Bongard-Tool

This is the repo for the implementation of PKU CoRe 22 Submission Bongard-Tool: Tool Concept Induction from Few-Shot Exemplars. Here you can find our data collection scripts and baseline experiment codebase. Work in progress, so please stay tuned.

Generating dataset

Stay tuned...

Running baselines

First, clone this repository.

git clone https://github.com/YuyangLee/Bongard-Tool.git
cd Bongard-Tool

To get a quick test on these baselines, you can download all images from this link.

wget https://aidenology-assets.s3.ap-southeast-1.amazonaws.com/dev/bongard-tools/datasets/FuncTools.1.2.Processed.zip -P Dataset/

Then, you can generate tasks by running:

python data/data_process.py --data_root Dataset/ --name_path toolnames/names.1.2.1.json

This will automatically read subdirectories of images under --data_root and dump task JSON files under --data_root.

After the task generation, you should change the configuration files before you run the scripts.

Take CNN-Baseline as an example.

1. Find the configuration file baselines/configs/configs_V2/train_cnn_shapebd.yaml and fill in data_root with the root directory of your dataset.
2. Run the training script bash baselines/scripts/run_cnn_model.sh