Closed martenlienen closed 2 years ago
The error estimates of the dopri5 and tsit5 solvers do not consider the step width dt in the error estimate which leads to overestimates of the error (assuming dt < 1)
dt
dt < 1
https://github.com/DiffEqML/torchdyn/blob/master/torchdyn/numerics/solvers/ode.py#L155
https://github.com/DiffEqML/torchdyn/blob/master/torchdyn/numerics/solvers/ode.py#L177
Thanks! #165 addresses this. We observe a small speedup across our benchmarks.
The error estimates of the dopri5 and tsit5 solvers do not consider the step width
dt
in the error estimate which leads to overestimates of the error (assumingdt < 1
)https://github.com/DiffEqML/torchdyn/blob/master/torchdyn/numerics/solvers/ode.py#L155
https://github.com/DiffEqML/torchdyn/blob/master/torchdyn/numerics/solvers/ode.py#L177