Torch dispatch for scipy-like functions and Softplus (#1066)

pytensor/link/pytorch/dispatch/elemwise.py

+import importlib
+
 import torch
 
 from pytensor.link.pytorch.dispatch.basic import pytorch_funcify
@@ -11,12 +13,26 @@ def pytorch_funcify_Elemwise(op, node, **kwargs):
     scalar_op = op.scalar_op
     base_fn = pytorch_funcify(scalar_op, node=node, **kwargs)
 
-    if hasattr(scalar_op, "nfunc_spec") and hasattr(torch, scalar_op.nfunc_spec[0]):
+    def check_special_scipy(func_name):
+        if "scipy." not in func_name:
+            return False
+        loc = func_name.split(".")[1:]
+        try:
+            mod = importlib.import_module(".".join(loc[:-1]), "torch")
+            return getattr(mod, loc[-1], False)
+        except ImportError:
+            return False
+
+    if hasattr(scalar_op, "nfunc_spec") and (
+        hasattr(torch, scalar_op.nfunc_spec[0])
+        or check_special_scipy(scalar_op.nfunc_spec[0])
+    ):
         # torch can handle this scalar
         # broadcast, we'll let it.
         def elemwise_fn(*inputs):
             Elemwise._check_runtime_broadcast(node, inputs)
             return base_fn(*inputs)
+
     else:
 
         def elemwise_fn(*inputs):

pytensor/link/pytorch/dispatch/scalar.py

+import importlib
+
 import torch
 
 from pytensor.link.pytorch.dispatch.basic import pytorch_funcify
@@ -5,6 +7,7 @@ from pytensor.scalar.basic import (
     Cast,
     ScalarOp,
 )
+from pytensor.scalar.math import Softplus
 
 
 @pytorch_funcify.register(ScalarOp)
@@ -19,9 +22,14 @@ def pytorch_funcify_ScalarOp(op, node, **kwargs):
     if nfunc_spec is None:
         raise NotImplementedError(f"Dispatch not implemented for Scalar Op {op}")
 
-    func_name = nfunc_spec[0]
+    func_name = nfunc_spec[0].replace("scipy.", "")
 
-    pytorch_func = getattr(torch, func_name)
+    if "." in func_name:
+        loc = func_name.split(".")
+        mod = importlib.import_module(".".join(["torch", *loc[:-1]]))
+        pytorch_func = getattr(mod, loc[-1])
+    else:
+        pytorch_func = getattr(torch, func_name)
 
     if len(node.inputs) > op.nfunc_spec[1]:
         # Some Scalar Ops accept multiple number of inputs, behaving as a variadic function,
@@ -49,3 +57,8 @@ def pytorch_funcify_Cast(op: Cast, node, **kwargs):
         return x.to(dtype=dtype)
 
     return cast
+
+
+@pytorch_funcify.register(Softplus)
+def pytorch_funcify_Softplus(op, node, **kwargs):
+    return torch.nn.Softplus()

tests/link/pytorch/test_basic.py

@@ -17,7 +17,7 @@ from pytensor.graph.op import Op
 from pytensor.ifelse import ifelse
 from pytensor.link.pytorch.linker import PytorchLinker
 from pytensor.raise_op import CheckAndRaise
-from pytensor.tensor import alloc, arange, as_tensor, empty, eye
+from pytensor.tensor import alloc, arange, as_tensor, empty, expit, eye, softplus
 from pytensor.tensor.type import matrices, matrix, scalar, vector
 
 
@@ -374,3 +374,17 @@ def test_pytorch_link_references():
     f = function([x], out, mode="PYTORCH")
     f(torch.ones(3))
     assert "inner_fn" not in dir(m), "function call reference leaked"
+
+
+def test_pytorch_scipy():
+    x = vector("a", shape=(3,))
+    out = expit(x)
+    f = FunctionGraph([x], [out])
+    compare_pytorch_and_py(f, [np.random.rand(3)])
+
+
+def test_pytorch_softplus():
+    x = vector("a", shape=(3,))
+    out = softplus(x)
+    f = FunctionGraph([x], [out])
+    compare_pytorch_and_py(f, [np.random.rand(3)])