|
| 1 | +"""Problem class: wraps a C problem capsule (objective + list of constraints). |
| 2 | +
|
| 3 | +Takes SparseDiffPy expressions (not CVXPY, not pre-built capsules). CVXPY-facing |
| 4 | +adapters live in downstream libraries (e.g. DNLP). |
| 5 | +""" |
| 6 | + |
| 7 | +import numpy as np |
| 8 | + |
| 9 | +from sparsediffpy import _sparsediffengine as _C |
| 10 | +from sparsediffpy._core._compile import _build_capsule, _collect_leaves |
| 11 | + |
| 12 | + |
| 13 | +class Problem: |
| 14 | + """A compiled NLP-style problem: one scalar objective plus a list of constraints. |
| 15 | +
|
| 16 | + Method names mirror DNLP's `C_problem` so a CVXPY adapter can return a |
| 17 | + Problem and existing solver callsites keep working. |
| 18 | + """ |
| 19 | + |
| 20 | + def __init__(self, objective, constraints=None, verbose=False): |
| 21 | + constraints = list(constraints) if constraints else [] |
| 22 | + |
| 23 | + if objective.shape != (1, 1): |
| 24 | + raise ValueError( |
| 25 | + f"Objective must be scalar (shape (1, 1)), got {objective.shape}" |
| 26 | + ) |
| 27 | + |
| 28 | + variables, parameters = [], [] |
| 29 | + visited = set() |
| 30 | + _collect_leaves(objective, variables, parameters, visited) |
| 31 | + for c in constraints: |
| 32 | + _collect_leaves(c, variables, parameters, visited) |
| 33 | + |
| 34 | + if not variables: |
| 35 | + raise ValueError("Problem must contain at least one Variable") |
| 36 | + |
| 37 | + scope = variables[0]._scope |
| 38 | + for v in variables[1:]: |
| 39 | + if v._scope is not scope: |
| 40 | + raise ValueError("All variables must belong to the same Scope") |
| 41 | + |
| 42 | + n_vars = scope._next_var_offset |
| 43 | + |
| 44 | + # One shared cache across objective + all constraints: CSE in both |
| 45 | + # directions (within an expression, and across the obj/constraint |
| 46 | + # boundary) is safe, and each Parameter capsule is appended to |
| 47 | + # param_caps exactly once. |
| 48 | + cache = {} |
| 49 | + param_caps, param_objs = [], [] |
| 50 | + obj_cap = _build_capsule(objective, n_vars, cache, param_caps, param_objs) |
| 51 | + constraint_caps = [ |
| 52 | + _build_capsule(c, n_vars, cache, param_caps, param_objs) |
| 53 | + for c in constraints |
| 54 | + ] |
| 55 | + |
| 56 | + self._capsule = _C.make_problem(obj_cap, constraint_caps, verbose) |
| 57 | + if param_caps: |
| 58 | + _C.problem_register_params(self._capsule, param_caps) |
| 59 | + |
| 60 | + self._scope = scope |
| 61 | + self._param_capsules = param_caps |
| 62 | + self._param_objects = param_objs |
| 63 | + self._n_vars = n_vars |
| 64 | + self._total_constraint_size = sum(c.size for c in constraints) |
| 65 | + self._jacobian_coo_initialized = False |
| 66 | + self._hessian_coo_initialized = False |
| 67 | + |
| 68 | + if param_caps: |
| 69 | + self._sync_params() |
| 70 | + |
| 71 | + # ------------------------------------------------------------------ |
| 72 | + # Internal |
| 73 | + # ------------------------------------------------------------------ |
| 74 | + |
| 75 | + def _sync_params(self): |
| 76 | + """Push current Parameter values to the C problem. |
| 77 | +
|
| 78 | + Called once at construction. After construction, callers invoke |
| 79 | + update_params(theta) explicitly (matching DNLP's solver-loop contract). |
| 80 | + """ |
| 81 | + for p in self._param_objects: |
| 82 | + if p._value_flat is None: |
| 83 | + raise ValueError( |
| 84 | + f"Parameter with shape {p.shape} has no value set. " |
| 85 | + f"Assign a value via parameter.value = ... before constructing Problem." |
| 86 | + ) |
| 87 | + theta = np.concatenate([p._value_flat for p in self._param_objects]) |
| 88 | + _C.problem_update_params(self._capsule, theta) |
| 89 | + self._scope._params_dirty = False |
| 90 | + |
| 91 | + # ------------------------------------------------------------------ |
| 92 | + # Parameter updates |
| 93 | + # ------------------------------------------------------------------ |
| 94 | + |
| 95 | + def update_params(self, theta): |
| 96 | + """Update parameter values in the C DAG from a flat theta vector. |
| 97 | +
|
| 98 | + Sparsity structures (Jacobian/Hessian) remain valid after this call. |
| 99 | + """ |
| 100 | + theta = np.asarray(theta, dtype=np.float64) |
| 101 | + _C.problem_update_params(self._capsule, theta) |
| 102 | + self._scope._params_dirty = False |
| 103 | + |
| 104 | + # ------------------------------------------------------------------ |
| 105 | + # Sparsity initialization (COO) |
| 106 | + # ------------------------------------------------------------------ |
| 107 | + |
| 108 | + def init_jacobian_coo(self): |
| 109 | + """Fill sparsity for the constraint Jacobian in COO format. |
| 110 | +
|
| 111 | + Must be called once before get_jacobian_sparsity_coo() or eval_jacobian_vals(). |
| 112 | + """ |
| 113 | + _C.problem_init_jacobian_coo(self._capsule) |
| 114 | + self._jacobian_coo_initialized = True |
| 115 | + |
| 116 | + def init_hessian_coo_lower_tri(self): |
| 117 | + """Fill sparsity for the Lagrangian Hessian (lower triangle, COO). |
| 118 | +
|
| 119 | + Must be called once before get_problem_hessian_sparsity_coo() or |
| 120 | + eval_hessian_vals_coo_lower_tri(). |
| 121 | + """ |
| 122 | + _C.problem_init_hessian_coo_lower_triangular(self._capsule) |
| 123 | + self._hessian_coo_initialized = True |
| 124 | + |
| 125 | + # ------------------------------------------------------------------ |
| 126 | + # Forward evaluation |
| 127 | + # ------------------------------------------------------------------ |
| 128 | + |
| 129 | + def objective_forward(self, u): |
| 130 | + """Evaluate the objective at variable values `u`. Returns a float.""" |
| 131 | + u = np.asarray(u, dtype=np.float64) |
| 132 | + return _C.problem_objective_forward(self._capsule, u) |
| 133 | + |
| 134 | + def constraint_forward(self, u): |
| 135 | + """Evaluate constraints at variable values `u`. Returns an np.ndarray.""" |
| 136 | + u = np.asarray(u, dtype=np.float64) |
| 137 | + return _C.problem_constraint_forward(self._capsule, u) |
| 138 | + |
| 139 | + def gradient(self): |
| 140 | + """Compute gradient of the objective. Call objective_forward first.""" |
| 141 | + return _C.problem_gradient(self._capsule) |
| 142 | + |
| 143 | + # ------------------------------------------------------------------ |
| 144 | + # Jacobian (COO path) |
| 145 | + # ------------------------------------------------------------------ |
| 146 | + |
| 147 | + def get_jacobian_sparsity_coo(self): |
| 148 | + """Return the sparsity pattern (rows, cols) of the constraint Jacobian. |
| 149 | +
|
| 150 | + Call init_jacobian_coo() first. |
| 151 | + """ |
| 152 | + rows, cols, _shape = _C.get_jacobian_sparsity_coo(self._capsule) |
| 153 | + return rows, cols |
| 154 | + |
| 155 | + def eval_jacobian_vals(self): |
| 156 | + """Evaluate the constraint Jacobian and return its nonzero values. |
| 157 | +
|
| 158 | + Values correspond to the sparsity pattern from get_jacobian_sparsity_coo(). |
| 159 | + Call constraint_forward() first to set the evaluation point. |
| 160 | + """ |
| 161 | + return _C.problem_eval_jacobian_vals(self._capsule) |
| 162 | + |
| 163 | + # ------------------------------------------------------------------ |
| 164 | + # Lagrangian Hessian (COO lower-triangular path) |
| 165 | + # ------------------------------------------------------------------ |
| 166 | + |
| 167 | + def get_problem_hessian_sparsity_coo(self): |
| 168 | + """Return the sparsity pattern (rows, cols) of the lower-triangular |
| 169 | + Lagrangian Hessian. |
| 170 | +
|
| 171 | + Call init_hessian_coo_lower_tri() first. |
| 172 | + """ |
| 173 | + rows, cols, _shape = _C.get_problem_hessian_sparsity_coo(self._capsule) |
| 174 | + return rows, cols |
| 175 | + |
| 176 | + def eval_hessian_vals_coo_lower_tri(self, obj_factor, lagrange): |
| 177 | + """Evaluate the lower-triangular Lagrangian Hessian values. |
| 178 | +
|
| 179 | + Computes obj_factor * H_f + sum_i lagrange[i] * H_gi, where f is the |
| 180 | + objective and g_i are the constraints. Values correspond to the sparsity |
| 181 | + pattern from get_problem_hessian_sparsity_coo(). |
| 182 | +
|
| 183 | + Call objective_forward() and constraint_forward() first to set the |
| 184 | + evaluation point. |
| 185 | + """ |
| 186 | + lagrange = np.asarray(lagrange, dtype=np.float64) |
| 187 | + if lagrange.size != self._total_constraint_size: |
| 188 | + raise ValueError( |
| 189 | + f"lagrange length {lagrange.size} != total_constraint_size " |
| 190 | + f"{self._total_constraint_size}" |
| 191 | + ) |
| 192 | + return _C.problem_eval_hessian_vals_coo( |
| 193 | + self._capsule, float(obj_factor), lagrange |
| 194 | + ) |
0 commit comments