Upgrade fbcode/pytorch to Python Scientific Stack 2 (#2584)

Summary:
X-link: pytorch/opacus#680

X-link: pytorch/captum#1387


X-link: pytorch/audio#3845

Differential Revision: D64008689

Author: igorsugak

botorch/exceptions/errors.py

@@ -10,7 +10,7 @@
 
 from typing import Any
 
-import numpy as np
+import numpy.typing as npt
 
 
 class BotorchError(Exception):
@@ -59,7 +59,7 @@ class OptimizationTimeoutError(BotorchError):
     r"""Exception raised when optimization times out."""
 
     def __init__(
-        self, /, *args: Any, current_x: np.ndarray, runtime: float, **kwargs: Any
+        self, /, *args: Any, current_x: npt.NDArray, runtime: float, **kwargs: Any
     ) -> None:
         r"""
         Args:
@@ -77,7 +77,7 @@ def __init__(
 class OptimizationGradientError(BotorchError, RuntimeError):
     r"""Exception raised when gradient array `gradf` containts NaNs."""
 
-    def __init__(self, /, *args: Any, current_x: np.ndarray, **kwargs: Any) -> None:
+    def __init__(self, /, *args: Any, current_x: npt.NDArray, **kwargs: Any) -> None:
         r"""
         Args:
             *args: Standard args to `BoTorchError`.

botorch/generation/gen.py

@@ -17,6 +17,7 @@
 from typing import Any, NoReturn
 
 import numpy as np
+import numpy.typing as npt
 import torch
 from botorch.acquisition import AcquisitionFunction
 from botorch.exceptions.errors import OptimizationGradientError
@@ -191,7 +192,7 @@ def gen_candidates_scipy(
     with_grad = options.get("with_grad", True)
     if with_grad:
 
-        def f_np_wrapper(x: np.ndarray, f: Callable):
+        def f_np_wrapper(x: npt.NDArray, f: Callable):
             """Given a torch callable, compute value + grad given a numpy array."""
             if np.isnan(x).any():
                 raise RuntimeError(
@@ -223,7 +224,7 @@ def f_np_wrapper(x: np.ndarray, f: Callable):
 
     else:
 
-        def f_np_wrapper(x: np.ndarray, f: Callable):
+        def f_np_wrapper(x: npt.NDArray, f: Callable):
             X = torch.from_numpy(x).to(initial_conditions).view(shapeX).contiguous()
             with torch.no_grad():
                 X_fix = fix_features(X=X, fixed_features=fixed_features)

botorch/models/pairwise_gp.py

@@ -26,6 +26,7 @@
 from typing import Any
 
 import numpy as np
+import numpy.typing as npt
 import torch
 from botorch.acquisition.objective import PosteriorTransform
 from botorch.exceptions import UnsupportedError
@@ -397,13 +398,13 @@ def _prior_predict(self, X: Tensor) -> tuple[Tensor, Tensor]:
 
     def _grad_posterior_f(
         self,
-        utility: Tensor | np.ndarray,
+        utility: Tensor | npt.NDArray,
         datapoints: Tensor,
         D: Tensor,
         covar_chol: Tensor,
         covar_inv: Tensor | None = None,
         ret_np: bool = False,
-    ) -> Tensor | np.ndarray:
+    ) -> Tensor | npt.NDArray:
         r"""Compute the gradient of S loss wrt to f/utility in [Chu2005preference]_.
 
         For finding f_map, which is negative of the log posterior, i.e., -log(p(f|D))
@@ -441,13 +442,13 @@ def _grad_posterior_f(
 
     def _hess_posterior_f(
         self,
-        utility: Tensor | np.ndarray,
+        utility: Tensor | npt.NDArray,
         datapoints: Tensor,
         D: Tensor,
         covar_chol: Tensor,
         covar_inv: Tensor,
         ret_np: bool = False,
-    ) -> Tensor | np.ndarray:
+    ) -> Tensor | npt.NDArray:
         r"""Compute the hessian of S loss wrt utility for finding f_map.
 
         which is negative of the log posterior, i.e., -log(p(f|D))

botorch/optim/closures/core.py

@@ -13,6 +13,8 @@
 from functools import partial
 from typing import Any
 
+import numpy.typing as npt
+
 import torch
 from botorch.optim.utils import (
     _handle_numerical_errors,
@@ -21,7 +23,7 @@
 )
 from botorch.optim.utils.numpy_utils import as_ndarray
 from botorch.utils.context_managers import zero_grad_ctx
-from numpy import float64 as np_float64, full as np_full, ndarray, zeros as np_zeros
+from numpy import float64 as np_float64, full as np_full, zeros as np_zeros
 from torch import Tensor
 
 
@@ -82,10 +84,10 @@ def __init__(
         self,
         closure: Callable[[], tuple[Tensor, Sequence[Tensor | None]]],
         parameters: dict[str, Tensor],
-        as_array: Callable[[Tensor], ndarray] = None,  # pyre-ignore [9]
-        as_tensor: Callable[[ndarray], Tensor] = torch.as_tensor,
-        get_state: Callable[[], ndarray] = None,  # pyre-ignore [9]
-        set_state: Callable[[ndarray], None] = None,  # pyre-ignore [9]
+        as_array: Callable[[Tensor], npt.NDArray] = None,  # pyre-ignore [9]
+        as_tensor: Callable[[npt.NDArray], Tensor] = torch.as_tensor,
+        get_state: Callable[[], npt.NDArray] = None,  # pyre-ignore [9]
+        set_state: Callable[[npt.NDArray], None] = None,  # pyre-ignore [9]
         fill_value: float = 0.0,
         persistent: bool = True,
     ) -> None:
@@ -140,11 +142,11 @@ def __init__(
 
         self.fill_value = fill_value
         self.persistent = persistent
-        self._gradient_ndarray: ndarray | None = None
+        self._gradient_ndarray: npt.NDArray | None = None
 
     def __call__(
-        self, state: ndarray | None = None, **kwargs: Any
-    ) -> tuple[ndarray, ndarray]:
+        self, state: npt.NDArray | None = None, **kwargs: Any
+    ) -> tuple[npt.NDArray, npt.NDArray]:
         if state is not None:
             self.state = state
 
@@ -164,14 +166,14 @@ def __call__(
         return value, grads
 
     @property
-    def state(self) -> ndarray:
+    def state(self) -> npt.NDArray:
         return self._get_state()
 
     @state.setter
-    def state(self, state: ndarray) -> None:
+    def state(self, state: npt.NDArray) -> None:
         self._set_state(state)
 
-    def _get_gradient_ndarray(self, fill_value: float | None = None) -> ndarray:
+    def _get_gradient_ndarray(self, fill_value: float | None = None) -> npt.NDArray:
         if self.persistent and self._gradient_ndarray is not None:
             if fill_value is not None:
                 self._gradient_ndarray.fill(fill_value)

botorch/optim/core.py

@@ -17,10 +17,12 @@
 from time import monotonic
 from typing import Any
 
+import numpy.typing as npt
+
 from botorch.optim.closures import NdarrayOptimizationClosure
 from botorch.optim.utils.numpy_utils import get_bounds_as_ndarray
 from botorch.optim.utils.timeout import minimize_with_timeout
-from numpy import asarray, float64 as np_float64, ndarray
+from numpy import asarray, float64 as np_float64
 from torch import Tensor
 from torch.optim.adam import Adam
 from torch.optim.optimizer import Optimizer
@@ -60,7 +62,7 @@ def scipy_minimize(
     parameters: dict[str, Tensor],
     bounds: dict[str, tuple[float | None, float | None]] | None = None,
     callback: Callable[[dict[str, Tensor], OptimizationResult], None] | None = None,
-    x0: ndarray | None = None,
+    x0: npt.NDArray | None = None,
     method: str = "L-BFGS-B",
     options: dict[str, Any] | None = None,
     timeout_sec: float | None = None,
@@ -98,7 +100,7 @@ def scipy_minimize(
     else:
         call_counter = count(1)  # callbacks are typically made at the end of each iter
 
-        def wrapped_callback(x: ndarray):
+        def wrapped_callback(x: npt.NDArray):
             result = OptimizationResult(
                 step=next(call_counter),
                 fval=float(wrapped_closure(x)[0]),

botorch/optim/parameter_constraints.py

@@ -16,6 +16,7 @@
 from typing import Union
 
 import numpy as np
+import numpy.typing as npt
 import torch
 from botorch.exceptions.errors import CandidateGenerationError, UnsupportedError
 from scipy.optimize import Bounds
@@ -131,7 +132,7 @@ def make_scipy_linear_constraints(
 
 
 def eval_lin_constraint(
-    x: np.ndarray, flat_idxr: list[int], coeffs: np.ndarray, rhs: float
+    x: npt.NDArray, flat_idxr: list[int], coeffs: npt.NDArray, rhs: float
 ) -> np.float64:
     r"""Evaluate a single linear constraint.
 
@@ -148,8 +149,8 @@ def eval_lin_constraint(
 
 
 def lin_constraint_jac(
-    x: np.ndarray, flat_idxr: list[int], coeffs: np.ndarray, n: int
-) -> np.ndarray:
+    x: npt.NDArray, flat_idxr: list[int], coeffs: npt.NDArray, n: int
+) -> npt.NDArray:
     r"""Return the Jacobian associated with a linear constraint.
 
     Args:
@@ -167,7 +168,7 @@ def lin_constraint_jac(
     return jac
 
 
-def _arrayify(X: Tensor) -> np.ndarray:
+def _arrayify(X: Tensor) -> npt.NDArray:
     r"""Convert a torch.Tensor (any dtype or device) to a numpy (double) array.
 
     Args:

botorch/optim/utils/common.py

@@ -14,12 +14,13 @@
 from warnings import warn_explicit, WarningMessage
 
 import numpy as np
+import numpy.typing as npt
 from linear_operator.utils.errors import NanError, NotPSDError
 
 
 def _handle_numerical_errors(
-    error: RuntimeError, x: np.ndarray, dtype: np.dtype | None = None
-) -> tuple[np.ndarray, np.ndarray]:
+    error: RuntimeError, x: npt.NDArray, dtype: np.dtype | None = None
+) -> tuple[npt.NDArray, npt.NDArray]:
     if isinstance(error, NotPSDError):
         raise error
     error_message = error.args[0] if len(error.args) > 0 else ""

botorch/optim/utils/numpy_utils.py

@@ -13,8 +13,8 @@
 from itertools import tee
 
 import numpy as np
+import numpy.typing as npt
 import torch
-from numpy import ndarray
 from torch import Tensor
 
 
@@ -35,7 +35,7 @@
 
 def as_ndarray(
     values: Tensor, dtype: np.dtype | None = None, inplace: bool = True
-) -> ndarray:
+) -> npt.NDArray:
     r"""Helper for going from torch.Tensor to numpy.ndarray.
 
     Args:
@@ -67,10 +67,10 @@ def as_ndarray(
 
 def get_tensors_as_ndarray_1d(
     tensors: Iterator[Tensor] | dict[str, Tensor],
-    out: ndarray | None = None,
+    out: npt.NDArray | None = None,
     dtype: np.dtype | str | None = None,
-    as_array: Callable[[Tensor], ndarray] = as_ndarray,
-) -> ndarray:
+    as_array: Callable[[Tensor], npt.NDArray] = as_ndarray,
+) -> npt.NDArray:
     # Create a pair of iterators, one for setup and one for data transfer
     named_tensors_iter, named_tensors_iter2 = tee(
         iter(tensors.items()) if isinstance(tensors, dict) else enumerate(tensors), 2
@@ -112,8 +112,8 @@ def get_tensors_as_ndarray_1d(
 
 def set_tensors_from_ndarray_1d(
     tensors: Iterator[Tensor] | dict[str, Tensor],
-    array: ndarray,
-    as_tensor: Callable[[ndarray], Tensor] = torch.as_tensor,
+    array: npt.NDArray,
+    as_tensor: Callable[[npt.NDArray], Tensor] = torch.as_tensor,
 ) -> None:
     r"""Sets the values of one more tensors based off of a vector of assignments."""
     named_tensors_iter = (
@@ -137,7 +137,7 @@ def set_tensors_from_ndarray_1d(
 def get_bounds_as_ndarray(
     parameters: dict[str, Tensor],
     bounds: dict[str, tuple[float | Tensor | None, float | Tensor | None]],
-) -> np.ndarray | None:
+) -> npt.NDArray | None:
     r"""Helper method for converting bounds into an ndarray.
 
     Args:

botorch/optim/utils/timeout.py

@@ -11,14 +11,14 @@
 from collections.abc import Callable, Sequence
 from typing import Any
 
-import numpy as np
+import numpy.typing as npt
 from botorch.exceptions.errors import OptimizationTimeoutError
 from scipy import optimize
 
 
 def minimize_with_timeout(
-    fun: Callable[[np.ndarray, ...], float],
-    x0: np.ndarray,
+    fun: Callable[[npt.NDArray, ...], float],
+    x0: npt.NDArray,
     args: tuple[Any, ...] = (),
     method: str | None = None,
     jac: str | Callable | bool | None = None,
@@ -45,7 +45,7 @@ def minimize_with_timeout(
         start_time = time.monotonic()
         callback_data = {"num_iterations": 0}  # update from withing callback below
 
-        def timeout_callback(xk: np.ndarray) -> bool:
+        def timeout_callback(xk: npt.NDArray) -> bool:
             runtime = time.monotonic() - start_time
             callback_data["num_iterations"] += 1
             if runtime > timeout_sec:
@@ -63,14 +63,14 @@ def timeout_callback(xk: np.ndarray) -> bool:
         elif method == "trust-constr":  # special signature
 
             def wrapped_callback(
-                xk: np.ndarray, state: optimize.OptimizeResult
+                xk: npt.NDArray, state: optimize.OptimizeResult
             ) -> bool:
                 # order here is important to make sure base callback gets executed
                 return callback(xk, state) or timeout_callback(xk=xk)
 
         else:
 
-            def wrapped_callback(xk: np.ndarray) -> None:
+            def wrapped_callback(xk: npt.NDArray) -> None:
                 timeout_callback(xk=xk)
                 callback(xk)
 

botorch/utils/sampling.py

@@ -24,6 +24,7 @@
 from typing import Any, TYPE_CHECKING
 
 import numpy as np
+import numpy.typing as npt
 import scipy
 import torch
 from botorch.exceptions.errors import BotorchError
@@ -374,11 +375,11 @@ def _convert_bounds_to_inequality_constraints(bounds: Tensor) -> tuple[Tensor, T
 
 
 def find_interior_point(
-    A: np.ndarray,
-    b: np.ndarray,
-    A_eq: np.ndarray | None = None,
-    b_eq: np.ndarray | None = None,
-) -> np.ndarray:
+    A: npt.NDArray,
+    b: npt.NDArray,
+    A_eq: npt.NDArray | None = None,
+    b_eq: npt.NDArray | None = None,
+) -> npt.NDArray:
     r"""Find an interior point of a polytope via linear programming.
 
     Args:

test/optim/test_parameter_constraints.py

@@ -8,6 +8,7 @@
 from itertools import product
 
 import numpy as np
+import numpy.typing as npt
 import torch
 from botorch.exceptions.errors import CandidateGenerationError, UnsupportedError
 from botorch.optim.parameter_constraints import (
@@ -55,7 +56,7 @@ def test_make_nonlinear_constraints(self):
         def nlc(x):
             return 4 - x.sum()
 
-        def f_np_wrapper(x: np.ndarray, f: Callable):
+        def f_np_wrapper(x: npt.NDArray, f: Callable):
             """Given a torch callable, compute value + grad given a numpy array."""
             X = (
                 torch.from_numpy(x)
@@ -114,7 +115,7 @@ def test_make_scipy_nonlinear_inequality_constraints(self):
         def nlc(x):
             return 4 - x.sum()
 
-        def f_np_wrapper(x: np.ndarray, f: Callable):
+        def f_np_wrapper(x: npt.NDArray, f: Callable):
             """Given a torch callable, compute value + grad given a numpy array."""
             X = (
                 torch.from_numpy(x)