search

ICB-DCM / pyPESTO

python Parameter EStimation TOolbox
https://pypesto.readthedocs.io
BSD 3-Clause "New" or "Revised" License
219 stars 47 forks source link

Support `sensi_orders` in gradient checks #1487

Open dilpath opened 3 weeks ago

dilpath commented 3 weeks ago

While debugging another model, the issue arose again that sometimes Objective(x, sensi_orders=(0,)) provides a different fval than Objective(x, sensi_orders=(0,1)). Tolerances etc. can be adjusted until these two are similar; however, pyPESTO's check_grad might incorrectly report that gradients are not close, because it computes fval with sensis https://github.com/ICB-DCM/pyPESTO/blob/ead29b36db80d2eeee91b5d29bfbf4430137f8aa/pypesto/objective/base.py#L535

but fval_p and fval_m (for the plus/minus steps used in finite difference approximations) without sensis https://github.com/ICB-DCM/pyPESTO/blob/ead29b36db80d2eeee91b5d29bfbf4430137f8aa/pypesto/objective/base.py#L556 https://github.com/ICB-DCM/pyPESTO/blob/ead29b36db80d2eeee91b5d29bfbf4430137f8aa/pypesto/objective/base.py#L561

This could be resolved by using the same sensi_orders in all function calls, or otherwise providing some way for the user to specify this.

dweindl commented 3 weeks ago

This could be resolved by using the same sensi_orders in all function calls, or otherwise providing some way for the user to specify this.

It might make sense to address that independently of gradient checks, since this will also be a problem for any optimizer that calls the objective with varying sensi_orders. Maybe a simple wrapper around Objective that always passes a certain sensi_orders internally and then filters the result to match the externally requested sensi_orders.