PNODE: A PETSc-based Neural ODE Implementation

This repository provides a neural ordinary differential equation (ODE) implementation using PyTorch and PETSc's discrete adjoint ODE solvers (TSAdjoint). The framework allows easy access to PETSc's time stepping algorithms, adjoint solvers and checkpointing algorithms. Switching between different algorithms can be simply done with command line options at runtime. All the examples are fully supported to run on the GPU with single or double precision.

Installation

petsc4py needs to be installed first. The suggested way to install petsc4py is to install it from PETSc source

git clone https://gitlab.com/petsc/petsc.git
cd petsc
./configure PETSC_ARCH=arch-linux-opt --with-debugging=0 --download-petsc4py

Here minimal configure options are provided. Several other widely used options are listed below:

--with-cude=1 Use CUDA if an NVIDIA GPU is available.
--with-fc=0 Disable Fortran if you do not have a Fortran compiler or do not need Fortran.
--with-precision=single Build PETSc with single precision (without his option, PETSc uses double precision). Note that PETSC_ARCH takes of your choice. One can create many different versions of PETSc, each of which can have a different list of options and be given a different name specified by PETSC_ARCH.

After configure, follow the instructions on screen to do make all and make check (optional).

After petsc4py is installed, install PNODE

git clone https://github.com/caidao22/pnode.git
cd pnode
pip3 install .

Examples

Examples are provided in the examples directory. They include:

examples/ode_demo_petsc.py Simplest example that demonstrate how to fit a spiral ODE
examples/train-Cifar10.py Image classification on the CIFAR-10 dataset
ffjord-pnode/train_tabular.py Continuous normalizing flow on the datasets from MAF

Basic usage

This library provides one class ODEPetsc with three interface functions:

setupTS setup function for the solver
odeint_adjoint the main interface function that solves a neural ODE and calculates the gradient with discrete adjoint methods
odeint general-purpose ODE/DAE solver (without gradient calculation)

The API is similar to NODE and ANODE by design.

To solve an ODE in the form

du/dt = f(t, u)    u(t_0) = u_0,

One needs to create an ODEPetsc object

from pnode import petsc_adjoint
ode = petsc_adjoint.ODEPetsc()

and then set up the solver with

ode.setupTS(u0, func, step_size=step_size, implicit_form=implicit_form, use_dlpack=use_dlpack)

where func is any callable implementing the ordinary differential equation f(t, u), u0 is an any-D Tensor or a tuple of any-D Tensors representing the initial values.

ode.odeint_adjoint(u0, t)

where t is a 1-D Tensor containing the evaluation points and t[0] is the initial time.

odeint_adjoint performs a forward sweep using the PETSc TS solver and a backward sweep using the PETSc TSAdjoint solver.

Keyword Arguments in the setup function

step_size Step size for time integration.
method One of the solvers among 'euler', 'midpoint', 'rk4', 'dopri5_fixed'.
enable_adjoint Switch for enabling or disabling the adjoint sweep.

These are just a few options for convenience. More PETSc settings can be used at runtime with command line options. For example, -ts_type cn will choose the Crank-Nicolson methods for time integration, and -ts_type rk -ts_rk_type 4 will choose the classic RK4 solver. For more details, we suggest reading the PETSc manual and the examples included.

Checkpointing

PNODE offers a variety of optimal checkpointing strategies through the PETSc TSAdjoint library. For best efficiency, we suggest users to start from the following options

-ts_trajectory_type memory -ts_trajectory_solution_only 0

This will make PNODE store the solutions and stage values in DRAM at each time step in the forward sweep. If there is a limited memory budget, one can specify the maximum number of allowable checkpoints by adding an extra option -ts_trajectory_max_cps_ram <num_of_checkpoints>.

If you found this tool useful in your research, please consider citing.