Reverse mode and exact hessians#22
Open
rafaqz wants to merge 8 commits into
Open
Conversation
- Drop Optimization.jl / OptimizationIpopt.jl / SciMLBase in favour of the direct Ipopt.jl C interface. Gives access to the primal+dual warm-start API (IpoptProblem.x / mult_g) that the SciMLBase wrappers do not expose, so consecutive solves at similar parameter settings reuse the previous solution as the new initial guess. - Export IPOPTSolverCache so callers can construct the cache once and pass it across sweeps. - ipopt.jl rewritten end-to-end: cached buffers for primals, duals, bounds, constraint values, and Enzyme.forward-over-reverse Hessian (replaces FiniteDiff). - Adopt HeatExchange Air/Water Fluid singletons in examples and test data: example_environment_pars defaults to Air(); endotherm test now maps the FLTYPE input flag to Water()/Air(). - Add bench_ipopt_sweep.jl and hessian_validation.jl examples for benchmarking warm-start speedup and Hessian correctness. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
mrke
approved these changes
May 14, 2026
Contributor
mrke
left a comment
mrke
left a comment
There was a problem hiding this comment.
Heavy stuff - I looked at all the code but probably need to spend a full week digesting how all this works.
Member
Author
|
I can probably split it up if that makes it easier to read? |
rafaqz
commented
May 18, 2026
| log_heat_flow_max, | ||
| flesh_conductivity_max, panting_rate_max, skin_wetness_max, | ||
| insulation_depth_max, aspect_ratio_max] | ||
| lb[1] = core_temperature_min; lb[2] = skin_temperature_min; lb[3] = skin_temperature_min |
Member
Author
There was a problem hiding this comment.
Note that this code is pretty bad. We only need it because we dont have the generalised (e.g. ModelParameters) way to get these thing into a vector.
This is a hack to get it working because its pretty hard to do that!
Later we can generalise it.
rafaqz
commented
May 18, 2026
| # through `_*_inputs!` to fill `cache.x_scaling`. Worth doing the next time | ||
| # we touch the trait schema. | ||
| # ────────────────────────────────────────────────────────────────────────── | ||
| function _scaling_weighted() |
Member
Author
There was a problem hiding this comment.
Again, as the comment mentions, this code sucks
Contributor
There was a problem hiding this comment.
Now that the trait examples are shifted into HeatExchange.jl we might be closer to revisiting the trait schema. I've been thinking about the whole trait database to parameters set up too - have some ideas.
Contributor
|
Spell out short identifiers in the IPOPT thermoregulation code: lb/ub/u0,
lcons/ucons, metab_pars/int_cond/evap_pars/opt_pars/nlp_pars, M/sT/iT
init shorthand, jac_*/hess_*/lag_* AD buffers, eval_f/eval_g/eval_grad_f/
eval_jac_g/eval_h closures, and the corresponding type parameters.
Drop the mutable-struct on IPOPTSolverCache: the only mutating field was
the warm-start flag, now stored as Base.RefValue{Bool}.
…o functors - `thermoregulate` and `IPOPTSolverCache` take a single `init` NamedTuple (`metabolic_heat_flow`, `skin_temperature`, `insulation_temperature`) instead of three loose positional args, threaded through the rule-based loop, IPOPT builder, `_inputs!`, and `_refresh!`. - Rename `_IPOPTState` to `IPOPTParameters` and its instance field to `ipopt_parameters` — what it holds is per-solve problem parameters, not optimizer state. The mutable cell is just so the callbacks can be reused across solves. - Replace the anonymous Ipopt callback closures with six named functor structs (`EvaluateObjective`, `EvaluateConstraints`, `EvaluateObjectiveGradient`, `EvaluateConstraintJacobian`, `EvaluateHessian`, `LagrangianGradient`), all subtypes of `Function` so `Ipopt.IpoptProblem`'s typed callback fields accept them. - Collapse `_build_cache` into two layered `IPOPTSolverCache` constructors and factor out `_problem_size` / `_constraint_bounds` so the strategy-specific dimensions live in one place. - Deduplicate `_inputs!`: one shared body computes all scalars and parameter NamedTuples; the strategy-specific index layout is delegated to a small `_write_layout!` dispatched on packed type. - Drop the `effectors[i]` magic numbers in `_objective_value` — name only the variables each method actually uses. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR adds proper reverse mode in Enzyme and exact hessians, and does some perfreoamance optimisation. Should be 4-5x faster.
We remove the OptimiastionIpopt package and just use Ipopt.jl directly, as OptimisationIpopt kind of just gets in the way.