Skip to content

Fix x.pow(0) gradient when x contains 0 #8945

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

Closed
vishwakftw wants to merge 7 commits into from
Closed

Fix x.pow(0) gradient when x contains 0 #8945

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
641ac38
Fix x.pow(0) gradient when x contains 0
vishwakftw Jun 27, 2018
e79bf25
Reduce allocations
vishwakftw Jun 27, 2018
7e5e32d
Fix error
vishwakftw Jun 27, 2018
15c0669
Fix error
vishwakftw Jun 27, 2018
cb45292
Add tests for the fix
vishwakftw Jun 27, 2018
d3dc77b
fix nit; randn --> zeros
vishwakftw Jun 28, 2018
fb97fae
Merge branch 'master' into pow-0-derivative
ezyang Jun 29, 2018
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
10 changes: 10 additions & 0 deletions test/test_autograd.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2030,6 +2030,16 @@ def run_test(input_size, norm_deg):
run_test((10,), 1)
run_test((10,), 1.5)

def test_pow_zero_tensor_gradient(self):
def run_test(input_size, exponent):
input = torch.zeros(*input_size, requires_grad=True)
input.pow(exponent).sum().backward()
self.assertEqual(input.grad.data.abs().sum(), 0)

run_test((10,), torch.zeros(10))
run_test((10, 10), torch.zeros(10, 10))
run_test((10,), 0)

def test_profiler(self):
x = torch.randn(10, 10)

Expand Down
6 changes: 3 additions & 3 deletions tools/autograd/derivatives.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -529,11 +529,11 @@
input2: not_implemented("potri")

- name: pow(Tensor self, Scalar exponent)
self: grad * exponent * self.pow(exponent.toDouble() - 1)
self: pow_backward(grad, self, exponent)

- name: pow(Tensor self, Tensor exponent)
self: grad * exponent * self.pow(exponent - 1)
exponent: grad * self.pow(exponent) * self.log()
self: pow_backward_self(grad, self, exponent)
exponent: pow_backward_exponent(grad, self, exponent)

- name: _prod(Tensor self, int64_t dim, bool keepdim)
self: prod_backward(grad, self, result, dim, keepdim)
Expand Down
17 changes: 17 additions & 0 deletions tools/autograd/templates/Functions.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,23 @@ Tensor norm_backward(Tensor grad, const Tensor & self, const Scalar & p_, Tensor
return norm_backward(grad, self, p_, norm);
}

Tensor pow_backward(Tensor grad, const Tensor & self, const Scalar & exponent_) {
double exponent = exponent_.toDouble();
if (exponent == 0.0) {
return zeros_like(self);
} else {
return grad * exponent * self.pow(exponent - 1);
}
}

Tensor pow_backward_self(Tensor grad, const Tensor & self, const Tensor & exponent) {
return at::where(exponent == 0.0, at::zeros({}, grad.type()), grad * exponent * self.pow(exponent - 1));
}

Tensor pow_backward_exponent(Tensor grad, const Tensor & self, const Tensor & exponent) {
return grad * self.pow(exponent) * self.log();
}

Tensor reduce_to(const Tensor & grad, IntList sizes) {
if (sizes.size() == 0) {
return grad.sum();
Expand Down