This is the pytorch implementation for the paper: Generalizable Mixed-Precision Quantization via Attribution Rank Preservation, which is accepted to ICCV2021.

GMPQ: Generalizable Mixed-Precision Quantization via Attribution Rank Preservation

This is the pytorch implementation for the paper: Generalizable Mixed-Precision Quantization via Attribution Rank Preservation, which is accepted to ICCV2021. This repo contains searching the quantization policy via attribution preservation on small datasets including CIFAR-10, Cars, Flowers, Aircraft, Pets and Food, and finetuning on largescale dataset like ImageNet using our proposed GMPQ.

Quick Start

Prerequisites

• python>=3.5
• pytorch>=1.1.0
• torchvision>=0.3.0
• other packages like numpy and sklearn

Dataset

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 don't have the ImageNet, you can use the following script to download it: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh

For small datasets which we search the quantization policy on, please follow the official instruction:

• CIFAR-10
• Stanford Cars
• Oxford 102 Flowers
• FGVC Aircraft
• Oxford-IIIT Pet
• Food-101

Searching the mixed-precision quantization policy

For a specific small dataset, you should first pretrain a full-precision model to provide supervision for attribution rank consistency preservation and save it to pretrain_model.pth.tar.

After that, you can start searching the quantization policy. Take ResNet18 and CIFAR-10 for example:

CUDA_VISIBLE_DEVICES=0,1 python search_attention.py \
-a mixres18_w2346a2346  -fa qresnet18_cifar  --epochs 25  --pretrained pretrain_model.pth.tar --aw 40 \
--dataname cifar10 --expname cifar10_resnet18  --cd 0.0003   --step-epoch 10    \
--batch-size 256   --lr 0.1   --lra 0.01 -j 16  \
  path/to/cifar10 \

It also supports other network architectures like ResNet50 and other small datasets like Cars, Flowers, Aircraft, Pets and Food.

Finetuning on ImageNet

After searching, you can get the optimal quantization policy, with the checkpoint arch_checkpoint.pth.tar. You can run the following command to finetune and evaluate the performance on ImageNet dataset.

CUDA_VISIBLE_DEVICES=0,1 python main.py     \
 -a qresnet18                 \
 --ac arch_checkpoint.pth.tar \
 -c checkpoints/train_resnet18   \
 --data_name imagenet          \
 --data path/to/imagenet           \
 --epochs 100                     \
 --pretrained pretrained.pth.tar
 --lr 0.01                    \
 --gpu_id 1,2,3     \
 --train_batch_per_gpu 192              \
 --wd 4e-5                       \
 --workers 32                    \