diff options
author | Jan Eilers <jan.eilers@arm.com> | 2021-06-02 12:01:25 +0100 |
---|---|---|
committer | Jan Eilers <jan.eilers@arm.com> | 2021-06-16 11:31:42 +0000 |
commit | 53ef79504b4c881c572735393c2eede5fa556c46 (patch) | |
tree | f6e0cd27c4d03075fa154074c5b12d7c8c3149f7 /src/backends/reference/test/RefPerChannelDecoderTests.cpp | |
parent | 77fe76bfa8cb798943821d1f3e432c228e1cdee3 (diff) | |
download | armnn-53ef79504b4c881c572735393c2eede5fa556c46.tar.gz |
IVGCVSW-5826 Change weights layout for depthwise to [1,H,W,I*M]
* This change is necessary because tflite uses a [1,H,W,I*M] format
and uses the I*M dimension for per axis quantization. Our previous
layout [M,I,H,W] can't handle the correlating quantization scales.
* Updates Onnx-, TfLiteParser and TfliteDelegate
* Updates the CpuRef, CpuAcc and GpuAcc backends
* Adjusts unit tests
* Adds test to ensure models with old layout can still be read and
executed
* Adds conversion function to previous layout [1,H,W,I*M] --> [M,I,H,W]
which can be used by backend developers
!android-nn-driver:5553
Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: Ifef23368b8c3702cf315a5838d214f7dc13c0152
Diffstat (limited to 'src/backends/reference/test/RefPerChannelDecoderTests.cpp')
-rw-r--r-- | src/backends/reference/test/RefPerChannelDecoderTests.cpp | 156 |
1 files changed, 156 insertions, 0 deletions
diff --git a/src/backends/reference/test/RefPerChannelDecoderTests.cpp b/src/backends/reference/test/RefPerChannelDecoderTests.cpp new file mode 100644 index 0000000000..c2e3cee7a0 --- /dev/null +++ b/src/backends/reference/test/RefPerChannelDecoderTests.cpp @@ -0,0 +1,156 @@ +// +// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include <reference/workloads/Decoders.hpp> +#include <armnn/utility/NumericCast.hpp> + +#include <fmt/format.h> + +#include <boost/test/unit_test.hpp> + +BOOST_AUTO_TEST_SUITE(RefPerChannelDecoder) + +template<typename T> +void CompareVector(std::vector<T> vec1, std::vector<T> vec2) +{ + BOOST_TEST(vec1.size() == vec2.size()); + + bool mismatch = false; + for (uint i = 0; i < vec1.size(); ++i) + { + if (vec1[i] != vec2[i]) + { + /*std::stringstream ss; + ss << "Vector value mismatch: index=" << i << " " << vec1[i] << "!=" << vec2[i];*/ + BOOST_TEST_MESSAGE(fmt::format("Vector value mismatch: index={} {} != {}", + i, + vec1[i], + vec2[i])); + mismatch = true; + } + } + + if (mismatch) + { + BOOST_FAIL("Error in CompareVector. Vectors don't match."); + } +} + +// Ensure quantization works for none depthwise convolutions +BOOST_AUTO_TEST_CASE(RefPerChannelDecoderTest1) +{ + using namespace armnn; + std::vector<int8_t> input = + { + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 + }; + + std::vector<float> expOutput = + { + 0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f, + 24.0f, 26.0f, 28.0f, 30.0f, 32.0f, 34.0f, 36.0f, 38.0f, 40.0f, 42.0f, 44.0f, 46.0f + }; + + TensorInfo tensorInfo ({2,2,2,3},DataType::QSymmS8,{1.0f, 2.0f},0); + auto decoder = MakeDecoder<float>(tensorInfo, input.data()); + + std::vector<float> output = decoder->DecodeTensor(tensorInfo.GetShape()); + + CompareVector(output, expOutput); +} + +// Ensure quantization works for depthwise convolutions M=1 +BOOST_AUTO_TEST_CASE(RefPerChannelDecoderTest2) +{ + using namespace armnn; + std::vector<int8_t> input = + { + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 + }; + + std::vector<float> expOutput = + { + 0.0f, 1.0f, 2.0f, 3.0f, + 8.0f, 10.0f, 12.0f, 14.0f, + 24.0f, 27.0f, 30.0f, 33.0f, + 48.0f, 52.0f, 56.0f, 60.0f + }; + + // [O,1,H,W] = [I*M,1,H,W] = [4*1,1,2,2] + TensorInfo tensorInfo ({4,1,2,2},DataType::QSymmS8,{1.0f, 2.0f, 3.0f, 4.0f},0); + auto decoder = MakeDecoder<float>(tensorInfo, input.data()); + + std::vector<float> output = decoder->DecodeTensor(tensorInfo.GetShape(), true); + + CompareVector(output, expOutput); +} + +// Ensure quantization works for depthwise convolutions M=2 +BOOST_AUTO_TEST_CASE(RefPerChannelDecoderTest3) +{ + using namespace armnn; + std::vector<int8_t> input = + { + 0, 1, 2, 3, + 4, 5, 6, 7, + 8, 9, 10, 11, + 12, 13, 14, 15, + 16, 17, 18, 19, + 20, 21, 22, 23 + }; + + std::vector<float> expOutput = + { + 0.0f, 1.0f, 2.0f, 3.0f, + 8.0f, 10.0f, 12.0f, 14.0f, + 24.0f, 27.0f, 30.0f, 33.0f, + 48.0f, 52.0f, 56.0f, 60.0f, + 80.0f, 85.0f, 90.0f, 95.0f, + 120.0f, 126.0f, 132.0f, 138.0f + }; + + // [O,1,H,W] = [I*M,1,H,W] = [3*2,1,2,2] + TensorInfo tensorInfo ({6,1,2,2},DataType::QSymmS8,{1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f},0); + auto decoder = MakeDecoder<float>(tensorInfo, input.data()); + + std::vector<float> output = decoder->DecodeTensor(tensorInfo.GetShape(), true); + + CompareVector(output, expOutput); +} + +// Ensure quantization works for depthwise convolutions M=2 for int32 +BOOST_AUTO_TEST_CASE(RefPerChannelDecoderTest4) +{ + using namespace armnn; + std::vector<int32_t> input = + { + 0, 1, 2, 3, + 4, 5, 6, 7, + 8, 9, 10, 11, + 12, 13, 14, 15, + 16, 17, 18, 19, + 20, 21, 22, 23 + }; + + std::vector<float> expOutput = + { + 0.0f, 1.0f, 2.0f, 3.0f, + 8.0f, 10.0f, 12.0f, 14.0f, + 24.0f, 27.0f, 30.0f, 33.0f, + 48.0f, 52.0f, 56.0f, 60.0f, + 80.0f, 85.0f, 90.0f, 95.0f, + 120.0f, 126.0f, 132.0f, 138.0f + }; + + // [O,1,H,W] = [I*M,1,H,W] = [3*2,1,2,2] + TensorInfo tensorInfo ({6,1,2,2},DataType::Signed32,{1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f},0); + auto decoder = MakeDecoder<float>(tensorInfo, input.data()); + + std::vector<float> output = decoder->DecodeTensor(tensorInfo.GetShape(), true); + + CompareVector(output, expOutput); +} + +BOOST_AUTO_TEST_SUITE_END() |