Momentum-frame Based 2D Physix
0. Prerequisites
From this directory, run below in your terminal:
$ mkdir solvers/data/model & mkdir solvers/data/gif_deeper & mkdir solvers/data/npy
And please make sure that you have python modules in "requirements.txt".
We do not recommend using "-r install", as it includes pytorch and taichi.
Please visit "https://docs.taichi.graphics/docs/" if you want to know about taichi for python.
1. Generate Ground-truth from Taichi Simulation
You can do it by running "solvers/1_generate_simulation_data.py".
Please run it in the local directory!
It will automatically generate momentum-frame dataset in .npy format.
You need at least 350 datasets to run the code without modifications.
2. Generate Pytorch Tensor Trainset/Testset
Please use "solvers/learning/2_generate_pt_dataset.ipynb".
If your .npy data are located correctly, it will save .pt datasets.
3. Training
Enjoy your training with "solvers/learning/3_training.ipynb".
In our experience, it will give satisfying results after about 15 epochs in the current setup.
We recommend you use RTX-3090 or other GPUs with larger memory sizes.
You may modify the "batch_size" hyperparameter if your GPU runs out of memory.
Please note that you may need to change learning rates also in that case.
4. Visualization
You can use "solvers/learning/4_visualization.ipynb".
It generates sequence images or gif files.
Please note that they use your testset, not the trainingset.