5, 5, 5, 3, 5, 4, 3, 5, 4, 3, 4, 4, 4, 2, 5, 4, 3, 3, 5, 3, 2, 5, 3, 2, 4, 3, 2, 5, 3, 2, 5, 3, 4, 2, 5, 2, 3, 4, 2, 3, 4]
You can follow the following bash file to finetune the K-Means quantized model to get a better performance:
bash run/run_kmeans_quantize_finetune.sh
# preserve ratio 60%
strategy = [[8, -1], [7, 7], [5, 6], [4, 6], [5, 6], [5, 7], [5, 6], [7, 4], [4, 6], [4, 6], [7, 7], [5, 6], [4, 6], [7, 3], [5, 7], [4, 7], [7, 3], [5, 7], [4, 7], [7, 7], [4, 7], [4, 7], [6, 4], [6, 7], [4, 7], [7, 4], [6, 7], [5, 7], [7, 4], [6, 7], [5, 7], [7, 4], [6, 7], [6, 7], [6, 4], [5, 7], [6, 7], [6, 4], [5, 7], [6, 7], [7, 7], [4, 7], [7, 7], [7, 7], [4, 7], [7, 7], [7, 7], [4, 7], [7, 7], [7, 7], [4, 7], [7, 7], [8, 8]]
You can follow the following bash file to finetune the linear quantized model to get a better performance:
bash run/run_linear_quantize_finetune.sh
python finetune.py --help
You can download the pretrained quantized model like this:
# download checkpoint
mkdir -p checkpoints/resnet50/
mkdir -p checkpoints/mobilenetv2/
cd checkpoints/resnet50/
wget https://hanlab.mit.edu/files/haq/resnet50_0.1_75.48.pth.tar
cd ../mobilenetv2/
wget https://hanlab.mit.edu/files/haq/qmobilenetv2_0.6_71.23.pth.tar
cd ../..
(If the server is down, you can download the pretrained model from google drive: qmobilenetv2_0.6_71.23.pth.tar)
You can evaluate the K-Means quantized model like this:
# evaluate K-Means quantization
bash run/run_kmeans_quantize_eval.sh
Models | preserve ratio | Top1 Acc (%) | Top5 Acc (%) |
---|---|---|---|
resnet50 (original) | 1.0 | 76.15 | 92.87 |
resnet50 (10x compress) | 0.1 | 75.48 | 92.42 |
You can evaluate the linear quantized model like this:
# evaluate linear quantization
bash run/run_linear_quantize_eval.sh
Models | preserve ratio | Top1 Acc (%) | Top5 Acc (%) |
---|---|---|---|
mobilenetv2 (original) | 1.0 | 72.05 | 90.49 |
mobilenetv2 (0.6x latency) | 0.6 | 71.23 | 90.00 |