[paper]
Physion: Evaluating Physical Prediction from Vision in Humans and MachinesDaniel M. Bear, Elias Wang, Damian Mrowca, Felix J. Binder, Hsiao-Yu Fish Tung, R.T. Pramod, Cameron Holdaway, Sirui Tao, Kevin Smith, Fan-Yun Sun, Li Fei-Fei, Nancy Kanwisher, Joshua B. Tenenbaum, Daniel L.K. Yamins, Judith E. Fan
This is the official implementation of particle-based models (GNS and DPI-Net) on the Physion dataset. The code is built based on the original implementation of DPI-Net (https://github.com/YunzhuLi/DPI-Net).
Contact: [email protected] (Fish Tung)
Papers of GNS and DPI-Net:
** Learning to Simulate Complex Physics with Graph Networks ** [paper]
Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, Peter W. Battaglia
** Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids ** [website] [paper]
Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B. Tenenbaum, Antonio Torralba **
Demo
Rollout from our learned model (left is ground truth, right is prediction)
Installation
Clone this repo:
git clone https://github.com/htung0101/DPI-Net-p.git
cd DPI-Net-p
git submodule update --init --recursive
Install Dependencies if using Conda
For Conda users, we provide an installation script:
bash ./scripts/conda_deps.sh
pip install pyyaml
To use tensorboard for training visualization
pip install tensorboardX
pip install tensorboard
Install binvox
We use binvox to transform object mesh into particles. To use binvox, please download binvox from https://www.patrickmin.com/binvox/, put it under ./bin, and include it in your path with
export PATH=$PATH:$PWD/bin.
You might need to do chmod 777 binvox in order to execute the file.
Setup your own data path
open paths.yaml and write your own path there. You can set up different paths for different machines under different user name.
Preprocessing the Physion dataset
1) We need to convert the mesh scenes into particle scenes. This line will generate a separate folder (dpi_data_dir specified in paths.yaml) that holds data for the particle-based models
bash run_preprocessing_tdw_cheap.sh [SCENARIO_NAME] [MODE]
e.g., bash run_preprocessing_tdw_cheap.sh Dominoes train
SCENARIO_NAME can be one of the following: Dominoes, Collide, Support, Link, Contain, Roll, Drop, or Drape. MODE can be either train or test
You can visualize the original videos and the generated particle scenes with
python preprocessing_tdw_cheap.py --scenario Dominones --mode "train" --visualization 1
There will be videos generated under the folder vispy.
2) Then, try generate a train.txt and valid.txt files that indicates the trials you want to use for training and validaiton.
python create_train_valid.py
You can also design your specific split. Just put the trial names into one txt file.
3) For evalution on the red-hits-yellow prediciton, we can get the binary red-hits-yellow label txt file from the test dataset with
bash run_get_label_txt.sh [SCENARIO_NAME] test
This will generate a folder called labels under your output_folder dpi_data_dir. In the folder, each scenario will have a corresponding label file called [SCENARIO_NAME].txt
Training
Ok, now we are ready to start training the models.You can use the following command to train from scratch.
- Train GNS
bash scripts/train_gns.sh [SCENARIO_NAME] [GPU_ID]
SCENARIO_NAME can be one of the following: Dominoes, Collide, Support, Link, Contain, Roll, Drop and Drape.
- Train DPI
bash scripts/train_dpi.sh [SCENARIO_NAME] [GPU_ID]
Our implementation is different from the original DPI paper in 2 ways: (1) our model takes as inputs relative positions as opposed to absolute positions, (2) our model is trained with injected noise. These two features are suggested in the GNS paper, and we found them to be critcial for the models to generalize well to unseen scenes.
- Train with multiple scenarios
You can also train with more than one scenarios by adding different scenario to the argument dataf
python train.py --env TDWdominoes --model_name GNS --log_per_iter 1000 --training_fpt 3 --ckp_per_iter 5000 --floor_cheat 1 --dataf "Dominoes, Collide, Support, Link, Roll, Drop, Contain, Drape" --outf "all_gns"
- Visualize your training progress
Models and model logs are saved under [out_dir]/dump/dump_TDWdominoes.
You can visualize the training progress using tensorboard
tensorboard --logdir MODEL_NAME/log
Evaluation
- Evaluate GNS
bash scripts/eval_gns.sh [TRAIN_SCENARIO_NAME] [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
You can get the prediction txt file under eval/eval_TDWdominoes/[MODEL_NAME]
, e.g., test-Drape.txt
, which contains results of testing the model on the Drape scenario. You can visualize the results with additional argument --vis 1
.
- Evaluate GNS-Ransac
bash scripts/eval_gns_ransac.sh [TRAIN_SCENARIO_NAME] [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
- Evaluate DPI
bash scripts/eval_dpi.sh [TRAIN_SCENARIO_NAME] [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
- Evaluate Models trained on multiple scenario Here we provide some example of evaluating on arbitray models trained on all scenarios.
bash eval_all_gns.sh [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
bash eval_all_dpi.sh [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
bash eval_all_gns_ransac.sh [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
- Visualize trained Models Here we provide an example of visualizing the rollout results from trained arbitray models.
bash vis_gns.sh [EPOCH] [ITER] [Test SCENARIO_NAME] [GPU_ID]
You can find the visualization under eval/eval_TDWdominoes/[MODEL_NAME]/test-[Scenario].
We should see a gif for the original RGB videos, and another gif for the side-by-side comparison of gt particle scenes and the predicted particle scenes.
Citing Physion
If you find this codebase useful in your research, please consider citing:
@inproceedings{bear2021physion,
Title={Physion: Evaluating Physical Prediction from Vision in Humans and Machines},
author= {Daniel M. Bear and
Elias Wang and
Damian Mrowca and
Felix J. Binder and
Hsiao{-}Yu Fish Tung and
R. T. Pramod and
Cameron Holdaway and
Sirui Tao and
Kevin A. Smith and
Fan{-}Yun Sun and
Li Fei{-}Fei and
Nancy Kanwisher and
Joshua B. Tenenbaum and
Daniel L. K. Yamins and
Judith E. Fan},
url = {https://arxiv.org/abs/2106.08261},
archivePrefix = {arXiv},
eprint = {2106.08261},
Year = {2021}
}