Improve logsumexp to work with infinite values (#4360)

* Make logsumexp work with inifinite values, matching scipy behavior

* Run pre-commit

* Add note to release_notes

RELEASE-NOTES.md

@@ -13,7 +13,7 @@ This is the first release to support Python3.9 and to drop Python3.6.
 - Removed `theanof.set_theano_config` because it illegally touched Theano's privates (see [#4329](https://github.com/pymc-devs/pymc3/pull/4329)).
 - In `sample_posterior_predictive` the `vars` kwarg was removed in favor of `var_names` (see [#4343](https://github.com/pymc-devs/pymc3/pull/4343)).
 - The notebook gallery has been moved to https://github.com/pymc-devs/pymc-examples (see [#4348](https://github.com/pymc-devs/pymc3/pull/4348)).
-
+- `math.logsumexp` now matches `scipy.special.logsumexp` when arrays contain infinite values (see [#4360](https://github.com/pymc-devs/pymc3/pull/4360)).
 
 ## PyMC3 3.10.0 (7 December 2020)
 

pymc3/math.py

@@ -175,6 +175,7 @@ def tround(*args, **kwargs):
 def logsumexp(x, axis=None, keepdims=True):
     # Adapted from https://github.com/Theano/Theano/issues/1563
     x_max = tt.max(x, axis=axis, keepdims=True)
+    x_max = tt.switch(tt.isinf(x_max), 0, x_max)
     res = tt.log(tt.sum(tt.exp(x - x_max), axis=axis, keepdims=True)) + x_max
     return res if keepdims else res.squeeze()
 

pymc3/tests/test_math.py

@@ -18,6 +18,8 @@
 import theano
 import theano.tensor as tt
 
+from scipy.special import logsumexp as scipy_logsumexp
+
 from pymc3.math import (
     LogDet,
     cartesian,
@@ -30,6 +32,7 @@
     log1mexp_numpy,
     log1pexp,
     logdet,
+    logsumexp,
     probit,
 )
 from pymc3.tests.helpers import SeededTest, verify_grad
@@ -207,3 +210,27 @@ def test_expand_packed_triangular():
     assert np.all(expand_upper.eval({packed: upper_packed}) == upper)
     assert np.all(expand_diag_lower.eval({packed: lower_packed}) == floatX(np.diag(vals)))
     assert np.all(expand_diag_upper.eval({packed: upper_packed}) == floatX(np.diag(vals)))
+
+
+@pytest.mark.parametrize(
+    "values, axis, keepdims",
+    [
+        (np.array([-4, -2]), None, True),
+        (np.array([-np.inf, -2]), None, True),
+        (np.array([-2, np.inf]), None, True),
+        (np.array([-np.inf, -np.inf]), None, True),
+        (np.array([np.inf, np.inf]), None, True),
+        (np.array([-np.inf, np.inf]), None, True),
+        (np.array([[-np.inf, -np.inf], [-np.inf, -np.inf]]), None, True),
+        (np.array([[-np.inf, -np.inf], [-np.inf, -np.inf]]), 0, True),
+        (np.array([[-np.inf, -np.inf], [-np.inf, -np.inf]]), 1, True),
+        (np.array([[-np.inf, -np.inf], [-np.inf, -np.inf]]), 0, False),
+        (np.array([[-np.inf, -np.inf], [-np.inf, -np.inf]]), 1, False),
+        (np.array([[-2, np.inf], [-np.inf, -np.inf]]), 0, True),
+    ],
+)
+def test_logsumexp(values, axis, keepdims):
+    npt.assert_almost_equal(
+        logsumexp(values, axis=axis, keepdims=keepdims).eval(),
+        scipy_logsumexp(values, axis=axis, keepdims=keepdims),
+    )