Changes for numpy 2.0 deprecations

- Replace np.cast with np.asarray: in numpy 2.0,
  `np.cast[new_dtype](arr)` is deprecated.
  The literal replacement is `np.asarray(arr, dtype=new_dtype)`.

- Replace np.sctype2char and np.obj2sctype.
  Added try/except to handle change in behavior
  of `np.dtype`

- Replace np.find_common_type with np.result_type

Further changes to `TensorType`:

TensorType.dtype must be a string, so the code
has been changed from `self.dtype = np.dtype(dtype).type`,
where the right-hand side is of type `np.generic`, to
`self.dtype = str(np.dtype(dtype))`, where the right-hand
side is a string that satisfies:

`self.dtype == str(np.dtype(self.dtype))`

This doesn't change the behavior of `np.array(..., dtype=self.dtype)`
etc.

Author: Armavica

pytensor/scalar/basic.py

@@ -2966,7 +2966,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / (x * np.asarray(math.log(2.0)).astype(x.dtype)),)
+        return (gz / (x * np.array(math.log(2.0), dtype=x.dtype)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3009,7 +3009,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / (x * np.asarray(math.log(10.0)).astype(x.dtype)),)
+        return (gz / (x * np.array(math.log(10.0), dtype=x.dtype)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3124,7 +3124,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz * exp2(x) * log(np.cast[x.type](2)),)
+        return (gz * exp2(x) * log(np.array(2, dtype=x.type)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3263,7 +3263,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz * np.asarray(np.pi / 180, gz.type),)
+        return (gz * np.array(np.pi / 180, dtype=gz.type),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3298,7 +3298,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz * np.asarray(180.0 / np.pi, gz.type),)
+        return (gz * np.array(180.0 / np.pi, dtype=gz.type),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3371,7 +3371,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (-gz / sqrt(np.cast[x.type](1) - sqr(x)),)
+        return (-gz / sqrt(np.array(1, dtype=x.type) - sqr(x)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3445,7 +3445,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / sqrt(np.cast[x.type](1) - sqr(x)),)
+        return (gz / sqrt(np.array(1, dtype=x.type) - sqr(x)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3517,7 +3517,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / (np.cast[x.type](1) + sqr(x)),)
+        return (gz / (np.array(1, dtype=x.type) + sqr(x)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3640,7 +3640,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / sqrt(sqr(x) - np.cast[x.type](1)),)
+        return (gz / sqrt(sqr(x) - np.array(1, dtype=x.type)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3717,7 +3717,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / sqrt(sqr(x) + np.cast[x.type](1)),)
+        return (gz / sqrt(sqr(x) + np.array(1, dtype=x.type)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs
@@ -3795,7 +3795,7 @@ def L_op(self, inputs, outputs, gout):
             else:
                 return [x.zeros_like()]
 
-        return (gz / (np.cast[x.type](1) - sqr(x)),)
+        return (gz / (np.array(1, dtype=x.type) - sqr(x)),)
 
     def c_code(self, node, name, inputs, outputs, sub):
         (x,) = inputs

pytensor/tensor/elemwise.py

@@ -668,7 +668,7 @@ def prepare_node(self, node, storage_map, compute_map, impl):
             and isinstance(self.nfunc, np.ufunc)
             and node.inputs[0].dtype in discrete_dtypes
         ):
-            char = np.sctype2char(out_dtype)
+            char = np.dtype(out_dtype).char
             sig = char * node.nin + "->" + char * node.nout
             node.tag.sig = sig
         node.tag.fake_node = Apply(

pytensor/tensor/type.py

@@ -4,6 +4,7 @@
 from typing import TYPE_CHECKING, Literal, Optional
 
 import numpy as np
+import numpy.typing as npt
 
 import pytensor
 from pytensor import scalar as ps
@@ -69,7 +70,7 @@ class TensorType(CType[np.ndarray], HasDataType, HasShape):
 
     def __init__(
         self,
-        dtype: str | np.dtype,
+        dtype: str | npt.DTypeLike,
         shape: Iterable[bool | int | None] | None = None,
         name: str | None = None,
         broadcastable: Iterable[bool] | None = None,
@@ -101,11 +102,11 @@ def __init__(
         if str(dtype) == "floatX":
             self.dtype = config.floatX
         else:
-            if np.obj2sctype(dtype) is None:
+            try:
+                self.dtype = str(np.dtype(dtype))
+            except TypeError:
                 raise TypeError(f"Invalid dtype: {dtype}")
 
-            self.dtype = np.dtype(dtype).name
-
         def parse_bcast_and_shape(s):
             if isinstance(s, bool | np.bool_):
                 return 1 if s else None
@@ -789,14 +790,16 @@ def tensor(
     **kwargs,
 ) -> "TensorVariable":
     if name is not None:
-        # Help catching errors with the new tensor API
-        # Many single letter strings are valid sctypes
-        if str(name) == "floatX" or (len(str(name)) > 1 and np.obj2sctype(name)):
-            np.obj2sctype(name)
-            raise ValueError(
-                f"The first and only positional argument of tensor is now `name`. Got {name}.\n"
-                "This name looks like a dtype, which you should pass as a keyword argument only."
-            )
+        try:
+            # Help catching errors with the new tensor API
+            # Many single letter strings are valid sctypes
+            if str(name) == "floatX" or (len(str(name)) > 1 and np.dtype(name).type):
+                raise ValueError(
+                    f"The first and only positional argument of tensor is now `name`. Got {name}.\n"
+                    "This name looks like a dtype, which you should pass as a keyword argument only."
+                )
+        except TypeError:
+            pass
 
     if dtype is None:
         dtype = config.floatX

tests/scan/test_rewriting.py

@@ -673,7 +673,7 @@ def test_machine_translation(self):
         zi = tensor3("zi")
         zi_value = x_value
 
-        init = pt.alloc(np.cast[config.floatX](0), batch_size, dim)
+        init = pt.alloc(np.asarray(0, dtype=config.floatX), batch_size, dim)
 
         def rnn_step1(
             # sequences

tests/tensor/test_extra_ops.py

@@ -708,7 +708,7 @@ def test_perform(self, shp):
         y = scalar()
         f = function([x, y], fill_diagonal(x, y))
         a = rng.random(shp).astype(config.floatX)
-        val = np.cast[config.floatX](rng.random())
+        val = rng.random(dtype=config.floatX)
         out = f(a, val)
         # We can't use np.fill_diagonal as it is bugged.
         assert np.allclose(np.diag(out), val)
@@ -720,7 +720,7 @@ def test_perform_3d(self):
         x = tensor3()
         y = scalar()
         f = function([x, y], fill_diagonal(x, y))
-        val = np.cast[config.floatX](rng.random() + 10)
+        val = rng.random(dtype=config.floatX) + 10
         out = f(a, val)
         # We can't use np.fill_diagonal as it is bugged.
         assert out[0, 0, 0] == val
@@ -782,7 +782,7 @@ def test_perform(self, test_offset, shp):
 
         f = function([x, y, z], fill_diagonal_offset(x, y, z))
         a = rng.random(shp).astype(config.floatX)
-        val = np.cast[config.floatX](rng.random())
+        val = rng.random(dtype=config.floatX)
         out = f(a, val, test_offset)
         # We can't use np.fill_diagonal as it is bugged.
         assert np.allclose(np.diag(out, test_offset), val)

tests/tensor/utils.py

@@ -152,7 +152,7 @@ def upcast_float16_ufunc(fn):
     """
 
     def ret(*args, **kwargs):
-        out_dtype = np.find_common_type([a.dtype for a in args], [np.float16])
+        out_dtype = np.result_type(np.float16, *args)
         if out_dtype == "float16":
             # Force everything to float32
             sig = "f" * fn.nin + "->" + "f" * fn.nout

tests/test_gradient.py

@@ -481,12 +481,12 @@ def make_grad_func(X):
         int_type = imatrix().dtype
         float_type = "float64"
 
-        X = np.cast[int_type](rng.standard_normal((m, d)) * 127.0)
-        W = np.cast[W.dtype](rng.standard_normal((d, n)))
-        b = np.cast[b.dtype](rng.standard_normal(n))
+        X = np.asarray(rng.standard_normal((m, d)) * 127.0, dtype=int_type)
+        W = rng.standard_normal((d, n), dtype=W.dtype)
+        b = rng.standard_normal(n, dtype=b.dtype)
 
         int_result = int_func(X, W, b)
-        float_result = float_func(np.cast[float_type](X), W, b)
+        float_result = float_func(np.asarray(X, dtype=float_type), W, b)
 
         assert np.allclose(int_result, float_result), (int_result, float_result)
 
@@ -508,7 +508,7 @@ def test_grad_disconnected(self):
         # the output
         f = pytensor.function([x], g)
         rng = np.random.default_rng([2012, 9, 5])
-        x = np.cast[x.dtype](rng.standard_normal(3))
+        x = rng.standard_normal(3, dtype=x.dtype)
         g = f(x)
         assert np.allclose(g, np.ones(x.shape, dtype=x.dtype))
 
@@ -631,7 +631,8 @@ def test_known_grads():
     rng = np.random.default_rng([2012, 11, 15])
     values = [rng.standard_normal(10), rng.integers(10), rng.standard_normal()]
     values = [
-        np.cast[ipt.dtype](value) for ipt, value in zip(inputs, values, strict=True)
+        np.asarray(value, dtype=ipt.dtype)
+        for ipt, value in zip(inputs, values, strict=True)
     ]
 
     true_grads = grad(cost, inputs, disconnected_inputs="ignore")
@@ -679,7 +680,7 @@ def test_known_grads_integers():
     f = pytensor.function([g_expected], g_grad)
 
     x = -3
-    gv = np.cast[config.floatX](0.6)
+    gv = np.asarray(0.6, dtype=config.floatX)
 
     g_actual = f(gv)
 
@@ -746,7 +747,8 @@ def test_subgraph_grad():
     rng = np.random.default_rng([2012, 11, 15])
     values = [rng.standard_normal(2), rng.standard_normal(3)]
     values = [
-        np.cast[ipt.dtype](value) for ipt, value in zip(inputs, values, strict=True)
+        np.asarray(value, dtype=ipt.dtype)
+        for ipt, value in zip(inputs, values, strict=True)
     ]
 
     wrt = [w2, w1]
@@ -1031,30 +1033,30 @@ def test_jacobian_scalar():
     # test when the jacobian is called with a tensor as wrt
     Jx = jacobian(y, x)
     f = pytensor.function([x], Jx)
-    vx = np.cast[pytensor.config.floatX](rng.uniform())
+    vx = np.asarray(rng.uniform(), dtype=pytensor.config.floatX)
     assert np.allclose(f(vx), 2)
 
     # test when the jacobian is called with a tuple as wrt
     Jx = jacobian(y, (x,))
     assert isinstance(Jx, tuple)
     f = pytensor.function([x], Jx[0])
-    vx = np.cast[pytensor.config.floatX](rng.uniform())
+    vx = np.asarray(rng.uniform(), dtype=pytensor.config.floatX)
     assert np.allclose(f(vx), 2)
 
     # test when the jacobian is called with a list as wrt
     Jx = jacobian(y, [x])
     assert isinstance(Jx, list)
     f = pytensor.function([x], Jx[0])
-    vx = np.cast[pytensor.config.floatX](rng.uniform())
+    vx = np.asarray(rng.uniform(), dtype=pytensor.config.floatX)
     assert np.allclose(f(vx), 2)
 
     # test when the jacobian is called with a list of two elements
     z = scalar()
     y = x * z
     Jx = jacobian(y, [x, z])
     f = pytensor.function([x, z], Jx)
-    vx = np.cast[pytensor.config.floatX](rng.uniform())
-    vz = np.cast[pytensor.config.floatX](rng.uniform())
+    vx = np.asarray(rng.uniform(), dtype=pytensor.config.floatX)
+    vz = np.asarray(rng.uniform(), dtype=pytensor.config.floatX)
     vJx = f(vx, vz)
 
     assert np.allclose(vJx[0], vz)

tests/typed_list/test_basic.py

@@ -577,10 +577,10 @@ def test_correct_answer(self):
         x = tensor3()
         y = tensor3()
 
-        A = np.cast[pytensor.config.floatX](np.random.random((5, 3)))
-        B = np.cast[pytensor.config.floatX](np.random.random((7, 2)))
-        X = np.cast[pytensor.config.floatX](np.random.random((5, 6, 1)))
-        Y = np.cast[pytensor.config.floatX](np.random.random((1, 9, 3)))
+        A = np.random.random((5, 3)).astype(pytensor.config.floatX)
+        B = np.random.random((7, 2)).astype(pytensor.config.floatX)
+        X = np.random.random((5, 6, 1)).astype(pytensor.config.floatX)
+        Y = np.random.random((1, 9, 3)).astype(pytensor.config.floatX)
 
         make_list((3.0, 4.0))
         c = make_list((a, b))