Support trilinear sampling in Resize CPU and CUDA kernels #5300
Conversation
| @@ -363,8 +385,8 @@ void UpsampleBilinear(int64_t batch_size, | |||
| in_x1[x] = std::min(static_cast<int64_t>(in_x), input_width - 1); | |||
| in_x2[x] = std::min(in_x1[x] + 1, input_width - 1); | |||
|
|
|||
| dx1[x] = std::abs(in_x - in_x1[x]); | |||
| dx2[x] = std::abs(in_x - in_x2[x]); | |||
| dx1[x] = std::fabs(in_x - in_x1[x]); | |||
There was a problem hiding this comment.
For consistency's sake - we use fabs everywhere else in this file
| @@ -197,7 +197,66 @@ TEST(ResizeOpTest, ResizeOpLineartUpSampleTest_2DBilinear_align_corners) { | |||
| test.Run(); | |||
| } | |||
|
|
|||
| TEST(ResizeOpTest, ResizeOpLineartScalesNoOpTest) { | |||
| TEST(ResizeOpTest, ResizeOpLinearDownSampleTest_3DTrilinear_pytorch_half_pixel) { | |||
There was a problem hiding this comment.
3D input + downsample
| test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider}); // TensorRT: results mismatch | ||
| } | ||
|
|
||
| TEST(ResizeOpTest, ResizeOpLinearUpSampleTest_5DTrilinear_pytorch_half_pixel) { |
There was a problem hiding this comment.
5D input + upsample
| @@ -23,7 +23,7 @@ void ResizeImpl( | |||
| TArray<int64_t>& input_strides, | |||
| TArray<fast_divmod>& output_div_pitches, | |||
| TArray<float>& scales_vals, | |||
| TArray<float>& roi, | |||
| TArray<float, 10>& roi, | |||
There was a problem hiding this comment.
default capacity is 8 which doesn't compile when input is 5D and roi vector contains 10 elements
| @@ -310,24 +310,46 @@ void UpsampleBilinear(int64_t batch_size, | |||
| AllocatorPtr& alloc, | |||
| GetOriginalCoordinateFunc get_original_coordinate) { | |||
| std::vector<float> y_original; | |||
| y_original.reserve(output_height); | |||
There was a problem hiding this comment.
best to reserve memory - we know the number of elements that we are going to put into this container.
|
Can you validate these newly supported scenarios using onnx's implementation of resize. Basically use the test data from your test cases with the resize implementation here https://github.com/onnx/onnx/blob/master/onnx/backend/test/case/node/resize.py and see whether the output's match. |
I believe I did that for the upsampling test. Let me run the downsampling test data for completeness’ sake. |
I ran both the test inputs to the |
Description: Support 3-D input (or 5-D input with outermost 2 scales being 1.0 - batched multi-channel volumes [N,C, D, H, W]) for the "Linear" mode in the CPU and CUDA Resize kernels
Motivation and Context
Support a first party ONNX model in ORT