joesonchan / hl-net

Heterophily Learning Network (HL-Net)

This is an official implementation for CVPR 2022 paper "HL-Net: Heterophily Learning Network for Scene Graph Generation".

Contents

0. Installation
1. Dataset setup
2. Config Introduction
3. Training
4. Validation and Testing
5. Models

Introduction

This project is based on the code of Scene Graph Benchmark in Pytorch.

Method overview

Installation

• Clone this repo

git clone https://github.com/siml3/HL-Net.git
cd HL-Net

• Create a conda virtual environment and activate it (optional)

conda create -n hlnet -y
conda activate hlnet

• Prepare proper pytorch environment for your device.

• Install dependencies

  conda install ipython
  conda install scipy
  conda install h5py
  pip install ninja yacs cython matplotlib tqdm opencv-python-headless overrides
  
  
  git clone https://github.com/cocodataset/cocoapi.git
  cd cocoapi/PythonAPI
  python setup.py build_ext install
  cd ../../
  
  # GPU and Cuda environment is necessary for proper compilation of this repo
  git clone https://github.com/NVIDIA/apex.git
  cd apex
  python setup.py install --cuda_ext --cpp_ext
  cd ../
  

• Compile extensions and install this repo in develop mode

  python setup.py build develop
  

Dataset

Please refer to the DATASET.md

Training

Detector Fine-tune (optional)

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --master_port 10001 --nproc_per_node=4 tools/detector_pretrain_net.py --config-file "configs/e2e_relation_detector_X_101_32_8_FPN_1x.yaml" SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.STEPS "(30000, 45000)" SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 MODEL.RELATION_ON False SOLVER.PRE_VAL False OUTPUT_DIR /path/to/output 

Training on Scene Graph Generation

The training scripts are available in the folder scripts

If you have multi gpus, try the command as the follows for SGCLS:

# four cards
PYTHONPATH=$PWD CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --master_port 10022 --nproc_per_node=4 tools/relation_train_net.py --config-file "configs/e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR "exp/HL-Net-sgcls" SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 4 DTYPE "float16" SOLVER.PRE_VAL True MODEL.ROI_RELATION_HEAD.PREDICTOR HLNetPredictor MODEL.ROI_RELATION_HEAD.MP_LAYER_NUM 5 SOLVER.BASE_LR 0.0025 GLOVE_DIR glove MODEL.ROI_RELATION_HEAD.PPR_ALPHA -0.5

Testing

gpu_id=0,1
port=10086
gpu_num=2
output_dir="home/XXXX/exp/.."

PYTHONPATH=$PWD CUDA_VISIBLE_DEVICES=${gpu_id} python -m torch.distributed.launch --master_port ${port} --nproc_per_node=${gpu_num} tools/relation_test_net.py --config-file "${output_dir}/config.yml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False OUTPUT_DIR "${output_dir}" TEST.IMS_PER_BATCH ${gpu_num} DTYPE "float16" GLOVE_DIR glove

Models

Here we provide our pretrained model via BaiduNetDisk:

Link：https://pan.baidu.com/s/1ujghYTjDB8MEgujQVVhCEQ Extraction code：1020

Citation

If HL-Net is helpful for your research, we'd really appreciate it if you could cite this paper:

@inproceedings{lin2022hl,
  title={HL-Net: Heterophily Learning Network for Scene Graph Generation},
  author={Lin, Xin and Ding, Changxing and Zhan, Yibing and Li, Zijian and Tao, Dacheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19476--19485},
  year={2022}
}