dfbench 0.2.2__tar.gz → 0.3.0__tar.gz

{dfbench-0.2.2 → dfbench-0.3.0}/.gitignore

@@ -23,11 +23,7 @@ temp/
 results/
 notebooks/
 .DS_Store
-
-runs/
-competition/
-thesis.typ
-training/
+.github/
 
 # IDE related
 .vscode/

{dfbench-0.2.2 → dfbench-0.3.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: dfbench
-Version: 0.2.2
-Summary: Differometor Benchmark - Optimization algorithms and problem definitions
+Version: 0.3.0
+Summary: dfbench: A Black-Box Optimization Framework for JAX-Differentiable Gravitational Wave Detector Discovery
 Project-URL: Homepage, https://github.com/artificial-scientist-lab/Differometor-Benchmark
 Project-URL: Repository, https://github.com/artificial-scientist-lab/Differometor-Benchmark
 Project-URL: Documentation, https://github.com/artificial-scientist-lab/Differometor-Benchmark/tree/main/docs
@@ -90,9 +90,6 @@ Provides-Extra: torch
 Requires-Dist: torch>=2.9.1; extra == 'torch'
 Description-Content-Type: text/markdown
 
-> [!IMPORTANT]
-> **BETA RELEASE:** This repository is currently in beta. The official competition will start on **July ~7**.
-
 # Differometor Benchmark
 
 [![image](https://img.shields.io/pypi/v/dfbench.svg)](https://pypi.python.org/pypi/dfbench)

{dfbench-0.2.2 → dfbench-0.3.0}/README.md

@@ -1,6 +1,3 @@
-> [!IMPORTANT]
-> **BETA RELEASE:** This repository is currently in beta. The official competition will start on **July ~7**.
-
 # Differometor Benchmark
 
 [![image](https://img.shields.io/pypi/v/dfbench.svg)](https://pypi.python.org/pypi/dfbench)

{dfbench-0.2.2 → dfbench-0.3.0}/docs/Implementing-a-New-Algorithm.md

@@ -98,7 +98,6 @@ class MyAlgorithm(OptimizationAlgorithm):
         """
         # ─── 1. Setup references ───
         obj = objective
-        problem = obj.problem
 
         # Sets unbounded mode, algorithm_str, seeds np/JAX, returns resolved seed + JAX key
         random_seed, key = self.prepare(
@@ -179,7 +178,6 @@ The base class `OptimizationAlgorithm.optimize()` contains a commented blueprint
 
 ```python
 obj = objective
-problem = obj.problem
 random_seed, key = self.prepare(
     obj,
     unbounded=False,
@@ -187,6 +185,12 @@ random_seed, key = self.prepare(
 )
 ```
 
+The problem instance wrapped by the `Objective` is private. Use the public
+Objective API directly: `obj.bounds` for the `(2, n_params)` bounds,
+`obj.n_params` for the dimension, `obj.problem_name` for the display name,
+`obj.value_function(unbounded=False)` for the raw bounded objective, and
+`obj.penalty_fn` / `obj.power_thresholds` for the penalty contract.
+
 `prepare()` is called as `prepare(obj, unbounded, random_seed, algorithm_str=None, **kwargs)`. It sets `obj.unbounded`, `obj.algorithm_str`, seeds `np.random` and JAX, and returns `(random_seed, key)`. If `random_seed=None` is passed, a seed is generated via system entropy. You can also configure the Objective manually instead of calling `prepare()`.
 
 **Choose `unbounded`:**

{dfbench-0.2.2 → dfbench-0.3.0}/docs/Objective-API-Reference.md

@@ -447,7 +447,7 @@ bounded ≈ lb + (ub - lb) * forward(random_params_unbounded(...))
 |----------|------|-------------|
 | `bounds` | `Array[2, n_params]` | Lower and upper parameter bounds (or $\pm\infty$ when unbounded). |
 | `n_params` | `int` | Number of optimizable parameters. |
-| `problem` | `ContinuousProblem` | The wrapped problem instance. |
+| `problem_name` | `str` | Display name of the wrapped problem. |
 | `penalty_fn` | `Callable \| None` | The wrapped problem's penalty function, or `None` if the problem does not expose one. Update it via `set_penalty_fn(fn)`. |
 | `power_thresholds` | `dict[str, float] \| None` | Per-group power thresholds (`hard`, `soft`, `detector`) for problems that opt into the penalty contract, or `None`. Constants; do not change across evaluations. |
 

{dfbench-0.2.2 → dfbench-0.3.0}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "hatchling.build"
 
 [project]
 name = "dfbench"
-version = "0.2.2"
-description = "Differometor Benchmark - Optimization algorithms and problem definitions"
+version = "0.3.0"
+description = "dfbench: A Black-Box Optimization Framework for JAX-Differentiable Gravitational Wave Detector Discovery"
 readme = "README.md"
 requires-python = ">=3.11,<3.14"
 license = "MIT"

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/derivative_free/_dfo_common.py

@@ -81,7 +81,7 @@ def random_bounded_start(
         (x0, new_key): x0 is a 1-D ``np.ndarray`` of shape ``(n_params,)``.
     """
     key, subkey = random.split(key)
-    lower, upper = obj.problem.bounds
+    lower, upper = obj.bounds
     x0_jax = random.uniform(subkey, shape=(obj.n_params,), minval=lower, maxval=upper)
     return np.asarray(x0_jax, dtype=np.float64), key
 
@@ -92,15 +92,15 @@ def clip_to_bounds(x: np.ndarray, obj: Objective) -> np.ndarray:
     Returns:
         Clipped copy of *x*.
     """
-    lower = np.asarray(obj.problem.bounds[0], dtype=np.float64)
-    upper = np.asarray(obj.problem.bounds[1], dtype=np.float64)
+    lower = np.asarray(obj.bounds[0], dtype=np.float64)
+    upper = np.asarray(obj.bounds[1], dtype=np.float64)
     return np.clip(x, lower, upper)
 
 
 def solver_bounds_np(obj: Objective) -> tuple[np.ndarray, np.ndarray]:
     """Return ``(lower, upper)`` as plain NumPy float64 arrays."""
-    lower = np.asarray(obj.problem.bounds[0], dtype=np.float64)
-    upper = np.asarray(obj.problem.bounds[1], dtype=np.float64)
+    lower = np.asarray(obj.bounds[0], dtype=np.float64)
+    upper = np.asarray(obj.bounds[1], dtype=np.float64)
     return lower, upper
 
 

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/derivative_free/_scipy_common.py

@@ -36,15 +36,15 @@ _NAN_PENALTY = 1e30
 
 def scipy_bounds(obj: Objective) -> Bounds:
     """Convert problem bounds to :class:`scipy.optimize.Bounds`."""
-    lb = np.asarray(obj.problem.bounds[0], dtype=np.float64)
-    ub = np.asarray(obj.problem.bounds[1], dtype=np.float64)
+    lb = np.asarray(obj.bounds[0], dtype=np.float64)
+    ub = np.asarray(obj.bounds[1], dtype=np.float64)
     return Bounds(lb, ub)
 
 
 def scipy_bounds_list(obj: Objective) -> list[tuple[float, float]]:
     """Return bounds as ``[(lb0, ub0), ...]`` (format used by *dual_annealing*)."""
-    lb = np.asarray(obj.problem.bounds[0], dtype=np.float64)
-    ub = np.asarray(obj.problem.bounds[1], dtype=np.float64)
+    lb = np.asarray(obj.bounds[0], dtype=np.float64)
+    ub = np.asarray(obj.bounds[1], dtype=np.float64)
     return list(zip(lb.tolist(), ub.tolist()))
 
 

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/derivative_free/omads_mads.py

@@ -102,11 +102,10 @@ def _run_omads(
     """
     _check_omads_available()
 
-    problem = obj.problem
-    bounds = problem.bounds  # (2, n_params)
+    bounds = obj.bounds  # (2, n_params)
     lower = np.array(bounds[0], dtype=np.float64)
     upper = np.array(bounds[1], dtype=np.float64)
-    n_params = problem.n_params
+    n_params = obj.n_params
 
     # Starting point: use best_params if available, else random bounded point
     if obj.best_params is not None:

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/derivative_free/pdfo/lincoa.py

@@ -1,14 +1,8 @@
 """LINCOA: LINearly Constrained Optimization Algorithm (via PDFO).
 
 LINCOA is a derivative-free trust-region method by M. J. D. Powell that
-handles *bound constraints* and *linear inequality constraints* natively.
-
-When the problem exposes only box bounds (the common case in this benchmark),
-LINCOA automatically uses them.  If the problem also exposes linear
-constraints via ``problem.linear_constraints`` (returning a dict with keys
-``A_ub``, ``b_ub`` for ``A_ub @ x <= b_ub``), those are forwarded to the
-solver as well.  Problems without any constraints are also supported -
-LINCOA will behave like a bounded NEWUOA.
+handles *bound constraints* natively.  Problems without any constraints are
+also supported - LINCOA will behave like a bounded NEWUOA.
 
 Defaults are conservative and benchmark-oriented: bounded physical space,
 one restart, ``rhobeg`` derived from the bound range.
@@ -76,7 +70,7 @@ class PDFOLINCOA(OptimizationAlgorithm):
             raise ImportError(
                 "PDFO is required for LINCOA.  Install with: uv add 'dfbench[dfo]'"
             ) from exc
-        from scipy.optimize import LinearConstraint, Bounds  # noqa: E402
+        from scipy.optimize import Bounds  # noqa: E402
 
         obj = objective
         random_seed, key = self.prepare(obj, unbounded=False, random_seed=random_seed)
@@ -92,15 +86,9 @@ class PDFOLINCOA(OptimizationAlgorithm):
         fun = dfo_objective_wrapper(obj)
         bounds = Bounds(lb=lower, ub=upper)
 
-        # Build constraint list: always include bounds; add linear if available.
+        # This batch only supports box bounds; problems with linear
+        # constraint metadata are not handled here.
         constraints: list = []
-        problem = obj.problem
-        if hasattr(problem, "linear_constraints"):
-            lc = problem.linear_constraints  # type: ignore[attr-defined]
-            if lc is not None:
-                A_ub = np.asarray(lc["A_ub"], dtype=np.float64)
-                b_ub = np.asarray(lc["b_ub"], dtype=np.float64)
-                constraints.append(LinearConstraint(A_ub, ub=b_ub))
 
         # JIT warmup
         obj.warmup_value()

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/cmaes_sep_cma.py

@@ -106,14 +106,13 @@ class CMAESSepCMA(OptimizationAlgorithm):
                 disables stagnation stopping.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         width = ub_np - lb_np
-        n = problem.n_params
+        n = obj.n_params
 
         # The cmaes package has no per-coordinate sigma.  On problems where
         # bound widths span orders of magnitude (e.g. Voyager), running the

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/evosax_es.py

@@ -198,15 +198,14 @@ class EvosaxMAES(OptimizationAlgorithm):
             max_iterations: Maximum CMA generations.  ``None`` = unlimited.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, rng = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         lb_jnp = jnp.asarray(lb_np)
         ub_jnp = jnp.asarray(ub_np)
-        n = problem.n_params
+        n = obj.n_params
 
         sigma = float(sigma0 if sigma0 is not None else np.mean(0.3 * (ub_np - lb_np)))
 
@@ -307,15 +306,14 @@ class EvosaxLMMAES(OptimizationAlgorithm):
             max_iterations: Maximum generations.  ``None`` = unlimited.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, rng = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         lb_jnp = jnp.asarray(lb_np)
         ub_jnp = jnp.asarray(ub_np)
-        n = problem.n_params
+        n = obj.n_params
 
         sigma = float(sigma0 if sigma0 is not None else np.mean(0.3 * (ub_np - lb_np)))
 

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/evox_es.py

@@ -68,7 +68,7 @@ class EvoxES(OptimizationAlgorithm):
         _variant (str): ES variant name (e.g., "CMAES", "OpenES").
 
     Note:
-        This algorithm uses `problem.objective_function` with the problem's bounds.
+        This algorithm uses the objective function with the problem's bounds.
         The population searches directly in the bounded parameter space.
 
     Example:
@@ -146,17 +146,11 @@ class EvoxES(OptimizationAlgorithm):
             **es_kwargs: Variant-specific keyword arguments passed to the EvoX algorithm.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
         torch.manual_seed(random_seed)
 
-        # Get bounds from problem
-        if not hasattr(problem, "bounds"):
-            raise ValueError(
-                f"Problem {type(problem).__name__} must have a 'bounds' attribute."
-            )
-        problem_bounds = problem.bounds
+        problem_bounds = obj.bounds
         lb_np = np.asarray(problem_bounds[0])
         ub_np = np.asarray(problem_bounds[1])
 
@@ -210,7 +204,7 @@ class EvoxES(OptimizationAlgorithm):
 
         def get_center_init():
             return torch.from_numpy(
-                np.random.uniform(lb_np, ub_np, size=problem.n_params)
+                np.random.uniform(lb_np, ub_np, size=obj.n_params)
             ).float()
 
         # Initialize based on variant-specific requirements
@@ -241,7 +235,7 @@ class EvoxES(OptimizationAlgorithm):
             if "init_mean" not in es_kwargs:
                 es_kwargs["init_mean"] = get_center_init()
             if "init_covar" not in es_kwargs:
-                es_kwargs["init_covar"] = torch.eye(problem.n_params).float() * (
+                es_kwargs["init_covar"] = torch.eye(obj.n_params).float() * (
                     default_sigma**2
                 )
             algorithm = AlgorithmClass(pop_size=pop_size, **es_kwargs)
@@ -250,9 +244,7 @@ class EvoxES(OptimizationAlgorithm):
             if "init_mean" not in es_kwargs:
                 es_kwargs["init_mean"] = get_center_init()
             if "init_std" not in es_kwargs:
-                es_kwargs["init_std"] = (
-                    torch.ones(problem.n_params).float() * default_sigma
-                )
+                es_kwargs["init_std"] = torch.ones(obj.n_params).float() * default_sigma
             algorithm = AlgorithmClass(pop_size=pop_size, **es_kwargs)
 
         elif self._variant == "DES":

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/evox_pso.py

@@ -41,7 +41,7 @@ class EvoxPSO(OptimizationAlgorithm):
         _variant (str): PSO variant name (uppercase, e.g., "PSO", "CLPSO").
 
     Note:
-        This algorithm uses `problem.objective_function` with the problem's bounds.
+        This algorithm uses the objective function with the problem's bounds.
         The swarm searches directly in the bounded parameter space.
 
     Example:
@@ -114,17 +114,11 @@ class EvoxPSO(OptimizationAlgorithm):
             **pso_kwargs: Variant-specific keyword arguments passed to the EvoX algorithm.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
         torch.manual_seed(random_seed)
 
-        # Get bounds from problem
-        if not hasattr(problem, "bounds"):
-            raise ValueError(
-                f"Problem {type(problem).__name__} must have a 'bounds' attribute."
-            )
-        problem_bounds = problem.bounds
+        problem_bounds = obj.bounds
         lb = j2t(problem_bounds[0])
         ub = j2t(problem_bounds[1])
 

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/jax_es.py

@@ -126,13 +126,12 @@ class JAXOnePlusOneES(OptimizationAlgorithm):
                 means unlimited (budget governs stopping).
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, rng = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb = jnp.asarray(problem.bounds[0])
-        ub = jnp.asarray(problem.bounds[1])
-        n = problem.n_params
+        lb = jnp.asarray(obj.bounds[0])
+        ub = jnp.asarray(obj.bounds[1])
+        n = obj.n_params
 
         if init_params is None:
             rng, init_rng = jax.random.split(rng)
@@ -279,13 +278,12 @@ class JAXMuLambdaES(OptimizationAlgorithm):
             raise ValueError(f"(μ,λ)-ES requires mu < lam, got mu={mu}, lam={lam}.")
 
         obj = objective
-        problem = obj.problem
 
         random_seed, rng = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb = jnp.asarray(problem.bounds[0])
-        ub = jnp.asarray(problem.bounds[1])
-        n = problem.n_params
+        lb = jnp.asarray(obj.bounds[0])
+        ub = jnp.asarray(obj.bounds[1])
+        n = obj.n_params
 
         if init_params is None:
             rng, init_rng = jax.random.split(rng)

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/nevergrad/ngopt.py

@@ -64,17 +64,13 @@ class NevergradNGOpt(OptimizationAlgorithm):
             n_restarts: Number of independent restarts. Budget is split evenly.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        if not hasattr(problem, "bounds"):
-            raise ValueError("NGOpt requires a bounded problem (problem.bounds).")
-
-        bounds = problem.bounds
+        bounds = obj.bounds
         lb = np.asarray(bounds[0], dtype=np.float64)
         ub = np.asarray(bounds[1], dtype=np.float64)
-        n_params = int(problem.n_params)
+        n_params = int(obj.n_params)
 
         budget_per_restart = max_iterations // n_restarts if max_iterations else None
 

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/nevergrad/oneplusone.py

@@ -64,17 +64,13 @@ class NevergradOnePlusOne(OptimizationAlgorithm):
             n_restarts: Number of independent restarts. Budget is split evenly.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        if not hasattr(problem, "bounds"):
-            raise ValueError("OnePlusOne requires a bounded problem (problem.bounds).")
-
-        bounds = problem.bounds
+        bounds = obj.bounds
         lb = np.asarray(bounds[0], dtype=np.float64)
         ub = np.asarray(bounds[1], dtype=np.float64)
-        n_params = int(problem.n_params)
+        n_params = int(obj.n_params)
 
         # Budget per restart
         budget_per_restart = max_iterations // n_restarts if max_iterations else None

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/nevergrad/tbpsa.py

@@ -73,17 +73,13 @@ class NevergradTBPSA(OptimizationAlgorithm):
                 counts against the Objective budget. Set >1 for noisy problems.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        if not hasattr(problem, "bounds"):
-            raise ValueError("TBPSA requires a bounded problem (problem.bounds).")
-
-        bounds = problem.bounds
+        bounds = obj.bounds
         lb = np.asarray(bounds[0], dtype=np.float64)
         ub = np.asarray(bounds[1], dtype=np.float64)
-        n_params = int(problem.n_params)
+        n_params = int(obj.n_params)
 
         budget_per_restart = max_iterations // n_restarts if max_iterations else None
 

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/evolutionary/pycma_cmaes.py

@@ -208,14 +208,13 @@ class PyCMACMAES(OptimizationAlgorithm):
                 means unlimited (budget and convergence govern stopping).
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         width = ub_np - lb_np
-        n = problem.n_params
+        n = obj.n_params
 
         # Build the initial mean in unit-cube coordinates.
         if init_params is None:
@@ -304,14 +303,13 @@ class PyCMAActiveCMAES(OptimizationAlgorithm):
             max_iterations: Maximum CMA generations.  ``None`` = unlimited.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         width = ub_np - lb_np
-        n = problem.n_params
+        n = obj.n_params
 
         # Build the initial mean in unit-cube coordinates.
         if init_params is None:
@@ -406,14 +404,13 @@ class PyCMAIPOP(OptimizationAlgorithm):
                 (independent of the overall budget).  ``None`` = unlimited.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         width = ub_np - lb_np
-        n = problem.n_params
+        n = obj.n_params
 
         base_pop = pop_size or _default_pop_size(n)
         # sigma is a dimensionless fraction of the unit cube.
@@ -521,14 +518,13 @@ class PyCMABIPOP(OptimizationAlgorithm):
                 ``None`` = unlimited.
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
-        lb_np = np.asarray(problem.bounds[0])
-        ub_np = np.asarray(problem.bounds[1])
+        lb_np = np.asarray(obj.bounds[0])
+        ub_np = np.asarray(obj.bounds[1])
         width = ub_np - lb_np
-        n = problem.n_params
+        n = obj.n_params
 
         base_pop = pop_size or _default_pop_size(n)
         # sigma is a dimensionless fraction of the unit cube.

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/generative/vae_sampling.py

@@ -326,12 +326,11 @@ class VAESampling(OptimizationAlgorithm):
             raise ValueError("top_k must be a fraction in (0, 1].")
 
         obj = objective
-        problem = obj.problem
 
         random_seed, _ = self.prepare(obj, unbounded=True, random_seed=random_seed)
         torch.manual_seed(random_seed)
 
-        input_dim = problem.n_params
+        input_dim = obj.n_params
         latent_dim = input_dim // latent_dim_factor + 1
 
         # Create VAE

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/global_search/basin_hopping.py

@@ -87,7 +87,6 @@ class BasinHopping(OptimizationAlgorithm):
                 of the bound range per dimension).
         """
         obj = objective
-        problem = obj.problem
 
         random_seed, _key = self.prepare(obj, unbounded=False, random_seed=random_seed)
 
@@ -120,8 +119,8 @@ class BasinHopping(OptimizationAlgorithm):
             minimizer_kwargs["bounds"] = scipy_bounds(obj)
 
         # ── Bounded step-taker ────────────────────────────────────────
-        lb = np.asarray(problem.bounds[0], dtype=np.float64)
-        ub = np.asarray(problem.bounds[1], dtype=np.float64)
+        lb = np.asarray(obj.bounds[0], dtype=np.float64)
+        ub = np.asarray(obj.bounds[1], dtype=np.float64)
         take_step = BoundedStep(stepsize, lb, ub)
 
         x0 = np.asarray(params, dtype=np.float64)

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/gradient_based/scipy/_common.py

@@ -95,27 +95,16 @@ class SciPyObjectiveAdapter:
         """Return SciPy bounds in physical space when requested."""
         if not self.config.use_bounds:
             return None
-        problem_bounds = np.asarray(self.obj.problem.bounds, dtype=float)
+        problem_bounds = np.asarray(self.obj.bounds, dtype=float)
         return Bounds(problem_bounds[0], problem_bounds[1], keep_feasible=True)
 
     @property
     def constraints(self) -> tuple:
-        """Return supported SciPy constraints or fail loudly."""
-        problem = self.obj.problem
-        for attr_name in (
-            "scipy_constraints",
-            "constraints",
-            "linear_constraints",
-            "nonlinear_constraints",
-        ):
-            if not hasattr(problem, attr_name):
-                continue
-            value = getattr(problem, attr_name)
-            if value not in (None, (), [], {}):
-                raise NotImplementedError(
-                    "Problem exposes constraint metadata via "
-                    f"'{attr_name}', but this batch only supports box constraints."
-                )
+        """Return supported SciPy constraints.
+
+        This batch only supports box bounds; problems with constraint
+        metadata are not handled here.
+        """
         return ()
 
     def warmup(self) -> None:

{dfbench-0.2.2 → dfbench-0.3.0}/src/dfbench/algorithms/surrogate_based/ax_baxus.py

@@ -133,14 +133,13 @@ class BAxUS(OptimizationAlgorithm):
             **bo_kwargs: Extra kwargs for acquisition optimisation.
         """
         obj = objective
-        problem = obj.problem
-        D = problem.n_params
+        D = obj.n_params
 
         random_seed, _ = self.prepare(obj, unbounded=False, random_seed=random_seed)
         torch.manual_seed(random_seed)
         rng = torch.Generator(device="cpu").manual_seed(random_seed)
 
-        bounds = get_problem_bounds_torch(problem, self.device, self.dtype)
+        bounds = get_problem_bounds_torch(obj.bounds, self.device, self.dtype)
 
         if d_init is None:
             d_init = min(5, D)