Skip to content

Latest commit

 

History

History
70 lines (63 loc) · 4.72 KB

README.md

File metadata and controls

70 lines (63 loc) · 4.72 KB
Raw

QuEST

The official repository for QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning [ArXiv]

Update Log

(2024.2.28) Reorganized the code structures.

Features

QuEST achieves state-of-the-art performance on mutiple high-resolution image generation tasks, including unconditional image generation, class-conditional image generation and text-to-image generation. We also achieve superior performance on full 4-bit (W4A4) generation.

drawing

On ImageNet 256*256:

drawing

On Stable Diffusion v1.4 (512*512):

drawing

Get Started

Prerequisites

Make sure you have conda installed first, then:

git clone https://github.com/hatchetProject/QuEST.git
cd QuEST
conda env create -f environment.yml
conda activate quest

Usage

  1. For Latent Diffusion and Stable Diffusion experiments, first download relvant checkpoints following the instructions in the latent-diffusion and stable-diffusion repos from CompVis. We currently use sd-v1-4.ckpt for Stable Diffusion.
  2. The calibration data for LSUN-Bedrooms/Churches and Stable Diffusion (COCO) can be downloaded from the Q-Diffusion repository. We will upload the calibration data for ImageNet soon.
  3. Use the following commands to reproduce the models. act_bit=4 additionally use channel-wise quantization on a more hardware-friendly dimension, which reduces computation cost. Also, exclude the --running_stat argument for W4A4 quantization.
  4. Change line 151 in ldm/models/diffusion/ddim.py according the the number of timesteps you use, where you can replace '10' with the resulting number of --c // --cali_st, e.g. 200 // 20 = 10. Comment line 149~158 if you would like to inference with the FP model.
  5. We highly recommend to use checkpoints by QDiffusion and resume it with the --resume_w command.
# LSUN-Bedrooms (LDM-4)
python sample_diffusion_ldm_bedroom.py -r models/ldm/lsun_beds256/model.ckpt -n 100 --batch_size 20 -c 200 -e 1.0  --seed 40 --ptq  --weight_bit <4 or 8> --quant_mode qdiff --cali_st 20 --cali_batch_size 32 --cali_n 256 --quant_act --act_bit <4 or 8> --a_sym --a_min_max --running_stat --cali_data_path <cali_data_path> -l <output_path>
# LSUN-Churches (LDM-8)
python scripts/sample_diffusion_ldm_church.py -r models/ldm/lsun_churches256/model.ckpt -n 50000 --batch_size 10 -c 500 -e 0.0 --seed 40 --ptq --weight_bit <4 or 8> --quant_mode qdiff --cali_st 20 --cali_batch_size 32 --cali_n 256 --quant_act --act_bit <4 or 8> --cali_data_path <cali_data_path> -l <output_path>
# ImageNet
python sample_diffusion_ldm_imagenet.py -r models/ldm/cin256-v2/model.ckpt -n 50 --batch_size 50 -c 20 -e 1.0  --seed 40 --ptq  --weight_bit <4 or 8> --quant_mode qdiff --cali_st 20 --cali_batch_size 32 --cali_n 256 --quant_act --act_bit <4 or 8> --a_sym --a_min_max --running_stat --cond --cali_data_path <cali_data_path> -l <output_path>
# Stable Diffusion
python txt2img.py --prompt "a photograph of an astronaut riding a horse" --plms --cond --ptq --weight_bit <4 or 8> --quant_mode qdiff --quant_act --act_bit <4 or 8> --cali_st 25 --cali_batch_size 8 --cali_n 128 --no_grad_ckpt --split --running_stat --sm_abit 16 --cali_data_path <cali_data_path> --outdir <output_path>

Calibration dataset

We will release the calibration data. But you can also generate them yourself by using the following command (10 images per class over all timesteps):

python get_calibration_set_imagenet_ddim.py -r <FP_Model_Path> -n 10 --batch_size 10 -c 20 -e 1.0 -seed 40 -l output/ --cond

Evaluation

We use the ADM’s TensorFlow evaluation suite link for evaluating FID, sFID and IS. For Stable Diffusion, we generate 10,000 samples based on the prompts from the COCO2014 dataset calculate the average CLIP score.

Acknowledgement

This project is heavily based on LDM and Q-Diffusion.

Citation

If you find this work helpful, please consider citing our paper:

@misc{wang2024quest,
      title={QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning}, 
      author={Haoxuan Wang and Yuzhang Shang and Zhihang Yuan and Junyi Wu and Yan Yan},
      year={2024},
      eprint={2402.03666},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}