A simple neural-network-driven method for recovering the coefficients of covariate-dependent transition-rate functions in Stochastic Petri Net (SPN) models directly from noisy, partially observed event trajectories, without ever computing an explicit likelihood.
Simulation is done in SPIKE (software by Chodak https://www-dssz.informatik.tu-cottbus.de/DSSZ/Software/Spike). The SPN model file is in a .andl format and is run with the .spc configuration file. All these setup is embedded in the Run_simulation.py file. The only things that need to be changed in the script are the file and folder paths. Everything else will run fine.
The entire simulation are run in parallel with 30 processes at a time. This ensured very efficient computations.
All source code for training and evaluating our neural surrogate model for dropout rates 0.1 and 0.2 are located in the model_src folder. First, create a python environment and install the required packages using the requirements.txt file. Use this command to run the model in terminal, python run_model.py for the default run. If you want to set your own arguments, use python run_model.py --epochs 100 --batch-size 16 --dropouts 0.1 0.3 0.5 --seed 42. For each dropout rate, the model_src contains the saved best model and scaler. One can call that directly without going through the full training procedure.
We also provide a jupyter notebook that contains our hyperparamter tuning procedure in our quest to find the best dropout rate that could produce the best uncertainty calibration and further assess multiple stochastic forward passes. Just open model_dp_tune.py run all to perform this procedure.
- Python 3.10 or newer
- This Spike version can only run on Windows
Python Packages:
numpy,pandas,matplotlibgeopandasscipysklearntorchjoblib