C++ API parity: Upsample (#28413)

Summary:
Adds `interpolate` functional and `Upsample` module support for the C++ API.

**Issue**: https://github.com/pytorch/pytorch/issues/25883

**Reviewer**: yf225
Pull Request resolved: https://github.com/pytorch/pytorch/pull/28413

Differential Revision: D18165014

Pulled By: yf225

fbshipit-source-id: ecae2f432a301b1f4afa7c038b2d104cbad139f2
This commit is contained in:
jon-tow
2019-10-28 21:32:30 -07:00
committed by Facebook Github Bot
parent 1e3e1f5bf9
commit 52dd587123
15 changed files with 546 additions and 1 deletions
+86
View File
@@ -1266,6 +1266,92 @@ TEST_F(FunctionalTest, Threshold) {
}
}
TEST_F(FunctionalTest, Interpolate) {
{
// 1D interpolation
auto input = torch::ones({1, 1, 2});
auto options = InterpolateOptions()
.size({4})
.mode(torch::kNearest);
auto output = F::interpolate(input, options);
auto expected = torch::ones({1, 1, 4});
ASSERT_TRUE(output.allclose(expected));
}
{
// 2D interpolation
for (const auto align_corners : {true, false}) {
// test float scale factor up & down sampling
for (const auto scale_factor : {0.5, 1.5, 2.0}) {
auto input = torch::ones({1, 1, 2, 2});
auto options = InterpolateOptions()
.scale_factor({scale_factor, scale_factor})
.mode(torch::kBilinear)
.align_corners(align_corners);
auto output = F::interpolate(input, options);
auto expected_size =
static_cast<int64_t>(std::floor(input.size(-1) * scale_factor));
auto expected = torch::ones({1, 1, expected_size, expected_size});
ASSERT_TRUE(output.allclose(expected));
}
}
}
{
// 3D interpolation
for (const auto align_corners : {true, false}) {
for (const auto scale_factor : {0.5, 1.5, 2.0}) {
auto input = torch::ones({1, 1, 2, 2, 2});
auto options =
InterpolateOptions()
.scale_factor({scale_factor, scale_factor, scale_factor})
.mode(torch::kTrilinear)
.align_corners(align_corners);
auto output = F::interpolate(input, options);
auto expected_size =
static_cast<int64_t>(std::floor(input.size(-1) * scale_factor));
auto expected =
torch::ones({1, 1, expected_size, expected_size, expected_size});
ASSERT_TRUE(output.allclose(expected));
}
}
}
{
auto input = torch::randn({3, 2, 2});
ASSERT_THROWS_WITH(
F::interpolate(input[0], InterpolateOptions().size({4, 4})),
"Input Error: Only 3D, 4D and 5D input Tensors supported (got 2D) "
"for the modes: nearest | linear | bilinear | bicubic | trilinear (got kNearest)");
ASSERT_THROWS_WITH(
F::interpolate(
torch::reshape(input, {1, 1, 1, 3, 2, 2}),
InterpolateOptions().size({1, 1, 1, 3, 4, 4})),
"Input Error: Only 3D, 4D and 5D input Tensors supported (got 6D) "
"for the modes: nearest | linear | bilinear | bicubic | trilinear (got kNearest)");
ASSERT_THROWS_WITH(
F::interpolate(input, InterpolateOptions()),
"either size or scale_factor should be defined");
ASSERT_THROWS_WITH(
F::interpolate(
input,
InterpolateOptions().size({3, 4, 4}).scale_factor({0.5})),
"only one of size or scale_factor should be defined");
ASSERT_THROWS_WITH(
F::interpolate(input, InterpolateOptions().scale_factor({3, 2})),
"scale_factor shape must match input shape. "
"Input is 1D, scale_factor size is 2");
ASSERT_THROWS_WITH(
F::interpolate(
input,
InterpolateOptions()
.mode(torch::kNearest)
.align_corners(true)),
"align_corners option can only be set with the "
"interpolating modes: linear | bilinear | bicubic | trilinear");
}
}
TEST_F(FunctionalTest, Pad) {
{
auto input = torch::arange(6, torch::kDouble).reshape({1, 2, 3});