Implement Dot and BatchedDot in PyTensor

pytensor/link/__init__.py

@@ -0,0 +1 @@
+from pytensor.link.pytorch.linker import PytorchLinker

pytensor/link/pytorch/dispatch/__init__.py

@@ -2,8 +2,11 @@
 from pytensor.link.pytorch.dispatch.basic import pytorch_funcify, pytorch_typify
 
 # # Load dispatch specializations
+import pytensor.link.pytorch.dispatch.blas
 import pytensor.link.pytorch.dispatch.scalar
 import pytensor.link.pytorch.dispatch.elemwise
+import pytensor.link.pytorch.dispatch.math
 import pytensor.link.pytorch.dispatch.extra_ops
 import pytensor.link.pytorch.dispatch.sort
+
 # isort: on

pytensor/link/pytorch/dispatch/blas.py

@@ -0,0 +1,14 @@
+import torch
+
+from pytensor.link.pytorch.dispatch import pytorch_funcify
+from pytensor.tensor.blas import BatchedDot
+
+
+@pytorch_funcify.register(BatchedDot)
+def pytorch_funcify_BatchedDot(op, **kwargs):
+    def batched_dot(a, b):
+        if a.shape[0] != b.shape[0]:
+            raise TypeError("Shapes must match in the 0-th dimension")
+        return torch.bmm(a, b)
+
+    return batched_dot

pytensor/link/pytorch/dispatch/math.py

@@ -0,0 +1,12 @@
+import torch
+
+from pytensor.link.pytorch.dispatch import pytorch_funcify
+from pytensor.tensor.math import Dot
+
+
+@pytorch_funcify.register(Dot)
+def pytorch_funcify_Dot(op, **kwargs):
+    def dot(x, y):
+        return torch.matmul(x, y)
+
+    return dot

tests/link/pytorch/test_blas.py

@@ -0,0 +1,35 @@
+import numpy as np
+import pytest
+
+from pytensor.compile.mode import Mode
+from pytensor.configdefaults import config
+from pytensor.graph.fg import FunctionGraph
+from pytensor.graph.op import get_test_value
+from pytensor.link.pytorch.linker import PytorchLinker
+from pytensor.tensor import blas as pt_blas
+from pytensor.tensor.type import tensor3
+from tests.link.pytorch.test_basic import compare_pytorch_and_py
+
+
+def test_pytorch_BatchedDot():
+    # tensor3 . tensor3
+    a = tensor3("a")
+    a.tag.test_value = (
+        np.linspace(-1, 1, 10 * 5 * 3).astype(config.floatX).reshape((10, 5, 3))
+    )
+    b = tensor3("b")
+    b.tag.test_value = (
+        np.linspace(1, -1, 10 * 3 * 2).astype(config.floatX).reshape((10, 3, 2))
+    )
+    out = pt_blas.BatchedDot()(a, b)
+    fgraph = FunctionGraph([a, b], [out])
+    pytensor_pytorch_fn, _ = compare_pytorch_and_py(
+        fgraph, [get_test_value(i) for i in fgraph.inputs]
+    )
+
+    # A dimension mismatch should raise a TypeError for compatibility
+    inputs = [get_test_value(a)[:-1], get_test_value(b)]
+    pytorch_mode_no_rewrites = Mode(PytorchLinker(), None)
+    pytensor_pytorch_fn.mode = pytorch_mode_no_rewrites
+    with pytest.raises(TypeError):
+        pytensor_pytorch_fn(*inputs)

tests/link/pytorch/test_math.py

@@ -0,0 +1,30 @@
+import numpy as np
+
+from pytensor.configdefaults import config
+from pytensor.graph.fg import FunctionGraph
+from pytensor.graph.op import get_test_value
+from pytensor.tensor.type import matrix, scalar, vector
+from tests.link.pytorch.test_basic import compare_pytorch_and_py
+
+
+def test_pytorch_dot():
+    y = vector("y")
+    y.tag.test_value = np.r_[1.0, 2.0].astype(config.floatX)
+    x = vector("x")
+    x.tag.test_value = np.r_[3.0, 4.0].astype(config.floatX)
+    A = matrix("A")
+    A.tag.test_value = np.array([[6, 3], [3, 0]], dtype=config.floatX)
+    alpha = scalar("alpha")
+    alpha.tag.test_value = np.array(3.0, dtype=config.floatX)
+    beta = scalar("beta")
+    beta.tag.test_value = np.array(5.0, dtype=config.floatX)
+
+    # 2D * 2D
+    out = A.dot(A * alpha) + beta * A
+    fgraph = FunctionGraph([A, alpha, beta], [out])
+    compare_pytorch_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
+
+    # 1D * 2D and 1D * 1D
+    out = y.dot(alpha * A).dot(x) + beta * y
+    fgraph = FunctionGraph([y, x, A, alpha, beta], [out])
+    compare_pytorch_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])