Skip to content

Added backward function for kl_div target #7839

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
weiyangfb merged 3 commits into from
Jun 8, 2018
Merged

Added backward function for kl_div target #7839

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e72c884
added backward fn for target
weiyangfb May 25, 2018
df15253
WIP: create correct testcase for gTarget/ggTarget
weiyangfb May 25, 2018
276c1ae
added module test for kl_div target, and assuming targets are probabi…
weiyangfb May 31, 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
13 changes: 13 additions & 0 deletions test/test_nn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5578,6 +5578,18 @@ def bce_with_logistic_no_reduce_scalar_test():
pickle=False)


def kldivloss_with_target_no_reduce_test():
i = torch.rand(10, 10).log()
return dict(
fullname='KLDivLoss_with_target_no_reduce',
constructor=wrap_functional(
lambda t: F.kl_div(i.type_as(t), t, reduce=False)),
input_fn=lambda: torch.rand(10, 10),
reference_fn=lambda t, _:
loss_reference_fns['KLDivLoss'](i.type_as(t), t, reduce=False),
pickle=False)


def kldivloss_no_reduce_test():
t = torch.randn(10, 10)
return dict(
Expand Down Expand Up @@ -6030,6 +6042,7 @@ def multimarginloss_weights_no_reduce_test():
bceloss_no_reduce_scalar_test(),
bceloss_weights_no_reduce_scalar_test(),
bce_with_logistic_no_reduce_scalar_test(),
kldivloss_with_target_no_reduce_test(),
kldivloss_no_reduce_test(),
kldivloss_no_reduce_scalar_test(),
l1loss_no_reduce_test(),
Expand Down
2 changes: 2 additions & 0 deletions tools/autograd/derivatives.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -718,6 +718,7 @@

- name: kl_div_forward(Tensor self, Tensor target, bool size_average, bool reduce)
self: kl_div_backward(grad, self, target, size_average, reduce)
target: kl_div_target_backward(grad, self, target, size_average, reduce)

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view


- name: l1_loss_forward(Tensor self, Tensor target, bool size_average, bool reduce)
self: l1_loss_backward(grad, self, target, size_average, reduce)
Expand Down Expand Up @@ -963,6 +964,7 @@
- name: kl_div_backward(Tensor grad_output, Tensor self, Tensor target, bool size_average, bool reduce)
grad_output: kl_div_double_backward_grad_output(grad, self, target, size_average, reduce)
self: zeros_like(grad)
target: zeros_like(grad)

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view


- name: l1_loss_backward(Tensor grad_output, Tensor self, Tensor target, bool size_average, bool reduce)
grad_output: l1_loss_double_backward_grad_output(grad, self, target, size_average, reduce)
Expand Down
12 changes: 12 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 @@ -665,6 +665,18 @@ Tensor kl_div_double_backward_grad_output(const Tensor & grad, const Tensor & in
return result;
}

// Compute derivatives for targets.
// Assume targets are given as probabilities (i.e. without taking the logarithm).
Tensor kl_div_target_backward(Tensor grad_output, Tensor self, Tensor target, bool size_average, bool reduce) {
if (!reduce) {
return grad_output.mul(target.log().add_(1).sub_(self)).masked_fill_(target == 0, 0.);
}
if (size_average) {
return grad_output.mul(target.log().add_(1).sub_(self)).div_(target.numel()).masked_fill_(target == 0, 0.);
}
return grad_output.mul(target.log().add_(1).sub_(self)).masked_fill_(target == 0, 0.);
}

Tensor log_sigmoid_double_backward(const Tensor & grad, const Tensor & input) {
auto z = input.sigmoid();
return grad * (z - 1) * z;
Expand Down