Run examples in float32 precision

pnkraemer / probdiffeq

Probabilistic solvers for differential equations in JAX. Adaptive ODE solvers with calibration, state-space model factorisations, and custom information operators. Compatible with the broader JAX scientific computing ecosystem.

https://pnkraemer.github.io/probdiffeq/

MIT License

32 stars 2 forks source link

Run examples in float32 precision #256

Open pnkraemer opened 1 year ago

pnkraemer commented 1 year ago

Not in float64 precision. This will also expose some potential issues in numerically stable implementations.

pnkraemer commented 1 year ago

At least the non-advanced examples that do not solve complicated problems.

pnkraemer commented 1 year ago

For future reference:

[x] Investigate all uses of jnp.sqrt(jnp.dot(...))
[ ] Investigate all uses of ** operations, e.g. l_obs**2
[ ] Change _adaptive.AdaptiveIVPSolver.numerical_zero to a value that depends on floating-point precision
[ ] Resolve #168 (which improves the debuggability of the "close-to-numerical-failure" cases)
[ ] Introduce a "success=True/False" flag in IVP solution (which improves the debuggability of the "close-to-numerical-failure" cases)