This repository contains the official implementation for the paper "I-ViT: Integer-only Quantization for Efficient Vision Transformer Inference". To the best of our knowledge, this is the first work on integer-only quantization for vision transformers.
Below are instructions of Pytorch code to reproduce the accuracy results of quantization-aware training (QAT). TVM benchmark is the TVM deployment project for reproducing latency results.
- To install I-ViT and develop locally:
git clone https://github.com/zkkli/I-ViT.git
cd I-ViT- You can quantize and fine-tune a single model using the following command:
python quant_train.py [--model] [--data] [--epochs] [--lr]
optional arguments:
--model: Model architecture, the choises can be:
deit_tiny, deit_small, deit_base, swin_tiny, swin_small, swin_base.
--data: Path to ImageNet dataset.
--epochs: recommended values are: [30, 60, 90], default=90.
--lr: recommended values are: [2e-7, 5e-7, 1e-6, 2e-6], default=1e-6.- Example: Quantize and fine-tune DeiT-T:
python quant_train.py --model deit_tiny --data <YOUR_DATA_DIR> --epochs 30 --lr 5e-7 Below are the Top-1 (%) accuracy results of our proposed I-ViT that you should get on ImageNet dataset.
| Model | FP32 | INT8 (I-ViT) | Diff. |
|---|---|---|---|
| ViT-S | 81.39 | 81.27 | -0.12 |
| ViT-B | 84.53 | 84.76 | +0.23 |
| DeiT-T | 72.21 | 72.24 | +0.03 |
| DeiT-S | 79.85 | 80.12 | +0.27 |
| DeiT-B | 81.85 | 81.74 | -0.11 |
| Swin-T | 81.35 | 81.50 | +0.15 |
| Swin-S | 83.20 | 83.01 | -0.19 |
We appreciate it if you would please cite the following paper if you found the implementation useful for your work:
@inproceedings{li2023vit,
title={I-vit: Integer-only quantization for efficient vision transformer inference},
author={Li, Zhikai and Gu, Qingyi},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={17065--17075},
year={2023}
}