ADEPT-Dataset-Release

This is the dataset generation code for ADEPT (Approximate Derenderer, Extended Physics, and Tracking).

Modeling Expectation Violation in Intuitive Physics with Coarse Probabilistic Object Representations

Kevin Smith*, Lingjie Mei*, Shunyu Yao, Jiajun Wu, Elizabeth S. Spelke, Joshua B. Tenenbaum, Tomer Ullman (* indicates equal contribution)

NeurIPS 2019, CogSci 2020

Paper BibTeX Website

For the model, see ADEPT-Model-Release

Getting started

1. Clone this directory and create a conda environment for ADEPT Dataset Generation.

   git clone https://github.com/JerryLingjieMei/ADEPT-Dataset-Release
   cd ADEPT-Dataset-Release
   conda create -n adept-dataset python=3.10
   conda activate adept-dataset

2. Run the setup script, which install Blender 3.1.2 and python packages used in the scripts.

    bash install.sh

3. Configure the ffmpeg in dataset.make_video to the one used by your system.

4. (Optional) By default, the script in run on one machine only. Set up get_host_id in util.misc to run the job on multiple devices. Append to each script with --stride argument will enable job assignment.

5. (Optional)

   • ShapeNet objects are by default not included unless the scripts below are run.

   • To render ShapeNet objects, please download ShapeNet Core V2 from its official website. Change SHAPE_NET_FOLDER in phys_sim/data/builder/collect_obj.sh to the path of ShapeNet meshes, and run thar script. To turn them into .blend files, run

    # Single machine
    ./blender/blender --background --python render/data/builder/collect_blend.py #Map phase
    ./blender/blender --background --python render/data/builder/collect_blend.py -- --reduce #Reduce phase
    # Multiple machines
    ./blender/blender --background --python render/data/builder/collect_blend.py -- --stride 8 #On each machine
    ./blender/blender --background --python render/data/builder/collect_blend.py -- --reduce --stride 8 #On a single machine

Dataset generation

1. Generate training set (e.g. with 1000 videos) by running

   # Single machine
   ./blender/blender --background --python dataset/generate_train.py --end 1000 --
   # Multiple machines
   ./blender/blender --background --python dataset/generate_train.py --end 1000 -- --stride 8 #On each machine

2. Generate human test set by running

   # Single machine
   ./blender/blender --background --python dataset/human/generate_human.py --
   # Multiple (e.g. 8) machines
   ./blender/blender --background --python ataset/human/generate_human.py -- --stride 8 #On each machine

Evaluation

1. Evaluating the relative accuracy on human test set. If you have a experiment output folder that contains .txt files containing the scores of all human test cases, run

   python3 dataset/human/collect_reuslts.py --summary_folder ${SUMMARY_FOLDER} #Score in SUMMARY_FOLDER/results
   python3 dataset/human/collect_reuslts.py --summary_folder ${SUMMARY_FOLDER} --output_folder ${OUTPUT_FOLDER} #Custom output folder

   Or you may get the relative accuracy from a json file that contains a dictionary mapping case name to its score:

   python3 dataset/human/collect_results.py --summary_file ${SUMMARY_FILE} --output_folder ${OUTPUT_FOLDER} #Custom output folder