juliusberner / deep_kolmogorov

Numerically Solving Parametric Families of High-Dimensional Kolmogorov Partial Differential Equations via Deep Learning (NeurIPS 2020)

Numerically Solving Parametric Families of High-Dimensional Kolmogorov Partial Differential Equations via Deep Learning

Accompanying code for NeurIPS 2020 paper (Poster). Deep Learning based algorithm for solving a parametrized family of high-dimensional Kolmogorov PDEs. Implemented in PyTorch and Tune.

Reproducing the Experiments

To run the experiments and visualize the results open the jupyter notebook experiments.ipynb. For reproducibility we recommend to use the docker container defined by Dockerfile (see Docker Tutorial).

Our setup:

• DGX-1 server
• Ubuntu 18.04.3, Python 3.6.9, Torch 1.5 (as given by the NVIDIA-Docker with base image nvcr.io/nvidia/pytorch:20.03-py3)
• additional requirements as specified by requirements.txt

Run experiments:

Experiment (and reference in the paper)	Command (adapt --gpus if necessary)
Black-Scholes model (Table 1, Fig. 3,4,5,6,7,8)	python main.py --mode=avg_bs --gpus=2
Heat-equation with paraboloid initial condition (Table 3)	python main.py --mode=avg_heat_paraboloid --gpus=2
Heat-equation with Gaussian initial condition (Table 4)	python main.py --mode=avg_heat_gaussian --gpus=2
Basket put option (Table 2)	python main.py --mode=avg_basket --gpus=4
Cost vs. input dimension (Fig. 9)	python main.py --mode=dims_heat_paraboloid --gpus=2
Ablation study Black-Scholes model (Table 7)	python main.py --mode=compare_nets_bs --gpus=2
Ablation study heat-equation (Table 8)	python main.py --mode=compare_nets_heat --gpus=2
Hyperparameter search (Table 6)	python main.py --mode=optimize_bs --gpus=2

Visualize results:

1. Jupyter notebook: Open the notebook experiments.ipynb and run section Analyze experiments
2. Tensorboard: Run tensorboard --logdir exp or tensorboard --logdir exp/experiment_xyz