[c++] Distance-agnostic triplet margin loss (#45377)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45377

This PR adds a C++ implementation of the TripletMarginWithDistanceLoss, for which the Python implementation was introduced in PR #43680.  It's based on PR #44072, but I'm resubmitting this to unlink it from Phabricator.

Test Plan: Imported from OSS

Reviewed By: izdeby

Differential Revision: D24003973

fbshipit-source-id: 2d9ada7260a6f27425ff2fdbbf623dad0fb79405
This commit is contained in:
Xinyu Li
2020-09-30 12:35:08 -07:00
committed by Facebook GitHub Bot
parent 181afd5220
commit c9bb990707
7 changed files with 379 additions and 11 deletions
+50
View File
@@ -682,6 +682,56 @@ TEST_F(FunctionalTest, TripletMarginLoss) {
ASSERT_TRUE(output.allclose(expected, 1e-04));
}
TEST_F(FunctionalTest, TripletMarginWithDistanceLossDefaultParity) {
// Check that if we use torch::pairwise_distance with the default
// TripletMarginLoss options as our distance function, the outputs
// are equal (i.e., equal under defaults).
std::vector<TripletMarginWithDistanceLossOptions::reduction_t>
reductions = {torch::kSum, torch::kMean, torch::kNone};
std::vector<float> margins = {0.5, 1.0, 1.5};
std::vector<bool> swaps = {true, false};
for (auto& reduction : reductions) {
for (auto& margin : margins) {
for (const auto& swap : swaps) {
auto anchor =
torch::randn({100, 128}, torch::dtype(torch::kFloat).requires_grad(true));
auto positive =
torch::randn({100, 128}, torch::dtype(torch::kFloat).requires_grad(true));
auto negative =
torch::randn({100, 128}, torch::dtype(torch::kFloat).requires_grad(true));
auto basicOptions = F::TripletMarginLossFuncOptions()
.reduction(reduction)
.margin(margin)
.swap(swap);
auto distanceOptions =
F::TripletMarginWithDistanceLossFuncOptions()
.reduction(reduction)
.margin(margin)
.swap(swap);
TripletMarginLoss basicLoss(basicOptions);
TripletMarginWithDistanceLoss distanceLoss(distanceOptions);
auto basicOutput =
F::triplet_margin_loss(anchor, positive, negative, basicOptions);
auto distanceOutput = F::triplet_margin_with_distance_loss(
anchor, positive, negative, distanceOptions);
ASSERT_TRUE(distanceOutput.allclose(basicOutput, 1e-6, 1e-6));
// handle for torch::kNone reduction
auto sum = distanceOutput.sum();
sum.backward();
ASSERT_EQ(anchor.sizes(), anchor.grad().sizes());
ASSERT_EQ(positive.sizes(), positive.grad().sizes());
ASSERT_EQ(negative.sizes(), negative.grad().sizes());
}
}
}
}
TEST_F(FunctionalTest, NLLLoss) {
auto input = torch::tensor({{-0.1315, -3.1315, -2.5315},
{-3.7038, -0.1038, -2.6038},