Added bicubic support for interpolation with AA #3810
+192
−75
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -141,66 +141,7 @@ void ti_cpu_upsample_generic_aa( | |
| // Helper structs to use with ti_upsample_generic_Nd_kernel_impl | ||
| template <typename index_t, typename scalar_t> | ||
| struct HelperInterpBase { | ||
| template <typename filter_fn_t> | ||
| static inline std::vector<Tensor> _compute_indices_weights_aa( | ||
| int64_t input_size, | ||
| int64_t output_size, | ||
|
|
@@ -209,14 +150,15 @@ struct HelperInterpLinear : public HelperInterpBase<index_t, scalar_t> { | |
| int64_t reshape_dim, | ||
| bool align_corners, | ||
| scalar_t scale, | ||
| int& in_out_interp_size, | ||
| filter_fn_t filter_fn) { | ||
| int interp_size = in_out_interp_size; | ||
| scalar_t support = | ||
| (scale >= 1.0) ? (interp_size * 0.5) * scale : interp_size * 0.5; | ||
| interp_size = (int)ceilf(support) * 2 + 1; | ||
|
|
||
| // return interp_size | ||
| in_out_interp_size = interp_size; | ||
|
|
||
| std::vector<Tensor> output; | ||
| auto new_shape = std::vector<int64_t>(ndims, 1); | ||
|
|
@@ -269,7 +211,7 @@ struct HelperInterpLinear : public HelperInterpBase<index_t, scalar_t> { | |
|
|
||
| total_w = 0.0; | ||
| for (j = 0; j < xmax; j++) { | ||
| scalar_t w = filter_fn((j + xmin - center + 0.5) * invscale); | ||
| wt_ptr[i * interp_size + j] = w; | ||
| total_w += w; | ||
| } | ||
|
|
@@ -287,6 +229,102 @@ struct HelperInterpLinear : public HelperInterpBase<index_t, scalar_t> { | |
| } | ||
| }; | ||
|
|
||
| template <typename index_t, typename scalar_t> | ||
| struct HelperInterpLinear : public HelperInterpBase<index_t, scalar_t> { | ||
| static const int interp_size = 2; | ||
|
|
||
| // taken from | ||
| // https://github.com/python-pillow/Pillow/blob/6812205f18ca4ef54372e87e1a13ce4a859434df/ | ||
| // src/libImaging/Resample.c#L20-L29 | ||
| static inline scalar_t _filter(scalar_t x) { | ||
| if (x < 0.0) { | ||
| x = -x; | ||
| } | ||
| if (x < 1.0) { | ||
| return 1.0 - x; | ||
| } | ||
| return 0.0; | ||
| } | ||
|
|
||
| static inline std::vector<Tensor> compute_indices_weights( | ||
| int64_t input_size, | ||
| int64_t output_size, | ||
| int64_t stride, | ||
| int64_t ndims, | ||
| int64_t reshape_dim, | ||
| bool align_corners, | ||
| const c10::optional<double> opt_scale, | ||
| bool antialias, | ||
| int& out_interp_size) { | ||
| TORCH_INTERNAL_ASSERT(antialias); | ||
| scalar_t scale = area_pixel_compute_scale<scalar_t>( | ||
| input_size, output_size, align_corners, opt_scale); | ||
|
|
||
| out_interp_size = HelperInterpLinear<index_t, scalar_t>::interp_size; | ||
| return HelperInterpLinear<index_t, scalar_t>::_compute_indices_weights_aa( | ||
| input_size, | ||
| output_size, | ||
| stride, | ||
| ndims, | ||
| reshape_dim, | ||
| align_corners, | ||
| scale, | ||
| out_interp_size, | ||
| _filter); | ||
| } | ||
| }; | ||
|
|
||
| template <typename index_t, typename scalar_t> | ||
| struct HelperInterpCubic : public HelperInterpBase<index_t, scalar_t> { | ||
| static const int interp_size = 4; | ||
|
|
||
| static inline std::vector<Tensor> compute_indices_weights( | ||
| int64_t input_size, | ||
| int64_t output_size, | ||
| int64_t stride, | ||
| int64_t ndims, | ||
| int64_t reshape_dim, | ||
| bool align_corners, | ||
| const c10::optional<double> opt_scale, | ||
| bool antialias, | ||
| int& out_interp_size) { | ||
| TORCH_INTERNAL_ASSERT(antialias); | ||
| scalar_t scale = area_pixel_compute_scale<scalar_t>( | ||
| input_size, output_size, align_corners, opt_scale); | ||
|
|
||
| out_interp_size = HelperInterpCubic<index_t, scalar_t>::interp_size; | ||
| return HelperInterpCubic<index_t, scalar_t>::_compute_indices_weights_aa( | ||
| input_size, | ||
| output_size, | ||
| stride, | ||
| ndims, | ||
| reshape_dim, | ||
| align_corners, | ||
| scale, | ||
| out_interp_size, | ||
| _filter); | ||
| } | ||
|
|
||
| // taken from | ||
| // https://github.com/python-pillow/Pillow/blob/6812205f18ca4ef54372e87e1a13ce4a859434df/ | ||
| // src/libImaging/Resample.c#L46-L62 | ||
| static inline scalar_t _filter(scalar_t x) { | ||
|
There was a problem hiding this comment. For the future: check if this is equivalent / similar to https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/UpSample.h#L324-L332 |
||
| // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm | ||
| #define a -0.5 | ||
| if (x < 0.0) { | ||
| x = -x; | ||
| } | ||
| if (x < 1.0) { | ||
| return ((a + 2.0) * x - (a + 3.0)) * x * x + 1; | ||
| } | ||
| if (x < 2.0) { | ||
| return (((x - 5) * x + 8) * x - 4) * a; | ||
| } | ||
| return 0.0; | ||
| #undef a | ||
| } | ||
| }; | ||
|
|
||
| template < | ||
| typename index_t, | ||
| int out_ndims, | ||
|
|
@@ -396,16 +434,15 @@ void ti_separable_upsample_generic_Nd_kernel_impl( | |
| index_t, | ||
| out_ndims, | ||
| scale_t, | ||
| F>( | ||
| temp_output, temp_input, interp_dim, align_corners, scales, antialias); | ||
| temp_input = temp_output; | ||
| } | ||
| _ti_separable_upsample_generic_Nd_kernel_impl_single_dim< | ||
| index_t, | ||
| out_ndims, | ||
| scale_t, | ||
| F>(output, temp_input, 2, align_corners, scales, antialias); | ||
| } | ||
|
|
||
| void _ti_upsample_bilinear2d_kernel_impl( | ||
|
|
@@ -423,6 +460,21 @@ void _ti_upsample_bilinear2d_kernel_impl( | |
| output, input, align_corners, {scales_h, scales_w}, antialias); | ||
| } | ||
|
|
||
| void _ti_upsample_bicubic2d_kernel_impl( | ||
| Tensor& output, | ||
| const Tensor& input, | ||
| bool align_corners, | ||
| c10::optional<double> scales_h, | ||
| c10::optional<double> scales_w, | ||
| bool antialias) { | ||
| ti_separable_upsample_generic_Nd_kernel_impl< | ||
| int64_t, | ||
| 2, | ||
| scale_t, | ||
| HelperInterpCubic>( | ||
| output, input, align_corners, {scales_h, scales_w}, antialias); | ||
| } | ||
|
|
||
| } // namespace internal_upsample | ||
| } // namespace native | ||
| } // namespace at | ||
|
|
@@ -463,6 +515,37 @@ at::Tensor interpolate_linear_aa_forward_kernel( | |
| return output; | ||
| } | ||
|
|
||
| at::Tensor interpolate_bicubic_aa_forward_kernel( | ||
| const at::Tensor& input, | ||
| at::IntArrayRef output_size, | ||
| bool align_corners) { | ||
| TORCH_CHECK(input.device().is_cpu(), "input must be a CPU tensor"); | ||
|
|
||
| c10::optional<c10::ArrayRef<double>> scale_factors = {}; | ||
|
|
||
| // Copied from UpSampleBilinear2d.cpp | ||
| auto output = at::empty({0}, input.options()); | ||
| auto osize = at::native::upsample::compute_output_size( | ||
| input.sizes(), output_size, scale_factors); | ||
| auto scale_h = at::native::upsample::get_scale_value(scale_factors, 0); | ||
| auto scale_w = at::native::upsample::get_scale_value(scale_factors, 1); | ||
| auto full_output_size = | ||
| at::native::upsample_2d_common_check(input.sizes(), osize); | ||
|
|
||
| // Allow for empty batch size but not other dimensions | ||
| TORCH_CHECK( | ||
| input.numel() != 0 || | ||
| c10::multiply_integers( | ||
| input.sizes().begin() + 1, input.sizes().end()), | ||
| "Non-empty 4D data tensor expected but got a tensor with sizes ", | ||
| input.sizes()); | ||
|
|
||
| output.resize_(full_output_size, input.suggest_memory_format()); | ||
| at::native::internal_upsample::_ti_upsample_bicubic2d_kernel_impl( | ||
| output, input, align_corners, scale_h, scale_w, /*antialias=*/true); | ||
| return output; | ||
| } | ||
|
|
||
| // TODO: Implement backward function | ||
| // at::Tensor interpolate_linear_aa_backward_kernel( | ||
| // const at::Tensor& grad) { | ||
|
|
@@ -475,6 +558,10 @@ TORCH_LIBRARY_IMPL(torchvision, CPU, m) { | |
| m.impl( | ||
| TORCH_SELECTIVE_NAME("torchvision::_interpolate_linear_aa"), | ||
| TORCH_FN(interpolate_linear_aa_forward_kernel)); | ||
| m.impl( | ||
| TORCH_SELECTIVE_NAME("torchvision::_interpolate_bicubic_aa"), | ||
| TORCH_FN(interpolate_bicubic_aa_forward_kernel)); | ||
|
Comment on lines
+561
to
+563
There was a problem hiding this comment. Up to you, but another option would be to handle this all inside a single function in C++, and don't expose multiple variants to be dispatched on the python side of things. |
||
|
|
||
| // TODO: Implement backward function | ||
| // m.impl( | ||
| // TORCH_SELECTIVE_NAME("torchvision::_interpolate_linear_aa_backward"), | ||
|
|
||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you share some image examples with me of the image difference between PIL and your implementation?
Something like
plt.imshow(pil_interp - tv_interp)so that I can see what types of differences we are seeing here?