Skip to content

Add rewrite for softplus(log(x)) -> log1p(x) #1452

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jul 30, 2025
Merged

Add rewrite for softplus(log(x)) -> log1p(x) #1452

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
10bcb41
Add rewrite for `softplus(log(x)) -> log1p(x)`
Jun 6, 2025
67d432d
Added log1mexp(log(x)) -> log1p(-x) and its test
Jun 23, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion pytensor/tensor/rewriting/math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -400,7 +400,7 @@ def local_exp_log(fgraph, node):


@register_specialize
@node_rewriter([exp, expm1])
@node_rewriter([exp, expm1, softplus])
def local_exp_log_nan_switch(fgraph, node):
# Rewrites of the kind exp(log...(x)) that require a `nan` switch
x = node.inputs[0]
Expand Down Expand Up @@ -453,6 +453,13 @@ def local_exp_log_nan_switch(fgraph, node):
new_out = switch(le(x, 0), neg(exp(x)), np.asarray(np.nan, old_out.dtype))
return [new_out]

# Case for softplus(log(x)) -> log1p(x)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nitpick I prefer to refer to it by log1pexp, which we have as an alias to softplus:

pytensor/pytensor/tensor/math.py

Line 2474 in ff98ab8

log1pexp = softplus

Also we can add a similar case for log1mexp?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I tested this (with the code below) and it works fine in its domain [0, 1]. I will add it too.

if isinstance(prev_op, ps.Log) and isinstance(node_op, ps_math.Softplus):
x = x.owner.inputs[0]
old_out = node.outputs[0]
new_out = switch(ge(x, 0), log1p(x), np.asarray(np.nan, old_out.dtype))
return [new_out]


@register_canonicalize
@register_specialize
Expand Down
21 changes: 21 additions & 0 deletions tests/tensor/rewriting/test_math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2010,6 +2010,27 @@ def test_exp_softplus(self, exp_op):
decimal=6,
)

def test_softplus_log(self):
# softplus(log(x)) -> log1p(x)
data_valid = np.random.random((4, 3)).astype("float32") * 2
data_valid[0, 0] = 0 # edge case
data_invalid = data_valid - 2

x = fmatrix()
f = function([x], softplus(log(x)), mode=self.mode)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if you want you can check against the expected graph directly, something like

assert equal_computations(f.maker.fgraph.outputs, [pt.switch(x > 0, pt.log1p(x), np.asarray([[np.nan]], dtype="float32")])

Or something like that. This is not a request!

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It took me a while to figure out how to make the test work.

It works fine if I apply it to a scalar rewritten output, like:

x = pt.scalar("x")
out = pt.softplus(pt.log(x))
new_out = rewrite_graph(out, include=("canonicalize", "stabilize", "specialize"))
equal_computations([new_out], [pt.switch(x >= 0, pt.log1p(x), pt.nan)])

But if I want to apply within the test (where the function is also applied to test data), I need to fix a few things.
First of all, the constants have to be matrices of the right type: int8 and float32, respectively (which was not obvious to me).
Then, for some reason, I could only make it work after overriding the compile mode:

        x = fmatrix()
        mode=get_mode('FAST_COMPILE').including("local_exp_log", "local_exp_log_nan_switch")
        f = function([x], softplus(log(x)), mode=mode)
        assert equal_computations(f.maker.fgraph.outputs, [pt.switch(x >= np.array([[0]], dtype=np.int8), pt.log1p(x), np.array([[np.nan]], dtype=np.float32))])

For some reason the mode used in the test (set at class level) seems to do something extra and equal_computations returns false, possibly due to:

nd_x.op: Elemwise(scalar_op=Switch,inplace_pattern=<frozendict {0: 1}>)
nd_y.op: Elemwise(scalar_op=Switch,inplace_pattern=<frozendict {}>)

I don't know enough of the internals of PyTensor to find a solution, except forcing the mode above inside this test.
Is it fine to override the class compile mode with the compile mode above?

Copy link
Member

@ricardoV94 ricardoV94 Jun 12, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah nvm. I usually use it with just the output of rewrite_graph with a subset of rewrites.

You're seeing inplace there

Not worth the trouble here.

graph = f.maker.fgraph.toposort()
ops_graph = [
node
for node in graph
if isinstance(node.op, Elemwise)
and isinstance(node.op.scalar_op, ps.Log | ps.Exp | ps.Softplus)
]
assert len(ops_graph) == 0

expected = np.log1p(data_valid)
np.testing.assert_almost_equal(f(data_valid), expected)
assert np.all(np.isnan(f(data_invalid)))

@pytest.mark.parametrize(
["nested_expression", "expected_switches"],
[
Expand Down