Skip to content
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

fix: Retain continuous approximation of Poisson in torch v1.8.0 #1351

Merged
merged 9 commits into from
Mar 5, 2021
Merged

fix: Retain continuous approximation of Poisson in torch v1.8.0 #1351

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
1a9b79b
Use astensor to force tensors to distributions
matthewfeickert Mar 5, 2021
df72c8d
Split pytorch test components with ValueError from other backends
matthewfeickert Mar 5, 2021
c966e15
Add variables names to make clearer
matthewfeickert Mar 5, 2021
9f0304a
Retain Poisson continuous approximation with validate_args=False
matthewfeickert Mar 5, 2021
3c61d19
Set minimum version of torch to be 1.8
matthewfeickert Mar 5, 2021
3d4974e
Revert "Set minimum version of torch to be 1.8"
matthewfeickert Mar 5, 2021
a1fa7ab
Reduce diff
matthewfeickert Mar 5, 2021
16122f3
Make seperate test for pytorch
matthewfeickert Mar 5, 2021
3040eac
Add comment
matthewfeickert Mar 5, 2021
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
19 changes: 16 additions & 3 deletions src/pyhf/tensor/pytorch_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -319,7 +319,8 @@ def einsum(self, subscripts, *operands):
return torch.einsum(subscripts, operands)

def poisson_logpdf(self, n, lam):
return torch.distributions.Poisson(lam).log_prob(n)
# validate_args=True disallows continuous approximation
return torch.distributions.Poisson(lam, validate_args=False).log_prob(n)

def poisson(self, n, lam):
r"""
Expand Down Expand Up @@ -347,9 +348,16 @@ def poisson(self, n, lam):
Returns:
PyTorch FloatTensor: Value of the continuous approximation to Poisson(n|lam)
"""
return torch.exp(torch.distributions.Poisson(lam).log_prob(n))
# validate_args=True disallows continuous approximation
return torch.exp(
torch.distributions.Poisson(lam, validate_args=False).log_prob(n)
)

def normal_logpdf(self, x, mu, sigma):
x = self.astensor(x)
mu = self.astensor(mu)
sigma = self.astensor(sigma)
matthewfeickert marked this conversation as resolved.
Show resolved Hide resolved

normal = torch.distributions.Normal(mu, sigma)
return normal.log_prob(x)

Expand Down Expand Up @@ -379,6 +387,10 @@ def normal(self, x, mu, sigma):
Returns:
PyTorch FloatTensor: Value of Normal(x|mu, sigma)
"""
x = self.astensor(x)
mu = self.astensor(mu)
sigma = self.astensor(sigma)

normal = torch.distributions.Normal(mu, sigma)
return self.exp(normal.log_prob(x))

Expand Down Expand Up @@ -433,7 +445,8 @@ def poisson_dist(self, rate):
PyTorch Poisson distribution: The Poisson distribution class

"""
return torch.distributions.Poisson(rate)
# validate_args=True disallows continuous approximation
return torch.distributions.Poisson(rate, validate_args=False)

def normal_dist(self, mu, sigma):
r"""
Expand Down
84 changes: 63 additions & 21 deletions tests/test_tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -252,25 +252,65 @@ def test_pdf_calculations(backend):
assert tb.tolist(tb.normal_cdf(tb.astensor([0.8]))) == pytest.approx(
[0.7881446014166034], 1e-07
)
assert tb.tolist(
tb.normal_logpdf(
tb.astensor([0, 0, 1, 1, 0, 0, 1, 1]),
tb.astensor([0, 1, 0, 1, 0, 1, 0, 1]),
tb.astensor([0, 0, 0, 0, 1, 1, 1, 1]),
if tb.name == "pytorch":
matthewfeickert marked this conversation as resolved.
Show resolved Hide resolved
with pytest.raises(ValueError):
_ = tb.normal_logpdf(
tb.astensor([0]),
tb.astensor([0]),
tb.astensor([0]),
)
with pytest.raises(ValueError):
_ = tb.normal_logpdf(
tb.astensor([0]),
tb.astensor([1]),
tb.astensor([0]),
)
with pytest.raises(ValueError):
_ = tb.normal_logpdf(
tb.astensor([1]),
tb.astensor([0]),
tb.astensor([0]),
)
with pytest.raises(ValueError):
_ = tb.normal_logpdf(
tb.astensor([1]),
tb.astensor([1]),
tb.astensor([0]),
)
assert tb.tolist(
tb.normal_logpdf(
tb.astensor([0, 0, 1, 1]),
tb.astensor([0, 1, 0, 1]),
tb.astensor([1, 1, 1, 1]),
)
) == pytest.approx(
[
-0.91893853,
-1.41893853,
-1.41893853,
-0.91893853,
],
)
else:
assert tb.tolist(
tb.normal_logpdf(
tb.astensor([0, 0, 1, 1, 0, 0, 1, 1]),
tb.astensor([0, 1, 0, 1, 0, 1, 0, 1]),
tb.astensor([0, 0, 0, 0, 1, 1, 1, 1]),
)
) == pytest.approx(
[
np.nan,
np.nan,
np.nan,
np.nan,
-0.91893853,
-1.41893853,
-1.41893853,
-0.91893853,
],
nan_ok=True,
)
) == pytest.approx(
[
np.nan,
np.nan,
np.nan,
np.nan,
-0.91893853,
-1.41893853,
-1.41893853,
-0.91893853,
],
nan_ok=True,
)
# poisson(lambda=0) is not defined, should return NaN
assert tb.tolist(
tb.poisson(tb.astensor([0, 0, 1, 1]), tb.astensor([0, 1, 0, 1]))
Expand All @@ -285,9 +325,11 @@ def test_pdf_calculations(backend):
)

# Ensure continuous approximation is valid
assert tb.tolist(
tb.poisson(tb.astensor([0.5, 1.1, 1.5]), tb.astensor(1.0))
) == pytest.approx([0.4151074974205947, 0.3515379040027489, 0.2767383316137298])
values = tb.astensor([0.5, 1.1, 1.5])
rates = tb.astensor(1.0)
assert tb.tolist(tb.poisson(values, rates)) == pytest.approx(
[0.4151074974205947, 0.3515379040027489, 0.2767383316137298]
)


def test_boolean_mask(backend):
Expand Down