Variational inference for nonlinear ordinary differential equations

This repository contains code supporting the above paper.

Dependencies

Generic scientific Python stack: numpy, scipy, matplotlib, seaborn and (optionally, for automatic Jacobians) sympy. Also rquires PyTorch.

For Pyro read: https://pyro.ai/ (we recommend installing from the source)

Usage

To run the SIR model inference, using forward sensitivity simply use the following command: python SIR_example.py --num_samples 1000 --warmup_steps 500 --iterations 10000 --num_qsamples 1000. The ProteinTransduction_example.py file can be run with similar arguments.

To run with adjoint sensitivity use:

python SIR_example.py --num_samples 1000 --warmup_steps 500 --iterations 10000 --num_qsamples 1000 --adjoint True

To run the lotka-Volterra model with ABC-SMC run within CPP directory:

python setup.py build_ext -i. Then rename the generated *.so file to lvssa.so and run LNA_abcsmc.py. Run LNA_variational.py for VI with LNA.

Using PyTorch's VJP.

All the examples in the paper were run using SymPy Lamdify function. To use PyTorch's VJP see the SIR_torch_jacobians.py script.