move linalg rewrites, delete sandbox

Moving caused a circular dependency with tensor.blas. It seems most
linalg rewrites are in the stablize set, so should run before the blas
specializers anyway, so these checks were removed.

This also deleted the unused `spectral_radius_bound` and dummy
`Minimal(Op)`.

Author: dehorsley

pytensor/sandbox/__init__.py

@@ -4,6 +4,7 @@
 
 # Register JAX specializations
 import pytensor.tensor.rewriting.jax
+import pytensor.tensor.rewriting.linalg
 import pytensor.tensor.rewriting.math
 import pytensor.tensor.rewriting.shape
 import pytensor.tensor.rewriting.special

pytensor/sandbox/linalg/__init__.py

@@ -2,7 +2,6 @@
 
 from pytensor.graph.rewriting.basic import node_rewriter
 from pytensor.tensor import basic as at
-from pytensor.tensor.blas import Dot22
 from pytensor.tensor.elemwise import DimShuffle
 from pytensor.tensor.math import Dot, Prod, dot, log
 from pytensor.tensor.math import pow as at_pow
@@ -32,12 +31,12 @@ def transinv_to_invtrans(fgraph, node):
 
 
 @register_stabilize
-@node_rewriter([Dot, Dot22])
+@node_rewriter([Dot])
 def inv_as_solve(fgraph, node):
     """
     This utilizes a boolean `symmetric` tag on the matrices.
     """
-    if isinstance(node.op, (Dot, Dot22)):
+    if isinstance(node.op, Dot):
         l, r = node.inputs
         if l.owner and isinstance(l.owner.op, MatrixInverse):
             return [solve(l.owner.inputs[0], r)]
@@ -123,7 +122,7 @@ def cholesky_ldotlt(fgraph, node):
         return
 
     A = node.inputs[0]
-    if not (A.owner and isinstance(A.owner.op, (Dot, Dot22))):
+    if not (A.owner and isinstance(A.owner.op, Dot)):
         return
 
     l, r = A.owner.inputs

pytensor/sandbox/minimal.py

@@ -8,10 +8,10 @@
 from pytensor import tensor as at
 from pytensor.compile import get_default_mode
 from pytensor.configdefaults import config
-from pytensor.sandbox.linalg.ops import inv_as_solve, spectral_radius_bound
 from pytensor.tensor.elemwise import DimShuffle
 from pytensor.tensor.math import _allclose
 from pytensor.tensor.nlinalg import MatrixInverse, matrix_inverse
+from pytensor.tensor.rewriting.linalg import inv_as_solve
 from pytensor.tensor.slinalg import Cholesky, Solve, solve
 from pytensor.tensor.type import dmatrix, matrix, vector
 from tests import unittest_tools as utt
@@ -68,53 +68,6 @@ def test_rop_lop():
     assert _allclose(v1, v2), f"LOP mismatch: {v1} {v2}"
 
 
-def test_spectral_radius_bound():
-    tol = 10 ** (-6)
-    rng = np.random.default_rng(utt.fetch_seed())
-    x = matrix()
-    radius_bound = spectral_radius_bound(x, 5)
-    f = pytensor.function([x], radius_bound)
-
-    shp = (3, 4)
-    m = rng.random(shp)
-    m = np.cov(m).astype(config.floatX)
-    radius_bound_pytensor = f(m)
-
-    # test the approximation
-    mm = m
-    for i in range(5):
-        mm = np.dot(mm, mm)
-    radius_bound_numpy = np.trace(mm) ** (2 ** (-5))
-    assert abs(radius_bound_numpy - radius_bound_pytensor) < tol
-
-    # test the bound
-    eigen_val = numpy.linalg.eig(m)
-    assert (eigen_val[0].max() - radius_bound_pytensor) < tol
-
-    # test type errors
-    xx = vector()
-    ok = False
-    try:
-        spectral_radius_bound(xx, 5)
-    except TypeError:
-        ok = True
-    assert ok
-    ok = False
-    try:
-        spectral_radius_bound(x, 5.0)
-    except TypeError:
-        ok = True
-    assert ok
-
-    # test value error
-    ok = False
-    try:
-        spectral_radius_bound(x, -5)
-    except ValueError:
-        ok = True
-    assert ok
-
-
 def test_transinv_to_invtrans():
     X = matrix("X")
     Y = matrix_inverse(X)