Fork of the yrcong/RelTR repository, optimised for production rather than evaluation.

1. (Optional, but recommended) Create conda environment with conda create -n reltr python=3.6 and activate it
2. Install PyTorch and PyVision with pip install torch==1.6.0 torchvision==0.7.0 --extra-index-url https://download.pytorch.org/whl/cpu
3. pip3 install -r requirements.txt
4. Download the RelTR model and place it into ckpt

For help installing PyTorch, follow PyTorch instructions

To produce a scene graph from an image, run

python3 mkgraph.py --img_path $IMAGE_PATH --resume $MODEL_PATH --device cpu [--export_path graph.json]

Or, if you have a CUDA-capable device, replace cpu by cuda.

Set an environment variable OPENAI_API_KEY=<YOUR API KEY> with your OpenAI key.

pip3 install openai flask
python server.py