discopt 0.2.4__tar.gz → 0.2.5__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {discopt-0.2.4 → discopt-0.2.5}/Cargo.lock +2 -2
- {discopt-0.2.4 → discopt-0.2.5}/Cargo.toml +1 -1
- {discopt-0.2.4 → discopt-0.2.5}/PKG-INFO +2 -1
- {discopt-0.2.4 → discopt-0.2.5}/README.md +1 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/bnb/branching.rs +8 -2
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/bnb/tree_manager.rs +5 -1
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/src/ripopt_bindings.rs +133 -55
- {discopt-0.2.4 → discopt-0.2.5}/pyproject.toml +2 -1
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/__init__.py +1 -1
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/ipm.py +12 -4
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/lp_ipm.py +54 -25
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/nl_evaluator.py +93 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/nlp_evaluator.py +130 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/problem_classifier.py +2 -2
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/qp_ipm.py +15 -19
- discopt-0.2.5/python/discopt/_jax/scaling.py +98 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/cli.py +69 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/interfaces/cutest.py +57 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/modeling/core.py +24 -5
- discopt-0.2.5/python/discopt/ro/ROADMAP.md +102 -0
- discopt-0.2.5/python/discopt/ro/__init__.py +83 -0
- discopt-0.2.5/python/discopt/ro/affine_policy.py +435 -0
- discopt-0.2.5/python/discopt/ro/counterpart.py +188 -0
- discopt-0.2.5/python/discopt/ro/formulations/__init__.py +11 -0
- discopt-0.2.5/python/discopt/ro/formulations/_common.py +273 -0
- discopt-0.2.5/python/discopt/ro/formulations/box.py +388 -0
- discopt-0.2.5/python/discopt/ro/formulations/ellipsoidal.py +240 -0
- discopt-0.2.5/python/discopt/ro/formulations/polyhedral.py +278 -0
- discopt-0.2.5/python/discopt/ro/uncertainty.py +314 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solver.py +674 -11
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/nlp_ipopt.py +11 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/qp_highs.py +8 -8
- {discopt-0.2.4 → discopt-0.2.5}/LICENSE +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/Cargo.toml +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/bnb/mod.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/bnb/node.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/bnb/pool.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/expr.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/lib.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/nl_parser.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/presolve/fbbt.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/presolve/mod.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/presolve/obbt.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/presolve/probing.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-core/src/presolve/simplify.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/Cargo.toml +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/src/batch.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/src/bnb_bindings.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/src/expr_bindings.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/src/lib.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/crates/discopt-python/src/nl_bindings.rs +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/PHASE3_MATH_NOTES.md +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/alphabb.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/batch_evaluator.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/convexity.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/cutting_planes.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/dag_compiler.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/differentiable.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/differentiable_lp.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/differentiable_qp.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/differentiable_solve.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/envelopes.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/gdp_reformulate.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/gnn_branching.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/gnn_policy.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/icnn.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/icnn_trainer.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/ipm_callbacks.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/ipm_iterative.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/learned_relaxations.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/mccormick.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/mccormick_nlp.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/nl_reconstruction.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/obbt.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/pcg.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/piecewise_mccormick.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/primal_heuristics.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/problem_graph.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/relaxation_compiler.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/sparse_ipm.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/sparse_jacobian.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/sparse_kkt.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/sparsity.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/_jax/strong_branching.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/benchmarks/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/benchmarks/metrics.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/benchmarks/runner.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/callbacks.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/constants.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/dae/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/dae/collocation.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/dae/finite_difference.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/dae/mol.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/dae/polynomials.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/export/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/export/_extract.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/export/lp.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/export/mps.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/interfaces/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/advisor.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/chat.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/commentary.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/diagnosis.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/prompts.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/provider.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/reformulation.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/safety.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/serializer.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/llm/tools.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/modeling/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/modeling/examples.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/bounds.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/formulations/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/formulations/base.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/formulations/full_space.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/formulations/reduced_space.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/formulations/relu_bigm.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/formulations/tree_ensemble.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/network.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/predictor.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/readers/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/readers/onnx_reader.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/readers/sklearn_reader.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/readers/torch_reader.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/scaling.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/nn/tree.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/__init__.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/gdpopt_loa.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/lp_highs.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/milp_highs.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/nlp_ripopt.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/oa.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/solvers/sipopt.py +0 -0
- {discopt-0.2.4 → discopt-0.2.5}/python/discopt/warm_start.py +0 -0
|
@@ -1067,9 +1067,9 @@ checksum = "a96887878f22d7bad8a3b6dc5b7440e0ada9a245242924394987b21cf2210a4c"
|
|
|
1067
1067
|
|
|
1068
1068
|
[[package]]
|
|
1069
1069
|
name = "ripopt"
|
|
1070
|
-
version = "0.6.
|
|
1070
|
+
version = "0.6.1"
|
|
1071
1071
|
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
1072
|
-
checksum = "
|
|
1072
|
+
checksum = "09db916a3dd0c64c14c2952b377296157b537f80da3753758e4a46a59f255b87"
|
|
1073
1073
|
dependencies = [
|
|
1074
1074
|
"env_logger",
|
|
1075
1075
|
"faer",
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: discopt
|
|
3
|
-
Version: 0.2.
|
|
3
|
+
Version: 0.2.5
|
|
4
4
|
Classifier: Development Status :: 3 - Alpha
|
|
5
5
|
Classifier: Intended Audience :: Science/Research
|
|
6
6
|
Classifier: License :: OSI Approved :: Eclipse Public License 2.0 (EPL-2.0)
|
|
@@ -61,6 +61,7 @@ Project-URL: Repository, https://github.com/jkitchin/discopt
|
|
|
61
61
|
|
|
62
62
|
# discopt
|
|
63
63
|
|
|
64
|
+
[](https://pypi.org/project/discopt/)
|
|
64
65
|
[](https://github.com/jkitchin/discopt/actions/workflows/ci.yml)
|
|
65
66
|
[](https://codecov.io/gh/jkitchin/discopt)
|
|
66
67
|
|
|
@@ -1,5 +1,6 @@
|
|
|
1
1
|
# discopt
|
|
2
2
|
|
|
3
|
+
[](https://pypi.org/project/discopt/)
|
|
3
4
|
[](https://github.com/jkitchin/discopt/actions/workflows/ci.yml)
|
|
4
5
|
[](https://codecov.io/gh/jkitchin/discopt)
|
|
5
6
|
|
|
@@ -287,6 +287,12 @@ pub fn create_children(
|
|
|
287
287
|
// Warm-start: pass parent's stored solution to children.
|
|
288
288
|
let parent_sol = parent.parent_solution.clone();
|
|
289
289
|
|
|
290
|
+
// Inherit parent's lower bound: the child's feasible region is a subset
|
|
291
|
+
// of the parent's, so the parent's lower bound is valid for the child.
|
|
292
|
+
// This prevents the global lower bound from dropping to -inf when
|
|
293
|
+
// unsolved children enter the pool.
|
|
294
|
+
let inherited_lb = parent.local_lower_bound;
|
|
295
|
+
|
|
290
296
|
// Left child: x_i <= floor(val)
|
|
291
297
|
let mut left_ub = parent.ub.clone();
|
|
292
298
|
left_ub[idx] = bp; // bp is already floor(val)
|
|
@@ -296,7 +302,7 @@ pub fn create_children(
|
|
|
296
302
|
depth: parent.depth + 1,
|
|
297
303
|
lb: parent.lb.clone(),
|
|
298
304
|
ub: left_ub,
|
|
299
|
-
local_lower_bound:
|
|
305
|
+
local_lower_bound: inherited_lb,
|
|
300
306
|
status: NodeStatus::Pending,
|
|
301
307
|
parent_solution: parent_sol.clone(),
|
|
302
308
|
};
|
|
@@ -310,7 +316,7 @@ pub fn create_children(
|
|
|
310
316
|
depth: parent.depth + 1,
|
|
311
317
|
lb: right_lb,
|
|
312
318
|
ub: parent.ub.clone(),
|
|
313
|
-
local_lower_bound:
|
|
319
|
+
local_lower_bound: inherited_lb,
|
|
314
320
|
status: NodeStatus::Pending,
|
|
315
321
|
parent_solution: parent_sol,
|
|
316
322
|
};
|
|
@@ -395,7 +395,11 @@ impl TreeManager {
|
|
|
395
395
|
return f64::INFINITY;
|
|
396
396
|
}
|
|
397
397
|
let denom = self.incumbent_value.abs().max(1.0);
|
|
398
|
-
(self.incumbent_value - self.global_lower_bound) / denom
|
|
398
|
+
let raw = (self.incumbent_value - self.global_lower_bound) / denom;
|
|
399
|
+
// Clamp to zero: a negative gap means the bound exceeds the
|
|
400
|
+
// incumbent due to numerical noise or suboptimal NLP solves.
|
|
401
|
+
// Don't report it as closed; keep exploring.
|
|
402
|
+
raw.max(0.0)
|
|
399
403
|
}
|
|
400
404
|
|
|
401
405
|
/// Get aggregate tree statistics.
|
|
@@ -11,6 +11,9 @@ use pyo3::types::PyDict;
|
|
|
11
11
|
///
|
|
12
12
|
/// Each callback acquires the GIL and calls back into the Python evaluator.
|
|
13
13
|
/// Bounds, structure, and initial point are cached as Rust-owned data.
|
|
14
|
+
/// When the evaluator provides sparse COO structure (via `has_sparse_structure`),
|
|
15
|
+
/// uses `evaluate_jacobian_values` / `evaluate_hessian_values` for efficient
|
|
16
|
+
/// value extraction. Otherwise falls back to dense evaluation.
|
|
14
17
|
struct PyNlpProblem {
|
|
15
18
|
evaluator: PyObject,
|
|
16
19
|
n: usize,
|
|
@@ -20,12 +23,15 @@ struct PyNlpProblem {
|
|
|
20
23
|
g_l: Vec<f64>,
|
|
21
24
|
g_u: Vec<f64>,
|
|
22
25
|
x0: Vec<f64>,
|
|
23
|
-
/// Cached Jacobian sparsity: (rows, cols)
|
|
26
|
+
/// Cached Jacobian sparsity: COO (rows, cols)
|
|
24
27
|
jac_rows: Vec<usize>,
|
|
25
28
|
jac_cols: Vec<usize>,
|
|
26
|
-
/// Cached Hessian sparsity: (rows, cols)
|
|
29
|
+
/// Cached Hessian sparsity: COO (rows, cols), lower triangle
|
|
27
30
|
hess_rows: Vec<usize>,
|
|
28
31
|
hess_cols: Vec<usize>,
|
|
32
|
+
/// Whether to use sparse value callbacks
|
|
33
|
+
use_sparse_jac: bool,
|
|
34
|
+
use_sparse_hess: bool,
|
|
29
35
|
}
|
|
30
36
|
|
|
31
37
|
impl ripopt::NlpProblem for PyNlpProblem {
|
|
@@ -107,19 +113,33 @@ impl ripopt::NlpProblem for PyNlpProblem {
|
|
|
107
113
|
}
|
|
108
114
|
Python::with_gil(|py| {
|
|
109
115
|
let x_py = PyArray1::from_slice(py, x);
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
116
|
+
if self.use_sparse_jac {
|
|
117
|
+
// Sparse path: evaluator returns 1-D values in COO order
|
|
118
|
+
let result = self
|
|
119
|
+
.evaluator
|
|
120
|
+
.call_method1(py, "evaluate_jacobian_values", (x_py,))
|
|
121
|
+
.expect("evaluate_jacobian_values failed");
|
|
122
|
+
let arr: PyReadonlyArray1<f64> = result
|
|
123
|
+
.bind(py)
|
|
124
|
+
.extract()
|
|
125
|
+
.expect("jacobian_values not ndarray");
|
|
126
|
+
let slice = arr.as_slice().expect("jacobian_values not contiguous");
|
|
127
|
+
values.copy_from_slice(slice);
|
|
128
|
+
} else {
|
|
129
|
+
// Dense path: flatten (m, n) Jacobian row-major
|
|
130
|
+
let result = self
|
|
131
|
+
.evaluator
|
|
132
|
+
.call_method1(py, "evaluate_jacobian", (x_py,))
|
|
133
|
+
.expect("evaluate_jacobian failed");
|
|
134
|
+
let arr: PyReadonlyArray1<f64> = result
|
|
135
|
+
.call_method0(py, "flatten")
|
|
136
|
+
.expect("flatten failed")
|
|
137
|
+
.bind(py)
|
|
138
|
+
.extract()
|
|
139
|
+
.expect("jacobian not ndarray");
|
|
140
|
+
let slice = arr.as_slice().expect("jacobian not contiguous");
|
|
141
|
+
values.copy_from_slice(slice);
|
|
142
|
+
}
|
|
123
143
|
})
|
|
124
144
|
}
|
|
125
145
|
|
|
@@ -132,30 +152,47 @@ impl ripopt::NlpProblem for PyNlpProblem {
|
|
|
132
152
|
let x_py = PyArray1::from_slice(py, x);
|
|
133
153
|
let lambda_py = PyArray1::from_slice(py, lambda);
|
|
134
154
|
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
155
|
+
if self.use_sparse_hess {
|
|
156
|
+
// Sparse path: evaluator returns 1-D values in COO order
|
|
157
|
+
let result = self
|
|
158
|
+
.evaluator
|
|
159
|
+
.call_method1(
|
|
160
|
+
py,
|
|
161
|
+
"evaluate_hessian_values",
|
|
162
|
+
(x_py, obj_factor, lambda_py),
|
|
163
|
+
)
|
|
164
|
+
.expect("evaluate_hessian_values failed");
|
|
165
|
+
let arr: PyReadonlyArray1<f64> = result
|
|
166
|
+
.bind(py)
|
|
167
|
+
.extract()
|
|
168
|
+
.expect("hessian_values not ndarray");
|
|
169
|
+
let slice = arr.as_slice().expect("hessian_values not contiguous");
|
|
170
|
+
values.copy_from_slice(slice);
|
|
171
|
+
} else {
|
|
172
|
+
// Dense path: extract lower triangle from (n, n) array
|
|
173
|
+
let result = self
|
|
174
|
+
.evaluator
|
|
175
|
+
.call_method1(
|
|
176
|
+
py,
|
|
177
|
+
"evaluate_lagrangian_hessian",
|
|
178
|
+
(x_py, obj_factor, lambda_py),
|
|
179
|
+
)
|
|
180
|
+
.expect("evaluate_lagrangian_hessian failed");
|
|
143
181
|
|
|
144
|
-
|
|
145
|
-
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
.
|
|
151
|
-
|
|
152
|
-
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
idx += 1;
|
|
182
|
+
let flat: PyReadonlyArray1<f64> = result
|
|
183
|
+
.call_method0(py, "flatten")
|
|
184
|
+
.expect("flatten failed")
|
|
185
|
+
.bind(py)
|
|
186
|
+
.extract()
|
|
187
|
+
.expect("hessian not ndarray");
|
|
188
|
+
let data = flat.as_slice().expect("hessian not contiguous");
|
|
189
|
+
let n = self.n;
|
|
190
|
+
let mut idx = 0;
|
|
191
|
+
for i in 0..n {
|
|
192
|
+
for j in 0..=i {
|
|
193
|
+
values[idx] = data[i * n + j];
|
|
194
|
+
idx += 1;
|
|
195
|
+
}
|
|
159
196
|
}
|
|
160
197
|
}
|
|
161
198
|
})
|
|
@@ -209,26 +246,65 @@ pub fn solve_ripopt(
|
|
|
209
246
|
let n = x0_vec.len();
|
|
210
247
|
let m = g_l_vec.len();
|
|
211
248
|
|
|
212
|
-
//
|
|
213
|
-
let
|
|
214
|
-
|
|
215
|
-
|
|
216
|
-
|
|
217
|
-
|
|
218
|
-
|
|
249
|
+
// Check if evaluator provides sparse COO structure
|
|
250
|
+
let use_sparse = evaluator
|
|
251
|
+
.call_method0(py, "has_sparse_structure")
|
|
252
|
+
.ok()
|
|
253
|
+
.and_then(|r| r.extract::<bool>(py).ok())
|
|
254
|
+
.unwrap_or(false);
|
|
255
|
+
|
|
256
|
+
let (jac_rows, jac_cols, use_sparse_jac) = if use_sparse && m > 0 {
|
|
257
|
+
let structure = evaluator
|
|
258
|
+
.call_method0(py, "jacobian_structure")
|
|
259
|
+
.expect("jacobian_structure failed");
|
|
260
|
+
let tuple = structure.bind(py).downcast::<pyo3::types::PyTuple>()
|
|
261
|
+
.expect("jacobian_structure must return tuple");
|
|
262
|
+
let rows: Vec<usize> = tuple
|
|
263
|
+
.get_item(0).expect("missing rows")
|
|
264
|
+
.extract().expect("rows not array of int");
|
|
265
|
+
let cols: Vec<usize> = tuple
|
|
266
|
+
.get_item(1).expect("missing cols")
|
|
267
|
+
.extract().expect("cols not array of int");
|
|
268
|
+
(rows, cols, true)
|
|
269
|
+
} else {
|
|
270
|
+
// Dense fallback: all m*n entries, row-major
|
|
271
|
+
let mut rows = Vec::with_capacity(m * n);
|
|
272
|
+
let mut cols = Vec::with_capacity(m * n);
|
|
273
|
+
for i in 0..m {
|
|
274
|
+
for j in 0..n {
|
|
275
|
+
rows.push(i);
|
|
276
|
+
cols.push(j);
|
|
277
|
+
}
|
|
219
278
|
}
|
|
220
|
-
|
|
279
|
+
(rows, cols, false)
|
|
280
|
+
};
|
|
221
281
|
|
|
222
|
-
|
|
223
|
-
|
|
224
|
-
|
|
225
|
-
|
|
226
|
-
|
|
227
|
-
|
|
228
|
-
|
|
229
|
-
|
|
282
|
+
let (hess_rows, hess_cols, use_sparse_hess) = if use_sparse {
|
|
283
|
+
let structure = evaluator
|
|
284
|
+
.call_method0(py, "hessian_structure")
|
|
285
|
+
.expect("hessian_structure failed");
|
|
286
|
+
let tuple = structure.bind(py).downcast::<pyo3::types::PyTuple>()
|
|
287
|
+
.expect("hessian_structure must return tuple");
|
|
288
|
+
let rows: Vec<usize> = tuple
|
|
289
|
+
.get_item(0).expect("missing rows")
|
|
290
|
+
.extract().expect("rows not array of int");
|
|
291
|
+
let cols: Vec<usize> = tuple
|
|
292
|
+
.get_item(1).expect("missing cols")
|
|
293
|
+
.extract().expect("cols not array of int");
|
|
294
|
+
(rows, cols, true)
|
|
295
|
+
} else {
|
|
296
|
+
// Dense fallback: full lower triangle
|
|
297
|
+
let hess_nnz = n * (n + 1) / 2;
|
|
298
|
+
let mut rows = Vec::with_capacity(hess_nnz);
|
|
299
|
+
let mut cols = Vec::with_capacity(hess_nnz);
|
|
300
|
+
for i in 0..n {
|
|
301
|
+
for j in 0..=i {
|
|
302
|
+
rows.push(i);
|
|
303
|
+
cols.push(j);
|
|
304
|
+
}
|
|
230
305
|
}
|
|
231
|
-
|
|
306
|
+
(rows, cols, false)
|
|
307
|
+
};
|
|
232
308
|
|
|
233
309
|
let problem = PyNlpProblem {
|
|
234
310
|
evaluator,
|
|
@@ -243,6 +319,8 @@ pub fn solve_ripopt(
|
|
|
243
319
|
jac_cols,
|
|
244
320
|
hess_rows,
|
|
245
321
|
hess_cols,
|
|
322
|
+
use_sparse_jac,
|
|
323
|
+
use_sparse_hess,
|
|
246
324
|
};
|
|
247
325
|
|
|
248
326
|
// Build SolverOptions from the Python dict
|
|
@@ -4,7 +4,7 @@ build-backend = "maturin"
|
|
|
4
4
|
|
|
5
5
|
[project]
|
|
6
6
|
name = "discopt"
|
|
7
|
-
version = "0.2.
|
|
7
|
+
version = "0.2.5"
|
|
8
8
|
description = "Hybrid MINLP solver combining Rust and JAX"
|
|
9
9
|
requires-python = ">=3.10"
|
|
10
10
|
authors = [{name = "John Kitchin"}]
|
|
@@ -69,6 +69,7 @@ include = [
|
|
|
69
69
|
[tool.ruff]
|
|
70
70
|
line-length = 100
|
|
71
71
|
target-version = "py310"
|
|
72
|
+
extend-exclude = ["adversary/"]
|
|
72
73
|
|
|
73
74
|
[tool.ruff.lint]
|
|
74
75
|
select = ["E", "F", "I", "W"]
|
|
@@ -1040,6 +1040,11 @@ def solve_nlp_ipm(
|
|
|
1040
1040
|
if fields:
|
|
1041
1041
|
ipm_opts = ipm_opts._replace(**fields)
|
|
1042
1042
|
|
|
1043
|
+
# Enforce max_wall_time by capping max_iter. The JIT-compiled
|
|
1044
|
+
# jax.lax.while_loop cannot check wall clock, so we limit iterations
|
|
1045
|
+
# upfront and verify after the solve (issue #5).
|
|
1046
|
+
max_wall_time = options.get("max_wall_time") if options else None
|
|
1047
|
+
|
|
1043
1048
|
obj_fn = evaluator._obj_fn
|
|
1044
1049
|
con_fn = evaluator._cons_fn if m > 0 else None
|
|
1045
1050
|
x0_jax = jnp.array(x0, dtype=jnp.float64)
|
|
@@ -1049,12 +1054,15 @@ def solve_nlp_ipm(
|
|
|
1049
1054
|
wall_time = time.perf_counter() - t0
|
|
1050
1055
|
|
|
1051
1056
|
conv = int(state.converged)
|
|
1052
|
-
|
|
1057
|
+
# Check wall-time limit (issue #5). The JIT loop cannot enforce this
|
|
1058
|
+
# mid-iteration, so we check post-hoc and downgrade the status.
|
|
1059
|
+
exceeded_time = max_wall_time is not None and wall_time > max_wall_time
|
|
1060
|
+
if conv in (1, 2) and not exceeded_time:
|
|
1053
1061
|
status = SolveStatus.OPTIMAL
|
|
1054
|
-
elif conv == 3:
|
|
1062
|
+
elif conv == 3 or exceeded_time:
|
|
1055
1063
|
# If the solution is primal-feasible despite hitting the iteration
|
|
1056
|
-
# limit, report as optimal.
|
|
1057
|
-
# to degenerate curvature but still produce a usable result.
|
|
1064
|
+
# or time limit, report as optimal. The IPM may stall near the
|
|
1065
|
+
# solution due to degenerate curvature but still produce a usable result.
|
|
1058
1066
|
feasible = True
|
|
1059
1067
|
if m > 0 and g_l is not None and g_u is not None and con_fn is not None:
|
|
1060
1068
|
g_final = con_fn(state.x)
|
|
@@ -114,6 +114,11 @@ def _make_problem_data(c, A, b, x_l, x_u):
|
|
|
114
114
|
m = b.shape[0]
|
|
115
115
|
has_lb = (x_l > -_INF).astype(jnp.float64)
|
|
116
116
|
has_ub = (x_u < _INF).astype(jnp.float64)
|
|
117
|
+
# Clamp infinite bounds to safe finite values so that arithmetic
|
|
118
|
+
# inside jax.jit (which evaluates both branches of jnp.where)
|
|
119
|
+
# never produces inf - x = inf, avoiding inf * 0 = NaN downstream.
|
|
120
|
+
x_l = jnp.where(has_lb > 0.5, x_l, -_INF)
|
|
121
|
+
x_u = jnp.where(has_ub > 0.5, x_u, _INF)
|
|
117
122
|
return LPProblemData(
|
|
118
123
|
c=c,
|
|
119
124
|
A=A,
|
|
@@ -235,8 +240,8 @@ def _initialize_state(pd, opts):
|
|
|
235
240
|
opts.bound_push,
|
|
236
241
|
)
|
|
237
242
|
|
|
238
|
-
s_l = jnp.maximum(x - pd.x_l, _SLACK_FLOOR)
|
|
239
|
-
s_u = jnp.maximum(pd.x_u - x, _SLACK_FLOOR)
|
|
243
|
+
s_l = jnp.where(pd.has_lb > 0.5, jnp.maximum(x - pd.x_l, _SLACK_FLOOR), 0.0)
|
|
244
|
+
s_u = jnp.where(pd.has_ub > 0.5, jnp.maximum(pd.x_u - x, _SLACK_FLOOR), 0.0)
|
|
240
245
|
z_l = jnp.where(
|
|
241
246
|
pd.has_lb > 0.5,
|
|
242
247
|
mu / jnp.maximum(s_l, _SLACK_FLOOR),
|
|
@@ -329,10 +334,14 @@ def _iteration_body(carry: LPCarry, tol: float, max_iter: int, tau_min: float) -
|
|
|
329
334
|
z_l, z_u = state.z_l, state.z_u
|
|
330
335
|
tau = jnp.maximum(1.0 - mu, tau_min)
|
|
331
336
|
|
|
332
|
-
s_l = jnp.maximum(x - pd.x_l, _SLACK_FLOOR)
|
|
333
|
-
s_u = jnp.maximum(pd.x_u - x, _SLACK_FLOOR)
|
|
337
|
+
s_l = jnp.where(pd.has_lb > 0.5, jnp.maximum(x - pd.x_l, _SLACK_FLOOR), 0.0)
|
|
338
|
+
s_u = jnp.where(pd.has_ub > 0.5, jnp.maximum(pd.x_u - x, _SLACK_FLOOR), 0.0)
|
|
334
339
|
|
|
335
|
-
Sig =
|
|
340
|
+
Sig = (
|
|
341
|
+
jnp.where(pd.has_lb > 0.5, z_l / jnp.maximum(s_l, _EPS), 0.0)
|
|
342
|
+
+ jnp.where(pd.has_ub > 0.5, z_u / jnp.maximum(s_u, _EPS), 0.0)
|
|
343
|
+
+ _EPS
|
|
344
|
+
)
|
|
336
345
|
Sig_inv = 1.0 / jnp.maximum(Sig, _EPS)
|
|
337
346
|
|
|
338
347
|
r_dual = c - z_l + z_u - A.T @ y
|
|
@@ -479,24 +488,24 @@ def _iteration_body(carry: LPCarry, tol: float, max_iter: int, tau_min: float) -
|
|
|
479
488
|
z_u_new = jnp.maximum(z_u + alpha_d * dz_u, _EPS) * pd.has_ub
|
|
480
489
|
y_new = y + alpha_d * dy
|
|
481
490
|
|
|
482
|
-
#
|
|
483
|
-
|
|
484
|
-
|
|
485
|
-
|
|
486
|
-
|
|
487
|
-
|
|
488
|
-
at_ub = pd.has_ub * (s_u_new <= _SLACK_FLOOR * 2.0).astype(jnp.float64)
|
|
489
|
-
at_lb = pd.has_lb * (s_l_new <= _SLACK_FLOOR * 2.0).astype(jnp.float64)
|
|
490
|
-
both = at_lb * at_ub
|
|
491
|
-
at_lb = at_lb * (1.0 - both)
|
|
492
|
-
at_ub = at_ub * (1.0 - both)
|
|
493
|
-
z_u_new = jnp.where(at_ub > 0.5, z_u_stat, z_u_new)
|
|
494
|
-
z_l_new = jnp.where(at_lb > 0.5, z_l_stat, z_l_new)
|
|
495
|
-
|
|
496
|
-
# --- Update barrier parameter ---
|
|
491
|
+
# Recompute slacks from updated x (without stationarity-based z recovery,
|
|
492
|
+
# which breaks the z*s = mu complementarity invariant).
|
|
493
|
+
s_l_new = jnp.where(pd.has_lb > 0.5, jnp.maximum(x_new - pd.x_l, _SLACK_FLOOR), 0.0)
|
|
494
|
+
s_u_new = jnp.where(pd.has_ub > 0.5, jnp.maximum(pd.x_u - x_new, _SLACK_FLOOR), 0.0)
|
|
495
|
+
|
|
496
|
+
# --- Update barrier parameter (with error gate) ---
|
|
497
497
|
compl = jnp.sum(pd.has_lb * z_l_new * s_l_new) + jnp.sum(pd.has_ub * z_u_new * s_u_new)
|
|
498
|
-
|
|
499
|
-
|
|
498
|
+
mu_candidate = compl / jnp.maximum(n_pairs, 1.0)
|
|
499
|
+
# Only decrease mu when current residuals are small relative to mu
|
|
500
|
+
# (prevents premature decrease that ill-conditions the KKT system).
|
|
501
|
+
r_dual_chk = c - z_l_new + z_u_new - A.T @ y_new
|
|
502
|
+
r_prim_chk = A @ x_new - b
|
|
503
|
+
barrier_err = jnp.maximum(
|
|
504
|
+
jnp.max(jnp.abs(r_prim_chk)) / (1.0 + jnp.max(jnp.abs(b))),
|
|
505
|
+
jnp.max(jnp.abs(r_dual_chk)) / (1.0 + jnp.max(jnp.abs(c))),
|
|
506
|
+
)
|
|
507
|
+
may_decrease = barrier_err < 10.0 * mu
|
|
508
|
+
mu_new = jnp.where(may_decrease, jnp.minimum(mu_candidate, mu), mu)
|
|
500
509
|
mu_new = jnp.maximum(mu_new, _EPS)
|
|
501
510
|
|
|
502
511
|
# --- Check convergence ---
|
|
@@ -603,14 +612,34 @@ def lp_ipm_solve(
|
|
|
603
612
|
x_l = jnp.asarray(x_l, dtype=jnp.float64)
|
|
604
613
|
x_u = jnp.asarray(x_u, dtype=jnp.float64)
|
|
605
614
|
|
|
615
|
+
# --- Problem scaling for improved conditioning ---
|
|
616
|
+
from discopt._jax.scaling import compute_lp_scaling, scale_lp, unscale_lp_solution
|
|
617
|
+
|
|
618
|
+
factors = compute_lp_scaling(c, A)
|
|
619
|
+
c, A, b = scale_lp(c, A, b, factors)
|
|
620
|
+
|
|
606
621
|
# m=0 dispatch — trivial bound-clamping, no need to JIT
|
|
607
622
|
if b.shape[0] == 0:
|
|
608
623
|
pd = _make_problem_data(c, A, b, x_l, x_u)
|
|
609
|
-
|
|
610
|
-
|
|
611
|
-
|
|
624
|
+
state = _solve_unconstrained(pd, opts)
|
|
625
|
+
if factors.applied:
|
|
626
|
+
x_us, y_us, zl_us, zu_us = unscale_lp_solution(
|
|
627
|
+
state.x, state.y, state.z_l, state.z_u, factors
|
|
628
|
+
)
|
|
629
|
+
state = state._replace(x=x_us, y=y_us, z_l=zl_us, z_u=zu_us)
|
|
630
|
+
return state # type: ignore[no-any-return]
|
|
631
|
+
|
|
632
|
+
state = _lp_ipm_solve_jit(
|
|
612
633
|
c, A, b, x_l, x_u, opts.tol, opts.max_iter, opts.tau_min, opts.bound_push
|
|
613
634
|
)
|
|
635
|
+
if factors.applied:
|
|
636
|
+
x_us, y_us, zl_us, zu_us = unscale_lp_solution(
|
|
637
|
+
state.x, state.y, state.z_l, state.z_u, factors
|
|
638
|
+
)
|
|
639
|
+
state = state._replace(x=x_us, y=y_us, z_l=zl_us, z_u=zu_us)
|
|
640
|
+
# Recompute objective on original scale
|
|
641
|
+
state = state._replace(obj=jnp.dot(jnp.asarray(c / factors.obj_scale), state.x))
|
|
642
|
+
return state # type: ignore[no-any-return]
|
|
614
643
|
|
|
615
644
|
|
|
616
645
|
def lp_ipm_solve_batch(
|
|
@@ -160,6 +160,99 @@ class NLPEvaluatorFromNl:
|
|
|
160
160
|
jac[:, j] = (cp - cm) / (2.0 * eps)
|
|
161
161
|
return jac
|
|
162
162
|
|
|
163
|
+
def has_sparse_structure(self) -> bool:
|
|
164
|
+
"""True if numeric probing detected useful sparsity."""
|
|
165
|
+
self._ensure_coo_cache()
|
|
166
|
+
return self._use_sparse
|
|
167
|
+
|
|
168
|
+
def jacobian_structure(self) -> tuple[np.ndarray, np.ndarray]:
|
|
169
|
+
"""Return Jacobian sparsity as COO (rows, cols) arrays."""
|
|
170
|
+
self._ensure_coo_cache()
|
|
171
|
+
return (self._jac_rows, self._jac_cols)
|
|
172
|
+
|
|
173
|
+
def hessian_structure(self) -> tuple[np.ndarray, np.ndarray]:
|
|
174
|
+
"""Return lower-triangle Hessian sparsity as COO (rows, cols) arrays."""
|
|
175
|
+
self._ensure_coo_cache()
|
|
176
|
+
return (self._hess_rows, self._hess_cols)
|
|
177
|
+
|
|
178
|
+
def evaluate_jacobian_values(self, x: np.ndarray) -> np.ndarray:
|
|
179
|
+
"""Return Jacobian values as 1-D array matching jacobian_structure order."""
|
|
180
|
+
self._ensure_coo_cache()
|
|
181
|
+
jac = self.evaluate_jacobian(x)
|
|
182
|
+
return jac[self._jac_rows, self._jac_cols].astype(np.float64)
|
|
183
|
+
|
|
184
|
+
def evaluate_hessian_values(
|
|
185
|
+
self, x: np.ndarray, obj_factor: float, lambda_: np.ndarray
|
|
186
|
+
) -> np.ndarray:
|
|
187
|
+
"""Return Lagrangian Hessian values as 1-D array matching hessian_structure."""
|
|
188
|
+
self._ensure_coo_cache()
|
|
189
|
+
h = self.evaluate_lagrangian_hessian(x, obj_factor, lambda_)
|
|
190
|
+
return h[self._hess_rows, self._hess_cols].astype(np.float64)
|
|
191
|
+
|
|
192
|
+
def _ensure_coo_cache(self) -> None:
|
|
193
|
+
"""Lazily detect sparsity via numeric probing and cache COO indices."""
|
|
194
|
+
if hasattr(self, "_jac_rows"):
|
|
195
|
+
return
|
|
196
|
+
|
|
197
|
+
n = self._n_variables
|
|
198
|
+
m = self._n_constraints
|
|
199
|
+
lb, ub = self.variable_bounds
|
|
200
|
+
|
|
201
|
+
# Clamp bounds for sampling
|
|
202
|
+
lb_safe = np.where(np.isfinite(lb), lb, -10.0)
|
|
203
|
+
ub_safe = np.where(np.isfinite(ub), ub, 10.0)
|
|
204
|
+
|
|
205
|
+
rng = np.random.RandomState(42)
|
|
206
|
+
n_probes = 3
|
|
207
|
+
threshold = 1e-10
|
|
208
|
+
|
|
209
|
+
jac_nonzero = np.zeros((m, n), dtype=bool)
|
|
210
|
+
hess_nonzero = np.zeros((n, n), dtype=bool)
|
|
211
|
+
|
|
212
|
+
for _ in range(n_probes):
|
|
213
|
+
x_probe = lb_safe + rng.rand(n) * (ub_safe - lb_safe)
|
|
214
|
+
if m > 0:
|
|
215
|
+
jac = self.evaluate_jacobian(x_probe)
|
|
216
|
+
jac_nonzero |= np.abs(jac) > threshold
|
|
217
|
+
lam = rng.randn(m) if m > 0 else np.array([], dtype=np.float64)
|
|
218
|
+
hess = self.evaluate_lagrangian_hessian(x_probe, 1.0, lam)
|
|
219
|
+
hess_nonzero |= np.abs(hess) > threshold
|
|
220
|
+
|
|
221
|
+
# Make Hessian pattern symmetric
|
|
222
|
+
hess_nonzero = hess_nonzero | hess_nonzero.T
|
|
223
|
+
|
|
224
|
+
# Decide whether sparsity is worth exploiting
|
|
225
|
+
jac_nnz = int(jac_nonzero.sum())
|
|
226
|
+
hess_nnz = int(hess_nonzero.sum())
|
|
227
|
+
jac_total = m * n if m > 0 else 1
|
|
228
|
+
hess_total = n * n if n > 0 else 1
|
|
229
|
+
jac_density = jac_nnz / jac_total
|
|
230
|
+
hess_density = hess_nnz / hess_total
|
|
231
|
+
self._use_sparse = n >= 50 and (jac_density < 0.15 or hess_density < 0.15)
|
|
232
|
+
|
|
233
|
+
if self._use_sparse:
|
|
234
|
+
if m > 0:
|
|
235
|
+
jac_r, jac_c = np.where(jac_nonzero)
|
|
236
|
+
self._jac_rows = jac_r.astype(np.intp)
|
|
237
|
+
self._jac_cols = jac_c.astype(np.intp)
|
|
238
|
+
else:
|
|
239
|
+
self._jac_rows = np.array([], dtype=np.intp)
|
|
240
|
+
self._jac_cols = np.array([], dtype=np.intp)
|
|
241
|
+
hess_r, hess_c = np.where(hess_nonzero)
|
|
242
|
+
lower_mask = hess_r >= hess_c
|
|
243
|
+
self._hess_rows = hess_r[lower_mask].astype(np.intp)
|
|
244
|
+
self._hess_cols = hess_c[lower_mask].astype(np.intp)
|
|
245
|
+
else:
|
|
246
|
+
# Dense fallback
|
|
247
|
+
if m > 0:
|
|
248
|
+
rows, cols = np.meshgrid(np.arange(m), np.arange(n), indexing="ij")
|
|
249
|
+
self._jac_rows = rows.flatten().astype(np.intp)
|
|
250
|
+
self._jac_cols = cols.flatten().astype(np.intp)
|
|
251
|
+
else:
|
|
252
|
+
self._jac_rows = np.array([], dtype=np.intp)
|
|
253
|
+
self._jac_cols = np.array([], dtype=np.intp)
|
|
254
|
+
self._hess_rows, self._hess_cols = np.tril_indices(n)
|
|
255
|
+
|
|
163
256
|
@property
|
|
164
257
|
def n_variables(self) -> int:
|
|
165
258
|
"""Total number of variables (flat)."""
|