Bridges DifferentialEquations.jl with PyTorch. Besides benefitting from the huge range of solvers available in DifferentialEquations.jl, this allows taking gradients through solvers using local sensitivity analysis/auto-diff. The package has only been tested with ODE problems, and in particular, automatic differentiation is only supported for ODEs using ForwardDiff.jl. This can be extended in the future, contributions are welcome.
- Simple ODE problem to demonstrate the interface and confirm gradients with analytical solution
- SIR model for a slighlty more complicated model with numerical gradient checking
- Hodgkin-Huxley model for a realistic example from Neuroscience
Prerequisites for using diffeqtorch are installation of Julia and Python. Note that the binary directory of julia needs to be in your PATH.
Install diffeqtorch:
$ pip install diffeqtorch
$ export JULIA_SYSIMAGE_DIFFEQTORCH="$HOME/.julia_sysimage_diffeqtorch.so"
$ python -c "from diffeqtorch.install import install_and_test; install_and_test()"
We recommend using a custom Julia system image containing dependencies. By setting the environment variable JULIA_SYSIMAGE_DIFFEQTORCH, an image will be created and used automatically. This may take a while but will improve speed afterwards.
from diffeqtorch import DiffEq
import torch
f = """
function f(du,u,p,t)
du[1] = p[1] * u[1]
end
"""
de = DiffEq(f)
u0 = torch.tensor([1.])
tspan = torch.tensor([0., 3.])
p = torch.tensor([1.01])
u, t = de(u0, tspan, p)See also help(DiffEq) and examples provided in notebooks/.
MIT