Skip to content

Adding Poisson NLL loss to libtorch #19316

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
interesaaat wants to merge 13 commits into from
Closed

Adding Poisson NLL loss to libtorch #19316

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
f24e799
* Added Poisson NLL loss c++ implementation
interesaaat Apr 16, 2019
d5f3c5d
Merge branch 'master' of https://github.com/interesaaat/pytorch
interesaaat Apr 16, 2019
6d79094
Fix typo
ezyang Apr 19, 2019
5b94f90
fixed problem with indexing non scalar tensors.
interesaaat Apr 24, 2019
d56d013
Merge branch 'master' of https://github.com/interesaaat/pytorch
interesaaat Apr 24, 2019
f1124c8
typed mask to long
interesaaat Apr 24, 2019
7465a9b
fixed typo
interesaaat Apr 25, 2019
0312a9f
fix whitespacing
interesaaat Apr 25, 2019
0dfd03f
change index_select to mask_select
interesaaat Apr 26, 2019
1944213
fix to how the full loss is computer
interesaaat Apr 27, 2019
f43913c
add missing define to fix compilation error in windows
interesaaat May 7, 2019
4697f66
add missing include
interesaaat May 7, 2019
ecfd9ba
Merge branch 'master' of https://github.com/pytorch/pytorch
interesaaat May 8, 2019
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
23 changes: 23 additions & 0 deletions aten/src/ATen/native/Loss.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,9 +1,15 @@
// define constants like M_PI and C keywords for MSVC
#ifdef _MSC_VER
#define _USE_MATH_DEFINES
#include <math.h>
#endif
#include <ATen/ATen.h>
#include <ATen/NativeFunctions.h>
#include <ATen/Dispatch.h>
#include <ATen/CPUApplyUtils.h>

#define EPSILON 1e-12
#define _USE_MATH_DEFINES

namespace {
static inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
Expand Down Expand Up @@ -125,4 +131,21 @@ Tensor binary_cross_entropy_with_logits_backward(const Tensor& grad, const Tenso

return grad_input;
}

Tensor poisson_nll_loss(const Tensor& input, const Tensor& target, const bool log_input, const bool full, const double eps, const int64_t reduction)
{
Tensor loss;
if (log_input) {
loss = at::exp(input) - target * input;
} else {
loss = input - target * at::log(input + eps);
}

if (full) {
auto mask1 = (target > 1);
loss.masked_select(mask1) += (target * at::log(target) - target + 0.5 * at::log(2 * M_PI * target)).masked_select(mask1);
}

return apply_loss_reduction(loss, reduction);
}
}} // namespace at::native
3 changes: 3 additions & 0 deletions aten/src/ATen/native/native_functions.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1411,6 +1411,9 @@
- func: pinverse(Tensor self, float rcond=1e-15) -> Tensor
variants: function, method

- func: poisson_nll_loss(Tensor input, Tensor target, bool log_input, bool full, float eps, int reduction) -> Tensor
variants: function

- func: scalar_tensor(Scalar s, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor

- func: rand(int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
Expand Down
17 changes: 3 additions & 14 deletions torch/nn/functional.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1930,22 +1930,11 @@ def poisson_nll_loss(input, target, log_input=True, full=False, size_average=Non
"""
if size_average is not None or reduce is not None:
reduction = _Reduction.legacy_get_string(size_average, reduce)
if log_input:
loss = torch.exp(input) - target * input
else:
loss = input - target * torch.log(input + eps)
if full:
mask = target > 1
loss[mask] += (target * torch.log(target) - target + 0.5 * torch.log(2 * math.pi * target))[mask]
if reduction == 'none':
ret = loss
elif reduction == 'mean':
ret = torch.mean(loss)
elif reduction == 'sum':
ret = torch.sum(loss)
else:
if reduction != 'none' and reduction != 'mean' and reduction != 'sum':
ret = input
raise ValueError(reduction + " is not valid")

ret = torch.poisson_nll_loss(input, target, log_input, full, eps, _Reduction.get_enum(reduction))
return ret


Expand Down