This repository contains the pytorch codes in our ACM MM 2019 paper "Visual Relationship Detection with Relative Location Mining", which is based on maskrcnn-benchmark.
Requirement(ours)
- pytorch,v1.0.1
- torchvision,v0.2.2
- CUDA, v9.0
- python3.6
- matlab
pip install yacs scipy tqdm
export INSTALL_DIR=$PWDcd $INSTALL_DIR
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext installcd $INSTALL_DIR
git clone https://github.com/zhouhaocv/RLM-Net.git
cd RLM-Net
python setup.py build develop-
See details from here
-
put "json_dataset" and "sg_dataset" in vrd/data/vrd/.
cd $INSTALL_DIR
cd vrd
a) generate vrd json
python gen_vrd_json.py
b) train the first stage(~4 hours)
python train_RLM_proposing_stage.py
c) train the second stage(~4.5 hours)
python train_RLM_predicate_stage.pypython test_RLM_Net.pyevaluate with predicate_step and relation_phrase_step.
If you have any problems, you can email to zhouhao_0039@sjtu.edu.cn.