PIML

The official PyTorch implementation of "Physics-infused Machine Learning for Crowd Simulation" (KDD'22)

To run the training, fine-tuning, and testing process and save the model, just run

cd src
python main.py

The result will be saved in the saved_model folder. If you want to change the configuration, see the get_args() function in src/main.py for the available options. Here are some important options:

• --finetune_flag to conduct the finetuning process after training.
• --data_config to set the train, valid, and test dataset for training.
• --ft_data_config to set the train, valid, and test dataset for finetuning.

Other Scripts

• src/symbolic_regression.py to conduct symbolic regression for the discovered MLAPM model.
• src/main_mlapm.py to run the simulation with the discovered MLAPM model.
• src/run_experiments.py to conduct grid searching for hyperparameters.

Requirements

python            3.12.2
matplotlib        3.8.4
numpy             1.26.4
PyYAML            6.0.1
setproctitle      1.3.3
torch             2.2.2
tqdm              4.66.2

About the Data

• GC Dataset: ./data/GC_Dataset/*.npy, e.g., GC_Dataset_ped1-12685_time1000-1060_interp9_xrange5-25_yrange15-35.npy
  • ped1-12685: Select all 12685 pedestrians.
  • time1000-1060: Use the time range from 1000s to 1060s, duration=1min.
  • interp9: Interplate 9 extra steps between time steps, such that $\Delta t = 0.08s$.
  • xrange5-25_yrange15-35: Use the region where $5m \leq x \leq 25m, 15, \leq y \leq 35m$.
• UCY Dataset: ./data/UCY_Dataset/*.npy, e.g., UCY_Dataset_time0-54_timeunit0.08.npy
  • time0-54: Use the time range from the initial to 54s
  • timeunit0.08: Inteplate along the time axis such that $\Delta t = 0.08s$.
• Synthetic Dataset: ./data/synthetic_data/*.npy, e.g., GC_Dataset_ped1-12685_time1560-1620_interp9_xrange5-25_yrange15-35_simulation.npy
  • GC_Dataset_ped1-12685_time1560-1620_interp9_xrange5-25_yrange15-35: Refer to the basic scenario.
  • `simulation: Simulate the basic scenario, where the number of pedestrians and their appearance time, initial position, and destination are the same as in the basic scenario.

For each *.npy file, the data is stored in the following format:

(metadata, trajectories, destinations, obstacles) = np.load(file_path, allow_pickle=True)

And here is the description of each element:

• metadata: A dictionary that contains the metadata of the dataset.
• trajectories: A list of trajectories. For each trajectory, np.array(trajectory) is a np.array with shape (T, 3), where T is the number of time steps where the pedestrian appears in the scenario, and each row is the x-position, y-position and time of that pedestrian.
• destinations: np.array(destinations) is an np.array with shape (N, 1, 3), where N is the number of pedestrians. np.array(destinations)[:, 0, :2] is the xy-position of the destination, while np.array(destinations)[:, 0, 2] is the time they reach the destination.
• obstacles: An np.array with shape (M, 2), where M is the number of obstacle points, and each row is the xy-position of the obstacle.