Skip to content

Allow censoring Categorical distributions #7662

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.

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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
45 changes: 37 additions & 8 deletions pymc/distributions/discrete.py
Original file line number Diff line number Diff line change
Expand Up @@ -1155,29 +1155,58 @@ def support_point(rv, size, p):
mode = pt.full(size, mode)
return mode

def logp(value, p):
k = pt.shape(p)[-1]
value_clip = pt.clip(value, 0, k - 1)
@staticmethod
def _safe_index_value_p(value, p):
# Find the probabily of the given value by indexing in p,
# after handling broadcasting and invalid values.

# In the standard case p has one more dimension than value
dim_diff = p.type.ndim - value.type.ndim
if dim_diff > 1:
# p brodacasts implicitly beyond value
value_clip = pt.shape_padleft(value_clip, dim_diff - 1)
value = pt.shape_padleft(value, dim_diff - 1)
elif dim_diff < 1:
# value broadcasts implicitly beyond p
p = pt.shape_padleft(p, 1 - dim_diff)

a = pt.log(pt.take_along_axis(p, value_clip[..., None], axis=-1).squeeze(-1))
k = pt.shape(p)[-1]
value_clip = pt.clip(value, 0, k - 1).astype(int)
return value, pt.log(pt.take_along_axis(p, value_clip[..., None], axis=-1).squeeze(-1))

res = pt.switch(
def logp(value, p):
k = pt.shape(p)[-1]
value, safe_value_p = Categorical._safe_index_value_p(value, p)
Copy link
Member

Choose a reason for hiding this comment

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

I don't love making _safe_index_value_p a method. What's the conceptual advantage vs just being a function?

Copy link
Member Author

Choose a reason for hiding this comment

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

doesn't clutter the namespace, that's the only difference


value_p = pt.switch(
pt.or_(pt.lt(value, 0), pt.gt(value, k - 1)),
-np.inf,
a,
safe_value_p,
)

return check_parameters(
res,
value_p,
0 <= p,
p <= 1,
pt.isclose(pt.sum(p, axis=-1), 1),
msg="0 <= p <=1, sum(p) = 1",
)

def logcdf(value, p):
k = pt.shape(p)[-1]
value, safe_value_p = Categorical._safe_index_value_p(value, p.cumsum(-1))

value_p = pt.switch(
pt.lt(value, 0),
-np.inf,
pt.switch(
pt.gt(value, k - 1),
0,
safe_value_p,
),
)

return check_parameters(
value_p,
0 <= p,
p <= 1,
pt.isclose(pt.sum(p, axis=-1), 1),
Expand Down
16 changes: 16 additions & 0 deletions tests/distributions/test_censored.py
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@

import pymc as pm

from pymc import logp
from pymc.distributions.shape_utils import change_dist_size


Expand Down Expand Up @@ -110,3 +111,18 @@ def test_dist_broadcasted_by_lower_upper(self):
pm.Normal.dist(size=(3, 4, 2)), lower=np.zeros((2,)), upper=np.zeros((4, 2))
)
assert tuple(x.owner.inputs[0].shape.eval()) == (3, 4, 2)

def test_censored_categorical(self):
cat = pm.Categorical.dist([0.1, 0.2, 0.2, 0.3, 0.2], shape=(5,))

np.testing.assert_allclose(
logp(cat, [-1, 0, 1, 2, 3, 4, 5]).exp().eval(),
[0, 0.1, 0.2, 0.2, 0.3, 0.2, 0],
)

censored_cat = pm.Censored.dist(cat, lower=1, upper=3, shape=(5,))

np.testing.assert_allclose(
logp(censored_cat, [-1, 0, 1, 2, 3, 4, 5]).exp().eval(),
[0, 0, 0.3, 0.2, 0.5, 0, 0],
)
55 changes: 35 additions & 20 deletions tests/distributions/test_discrete.py
Original file line number Diff line number Diff line change
Expand Up @@ -367,43 +367,58 @@ def test_poisson(self):

@pytest.mark.parametrize("n", [2, 3, 4])
def test_categorical(self, n):
domain = Domain(range(n), dtype="int64", edges=(0, n))
paramdomains = {"p": Simplex(n)}

check_logp(
pm.Categorical,
Domain(range(n), dtype="int64", edges=(0, n)),
{"p": Simplex(n)},
domain,
paramdomains,
lambda value, p: categorical_logpdf(value, p),
)

def test_categorical_logp_batch_dims(self):
check_selfconsistency_discrete_logcdf(
pm.Categorical,
domain,
paramdomains,
)

@pytest.mark.parametrize("method", (logp, logcdf), ids=lambda x: x.__name__)
def test_categorical_logp_batch_dims(self, method):
# Core case
p = np.array([0.2, 0.3, 0.5])
value = np.array(2.0)
logp_expr = logp(pm.Categorical.dist(p=p, shape=value.shape), value)
assert logp_expr.type.ndim == 0
np.testing.assert_allclose(logp_expr.eval(), np.log(0.5))
expr = method(pm.Categorical.dist(p=p, shape=value.shape), value)
assert expr.type.ndim == 0
expected_p = 0.5 if method is logp else 1.0
np.testing.assert_allclose(expr.exp().eval(), expected_p)

# Explicit batched value broadcasts p
bcast_p = p[None] # shape (1, 3)
batch_value = np.array([0, 1]) # shape(3,)
logp_expr = logp(pm.Categorical.dist(p=bcast_p, shape=batch_value.shape), batch_value)
assert logp_expr.type.ndim == 1
np.testing.assert_allclose(logp_expr.eval(), np.log([0.2, 0.3]))
expr = method(pm.Categorical.dist(p=bcast_p, shape=batch_value.shape), batch_value)
assert expr.type.ndim == 1
expected_p = [0.2, 0.3] if method is logp else [0.2, 0.5]
np.testing.assert_allclose(expr.exp().eval(), expected_p)

# Implicit batch value broadcasts p
expr = method(pm.Categorical.dist(p=p, shape=()), batch_value)
assert expr.type.ndim == 1
expected_p = [0.2, 0.3] if method is logp else [0.2, 0.5]
np.testing.assert_allclose(expr.exp().eval(), expected_p)

# Explicit batched value and batched p
batch_p = np.array([p[::-1], p])
logp_expr = logp(pm.Categorical.dist(p=batch_p, shape=batch_value.shape), batch_value)
assert logp_expr.type.ndim == 1
np.testing.assert_allclose(logp_expr.eval(), np.log([0.5, 0.3]))

# Implicit batch value broadcasts p
logp_expr = logp(pm.Categorical.dist(p=p, shape=()), batch_value)
assert logp_expr.type.ndim == 1
np.testing.assert_allclose(logp_expr.eval(), np.log([0.2, 0.3]))
expr = method(pm.Categorical.dist(p=batch_p, shape=batch_value.shape), batch_value)
assert expr.type.ndim == 1
expected_p = [0.5, 0.3] if method is logp else [0.5, 0.5]
np.testing.assert_allclose(expr.exp().eval(), expected_p)

# Implicit batch p broadcasts value
logp_expr = logp(pm.Categorical.dist(p=batch_p, shape=None), value)
assert logp_expr.type.ndim == 1
np.testing.assert_allclose(logp_expr.eval(), np.log([0.2, 0.5]))
expr = method(pm.Categorical.dist(p=batch_p, shape=None), value)
assert expr.type.ndim == 1
expected_p = [0.2, 0.5] if method is logp else [1.0, 1.0]
np.testing.assert_allclose(expr.exp().eval(), expected_p)

@pytensor.config.change_flags(compute_test_value="raise")
def test_categorical_bounds(self):
Expand Down