MeshGraphNetRP: Improving Generalization of GNN-based Cloth Simulation

• Paper: https://dl.acm.org/doi/10.1145/3623264.3624441
• Project: https://lava.kaist.ac.kr/?page_id=7246

File Locations

input

• download here: https://kaist.gov-dooray.com/share/drive-files/kkmh2qu2awzo.4fpYw3WzRKqYVxbJtkWNYg
• contains the gt_data
• gt_data folder includes the data gathered from Arcsim, it includes the training dataset as well as test dataset
• unity_demo includes different data for the arbitrary mode (refer to Runtime code). It includes the data of different unseen topologies

output

• download here: https://kaist.gov-dooray.com/share/drive-files/kkmh2qu2awzo.B0z0QN1PRZmbquL6csUVhw
• contains the basic_cloth, npy_results, and renders
• basic_cloth folder has the saved trained models of (1) final RNN-10 model (2) MLP encoder model and (3) original MeshGraphNets model
• npy_results contains the runtime results saved as npy. These npy files would be the input for the rendering
• renders contains the videos of the output

params

• these include the json files that describes the configurations of the different models
• it has the json file for (1) final RNN-10 model (2) MLP encoder model and (3) original MeshGraphNets model that could be used for training

Runtime code

• run the new_runtime_code.py for running the runtime
• it basically needs the name of the model (”json_name”) and the configuration of the runtime (”input_data”)
  • input_data
    • includes the mode, obj_code, and motion_code
    • mode
      • either “arbitrary” or “with_gt”
      • “arbitrary” is for configurations without gt
      • “with_gt” is for configurations with gt
        • refer to input>gt_data folder to see thek different ground truth data available
    • obj_code
      • the code for the cloth topology
      • for “with_gt” mode only “square_1024” as obj_code is available
      • for “arbitrary” refer to obj_name_to_handle_ind_list_dict variable in new_runtime_code.py for the list of different available obj_code
    • motion_code
      • the code for the motion
      • refer to input>gt_data folder to see the different motion available
        • specifically, the meta.json includes the motion code of these gt_data
      • for arbitrary motions, refer to make_input_dict_from_motion_code in generate_json_conf.py for the different available motion_code
        • example motion codes:
          • directional code: "fwd", "side", "updown", "xy", "yz", "xz", "xyz"
          • "bouncy"
          • "rotate_bounce"
          • "spiral"
          • "snake"
          • "spiral_upward"
• it would output an npy file to output>npy_results
• properly edit the "runtime" section of the json files if you want to run a different saved epoch or different model name

Rendering

• run render_plot.py to render the output>npy_results
• it needs (1) folder name in output>npy_results and (2) viewport:
• viewport could be either:
  • ‘def’ for default view
  • ‘side’
  • ‘front’
• configure the ffmpeg directory to run the code properly

Training

• run the train.py to train a model given a json file containing the configuration
• it saves the losses using wandb
  • change the wandb.login and wandb.init in train.py with your proper wandb credentials
• once you properly assign the wandb credentials, change the ‘is_save_output_on’ in the params folder json files from false to true
• to train simply run with the following format:
  • python train.py --params final_model
    • params is the name of the json_file
• the models would be saved to output>basic_cloth

Other

• please refer to environment.yml to see the different libraries used for the python environment

Acknowledgments

• This repository contains pieces of code from the following repository:
• PyTorch version of Learning Mesh-Based Simulation with Graph Networks (ICLR 2021)