Update np.Inf -> np.inf per numpy 2.0 release (#2527)

Summary:
Pull Request resolved: #2527

Numpy 2.0 was released yesterday ([post](https://numpy.org/news/#numpy-20-release-date-june-16)) and accordingly a couple of tests started breaking due to usage of a now-outdated API. This updates references to the now-deprecated `np.Inf` to `np.inf`, which is available in both old and new versions of numpy and so shouldn't create any issues.

Our early-stopping tutorial is still breaking due to a similar issue in TensorBoard (tensorflow/tensorboard#6869), so not all tests are expected to pass as of now.

Reviewed By: mpolson64

Differential Revision: D58672474

fbshipit-source-id: 717b751ac9608867976923bc60ecb5656cc36e08

ax/core/utils.py

@@ -141,7 +141,7 @@ def best_feasible_objective(
  objective = optimization_config.objective
  f = values[objective.metric.name]
  # Set infeasible points to have infinitely bad values
- infeas_val = np.Inf if objective.minimize else -np.Inf
+ infeas_val = np.inf if objective.minimize else -np.inf
  for oc in optimization_config.outcome_constraints:
  if oc.relative:
  raise ValueError(

ax/models/discrete/thompson.py

@@ -213,9 +213,9 @@ def _produce_samples(
  objective_values = np.dot(
  samples_per_metric, objective_weights
  ) # (k x num_samples)
- objective_values[any_violation] = -np.Inf
+ objective_values[any_violation] = -np.inf
  best_arm = objective_values.max(axis=0) # (num_samples,)
- all_arms_infeasible = best_arm == -np.Inf # (num_samples,)
+ all_arms_infeasible = best_arm == -np.inf # (num_samples,)
  fraction_all_infeasible = all_arms_infeasible.mean()
  filtered_objective = objective_values[:, ~all_arms_infeasible] # (k x ?)
  return filtered_objective, fraction_all_infeasible

ax/models/model_utils.py

@@ -453,15 +453,15 @@ def best_in_sample_point(
  # Identify best point
  if method == "feasible_threshold":
  utility = obj
- utility[pfeas < threshold] = -np.Inf
+ utility[pfeas < threshold] = -np.inf
  elif method == "max_utility":
  if B is None:
  B = obj.min()
  utility = (obj - B) * pfeas
  else: # pragma: no cover
  raise UnsupportedError(f"Unknown best point method {method}.")
  i = np.argmax(utility)
- if utility[i] == -np.Inf:
+ if utility[i] == -np.inf:
  return None
  else:
  return X_obs[i, :], utility[i]