Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation [ICML 2025]

October 8, 2025 · View on GitHub

This repository is the official implementaiton of our paper Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation

Setup Environment

For this project, we used python 3.8.5. We recommend setting up a new virtual environment:

In that environment, the requirements can be installed with:

pytorch:1.7.1-cuda11.0-cudnn8-devel

docker run --shm-size=8g --gpus all -e NVIDIA_VISIBLE_DEVICES=$GPU -it pytorch/pytorch:1.7.1-cuda11.0-cudnn8-devel

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv-full==1.3.7  # requires the other packages to be installed first
# or
pip install mmcv-full==1.3.7 -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7.0/index.html

pytorch/pytorch:2.0.0-cuda11.7-cudnn8-devel

docker run --shm-size=8g --gpus all -e NVIDIA_VISIBLE_DEVICES=$GPU -v /NAS:/NAS -v /SSD:/SSD -it pytorch/pytorch:2.0.0-cuda11.7-cudnn8-devel

pip install cityscapesscripts==2.2.0 cycler==0.10.0 gdown==4.2.0 humanfriendly==9.2 kiwisolver==1.2.0 kornia==0.5.8 matplotlib==3.4.2 numpy opencv-python pandas Pillow prettytable==2.1.0 pyparsing easydict pytz PyYAML scipy seaborn timm tqdm typing-extensions wcwidth yapf linklink
pip install mmcv-full==1.7.2 -f https://download.openmmlab.com/mmcv/dist/cu117/torch2.0.0/index.html
apt-get update && apt-get install -y libgl1-mesa-glx libglib2.0-0

Please, download the MiT-B5 ImageNet weights provided by SegFormer from their OneDrive and put them in the folder pretrained/. Further, download the checkpoint of HRDA on GTA→Cityscapes and extract it to the folder work_dirs/.

Trouble Shooting in Setup Environment

ImportError: libGL.so.1 is occured

apt-get install -y libgl1-mesa-glx

ImportError: libgthread-2.0.so.0 is occured

apt-get install -y libglib2.0-0

Setting

Dataset - GTA

Dataset - CityScapes (CS)

Dataset - IDD

Data Processing

  • GTA
    • python tools/convert_datasets/gta.py data/gta --nproc 8
  • CS
    • python tools/convert_datasets/cityscapes.py data/cityscapes --nproc 8
  • IDD
    • python tools/convert_datasets/idd.py data/idd --nproc 8

For Quantization

Create log, work_dirs, ckpt direcotries

  • log: log text
  • work_dirs: fp models
  • ckpt: save models dir
    • ckpt/CS: for CS dataset
    • ckpt/IDD: for IDD dataset
mkdir log
mkdir work_dirs
mkdir ckpt
cd ckpt; mkdir CS
cd ckpt; mkidr IDD

For Evaluation

Create result_txts

  • result_txts: save result texts of evalution
mkdir result_txts

Checkpoints

FP and quantized checkpoints are here

Quantization

Example

  • BRECQ
# Example commands (GTA -> CS, BRECQ, W8A8, seed: 2025)
./test_brecq.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 8 8 cs 2025 fixed

# Example commands (GTA -> IDD, BRECQ, W8A8, seed: 2025)
./test_brecq.sh work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a 8 8 idd 2025 fixed
  • QDROP
# Example commands (GTA -> CS, QDrop, W8A8, seed: 2025)
./test_qdrop.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 8 8 cs 2025 fixed

# Example commands (GTA -> IDD, QDrop, W8A8, seed: 2025)
./test_qdrop.sh work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a 8 8 idd 2025 fixed
  • HDRQ
# Example commands (GTA -> CS, HDRQ, W8A8, seed: 2025)
./test_hdrq.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 8 8 cs 2025 fixed

# Example commands (GTA -> IDD, HDRQ, W8A8, seed: 2025)
./test_hdrq.sh work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a 8 8 idd 2025 fixed

Evaluation (Each Quantization)

./eval.sh [config path] [model_path] [n_gpus]

ex1) HDRQ (IDD)

# Example commands (HDRQ, W4A4, 4 GPUS, quant_seed: 1005)
./eval.sh work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a ckpt/IDD/HDRQ_PTQTestR101_W4A4_fixed_IDD_seed1005.pt 4

ex2) HDRQ (CS)

# Example commands (HDRQ, W4A4, 4 GPUS, quant_seed: 1005)
./eval.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 ckpt/CS/HDRQ_PTQTestR101_W4A4_fixed_IDD_seed1005.pt 4

ex3) QDROP (IDD)

# Example commands (QDROP, W4A4, 4 GPUS, quant_seed: 1005)
./eval.sh work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a ckpt/IDD/QDROP_PTQTestR101_W4A4_fixed_IDD_seed1005.pt 4

ex4) QDROP (CS)

# Example commands (QDROP, W4A4, 4 GPUS, quant_seed: 1005)
./eval.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 ckpt/CS/QDROP_PTQTestR101_W4A4_fixed_IDD_seed1005.pt 4

Evaluation (Model Merging)

# Quantization Type: BRECQ, QDROP, HDRQ, FLEX, SMQ

./eval_merge_seed.sh [config path of dataset1] [config path of dataset2] [model_path of dataset1] [model_path of dataset2] [n_gpus] [quantization type] [port_num (for ddp)] [eval_seed]

# using default port num (29703)
./eval_merge_seed.sh [config path of dataset1] [config path of dataset2] [model_path of dataset1] [model_path of dataset2] [n_gpus] [quantization type] [eval_seed]


# eval_merge.sh also working! (without setting seed)
# you can skip a "port_num" option (default port_num is 29703)
./eval_merge.sh [config path of dataset1] [config path of dataset2] [model_path of dataset1] [model_path of dataset2] [n_gpus] [quantization type] [port_num (for ddp)]

ex1) QDROP (./eval_merge_seed.sh)

# Example commands (QDROP, W4A4, 4 GPUS, quant_seed: 200, eval_seed: 5000)
./eval_merge_seed.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a ckpt/CS/QDROP_PTQTestR101_W4A4_fixed_CS_seed200.pt ckpt/IDD/QDROP_PTQTestR101_W4A4_fixed_IDD_seed200.pt 4 QDROP 5000

ex2) HDRQ (./eval_merge.sh)

# Example commands (HDRQ, W4A4, 4 GPUS, quant_seed: 200, eval_seed: 5000)
./eval_merge.sh work_dirs/fp_gtaHR2csHR_hrda_r101_a5271 work_dirs/fp_gtaHR2iddHR_hrda_r101_b958a ckpt/CS/HDRQ_PTQTestR101_W4A4_fixed_CS_seed200.pt ckpt/IDD/HDRQ_PTQTestR101_W4A4_fixed_IDD_seed200.pt 4 HDRQ 29700

Reference