From ba424d2aff70a13e1a16c4f9018a0bea4e5c11b3 Mon Sep 17 00:00:00 2001 From: josh minor Date: Wed, 13 Nov 2019 10:55:17 -0600 Subject: IVGCVSW-1530 Add TfLite slice parser and fix transpose perm vector creation * TfLite slice parser and relevant tests added * TfLite transpose parser logic added to translate Tf/np permutation vector definitions to Armnn definitions * TfLite transpose parser no permute data test modified to include data for default permutation vector when none specified Signed-off-by: josh minor Change-Id: Iebd30971bd180593dc6b8f0d5be1d1bc61a3a5bf --- src/armnnTfLiteParser/test/Slice.cpp | 176 +++++++++++++++++++++++++++++++ src/armnnTfLiteParser/test/Transpose.cpp | 55 +++++----- 2 files changed, 202 insertions(+), 29 deletions(-) create mode 100644 src/armnnTfLiteParser/test/Slice.cpp (limited to 'src/armnnTfLiteParser/test') diff --git a/src/armnnTfLiteParser/test/Slice.cpp b/src/armnnTfLiteParser/test/Slice.cpp new file mode 100644 index 0000000000..17d1b1a68c --- /dev/null +++ b/src/armnnTfLiteParser/test/Slice.cpp @@ -0,0 +1,176 @@ +// +// Copyright © 2019 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include +#include "ParserFlatbuffersFixture.hpp" +#include "../TfLiteParser.hpp" + +BOOST_AUTO_TEST_SUITE(TensorflowLiteParser) + +struct SliceFixture : public ParserFlatbuffersFixture +{ + explicit SliceFixture(const std::string & inputShape, + const std::string & outputShape, + const std::string & beginData, + const std::string & sizeData) + { + m_JsonString = R"( + { + "version": 3, + "operator_codes": [ + { + "builtin_code": "SLICE", + "version": 1 + } + ], + "subgraphs": [ + { + "tensors": [ + { + "shape": )" + inputShape + R"(, + "type": "FLOAT32", + "buffer": 0, + "name": "inputTensor", + "quantization": { + "min": [ + 0.0 + ], + "max": [ + 255.0 + ], + "details_type": 0, + "quantized_dimension": 0 + }, + "is_variable": false + }, + { + "shape": )" + outputShape + R"(, + "type": "FLOAT32", + "buffer": 1, + "name": "outputTensor", + "quantization": { + "details_type": 0, + "quantized_dimension": 0 + }, + "is_variable": false + })"; + m_JsonString += R"(, + { + "shape": [ + 3 + ], + "type": "INT32", + "buffer": 2, + "name": "beginTensor", + "quantization": { + } + })"; + m_JsonString += R"(, + { + "shape": [ + 3 + ], + "type": "INT32", + "buffer": 3, + "name": "sizeTensor", + "quantization": { + } + })"; + m_JsonString += R"(], + "inputs": [ + 0 + ], + "outputs": [ + 1 + ], + "operators": [ + { + "opcode_index": 0, + "inputs": [ + 0, + 2, + 3)"; + m_JsonString += R"(], + "outputs": [ + 1 + ], + mutating_variable_inputs: [ + ] + } + ] + } + ], + "description": "TOCO Converted.", + "buffers": [ + { }, + { })"; + m_JsonString += R"(,{"data": )" + beginData + R"( })"; + m_JsonString += R"(,{"data": )" + sizeData + R"( })"; + m_JsonString += R"( + ] + } + )"; + SetupSingleInputSingleOutput("inputTensor", "outputTensor"); + } +}; + +struct SliceFixtureSingleDim : SliceFixture +{ + SliceFixtureSingleDim() : SliceFixture("[ 3, 2, 3 ]", + "[ 1, 1, 3 ]", + "[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]", + "[ 1, 0, 0, 0, 1, 0, 0, 0, 3, 0, 0, 0 ]") {} +}; + +BOOST_FIXTURE_TEST_CASE(SliceSingleDim, SliceFixtureSingleDim) +{ + RunTest<3, armnn::DataType::Float32>( + 0, + {{"inputTensor", { 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6 }}}, + {{"outputTensor", { 3, 3, 3 }}}); + + BOOST_TEST((m_Parser->GetNetworkOutputBindingInfo(0, "outputTensor").second.GetShape() + == armnn::TensorShape({1,1,3}))); +} + +struct SliceFixtureD123 : SliceFixture +{ + SliceFixtureD123() : SliceFixture("[ 3, 2, 3 ]", + "[ 1, 2, 3 ]", + "[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]", + "[ 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0 ]") {} +}; + +BOOST_FIXTURE_TEST_CASE(SliceD123, SliceFixtureD123) +{ + RunTest<3, armnn::DataType::Float32>( + 0, + {{"inputTensor", { 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6 }}}, + {{"outputTensor", { 3, 3, 3, 4, 4, 4 }}}); + + BOOST_TEST((m_Parser->GetNetworkOutputBindingInfo(0, "outputTensor").second.GetShape() + == armnn::TensorShape({1,2,3}))); +} + +struct SliceFixtureD213 : SliceFixture +{ + SliceFixtureD213() : SliceFixture("[ 3, 2, 3 ]", + "[ 2, 1, 3 ]", + "[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]", + "[ 2, 0, 0, 0, 1, 0, 0, 0, 3, 0, 0, 0 ]") {} +}; + +BOOST_FIXTURE_TEST_CASE(SliceD213, SliceFixtureD213) +{ + RunTest<3, armnn::DataType::Float32>( + 0, + {{"inputTensor", { 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6 }}}, + {{"outputTensor", { 3, 3, 3, 5, 5, 5 }}}); + + BOOST_TEST((m_Parser->GetNetworkOutputBindingInfo(0, "outputTensor").second.GetShape() + == armnn::TensorShape({2,1,3}))); +} + +BOOST_AUTO_TEST_SUITE_END() \ No newline at end of file diff --git a/src/armnnTfLiteParser/test/Transpose.cpp b/src/armnnTfLiteParser/test/Transpose.cpp index 2e3190b62e..b2f953e75d 100644 --- a/src/armnnTfLiteParser/test/Transpose.cpp +++ b/src/armnnTfLiteParser/test/Transpose.cpp @@ -55,24 +55,20 @@ struct TransposeFixture : public ParserFlatbuffersFixture }, "is_variable": false })"; - if (!permuteData.empty()) - { - m_JsonString += R"(, - { - "shape": [ - 3 - ], - "type": "INT32", - "buffer": 2, - "name": "permuteTensor", - "quantization": { - "details_type": 0, - "quantized_dimension": 0 - }, - "is_variable": false - })"; - } - + m_JsonString += R"(, + { + "shape": [ + 3 + ], + "type": "INT32", + "buffer": 2, + "name": "permuteTensor", + "quantization": { + "details_type": 0, + "quantized_dimension": 0 + }, + "is_variable": false + })"; m_JsonString += R"(], "inputs": [ 0 @@ -85,10 +81,7 @@ struct TransposeFixture : public ParserFlatbuffersFixture "opcode_index": 0, "inputs": [ 0)"; - if (!permuteData.empty()) - { - m_JsonString += R"(,2)"; - } + m_JsonString += R"(,2)"; m_JsonString += R"(], "outputs": [ 1 @@ -117,6 +110,7 @@ struct TransposeFixture : public ParserFlatbuffersFixture } }; +// Note that this assumes the Tensorflow permutation vector implementation as opposed to the armnn implemenation. struct TransposeFixtureWithPermuteData : TransposeFixture { TransposeFixtureWithPermuteData() : TransposeFixture("[ 2, 2, 3 ]", @@ -128,29 +122,32 @@ BOOST_FIXTURE_TEST_CASE(TransposeWithPermuteData, TransposeFixtureWithPermuteDat { RunTest<3, armnn::DataType::Float32>( 0, - {{"inputTensor", { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 }}}, - {{"outputTensor", { 1, 4, 2, 5, 3, 6, 7, 10, 8, 11, 9, 12 }}}); + {{"inputTensor", { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 }}}, + {{"outputTensor", { 1, 4, 2, 5, 3, 6, 7, 10, 8, 11, 9, 12 }}}); BOOST_TEST((m_Parser->GetNetworkOutputBindingInfo(0, "outputTensor").second.GetShape() == armnn::TensorShape({2,3,2}))); } +// Tensorflow default permutation behavior assumes no permute argument will create permute vector [n-1...0], +// where n is the number of dimensions of the input tensor +// In this case we should get output shape 3,2,2 given default permutation vector 2,1,0 struct TransposeFixtureWithoutPermuteData : TransposeFixture { TransposeFixtureWithoutPermuteData() : TransposeFixture("[ 2, 2, 3 ]", - "", - "[ 2, 3, 2 ]") {} + "[ 2, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ]", + "[ 3, 2, 2 ]") {} }; BOOST_FIXTURE_TEST_CASE(TransposeWithoutPermuteDims, TransposeFixtureWithoutPermuteData) { RunTest<3, armnn::DataType::Float32>( 0, - {{"inputTensor", { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 }}}, - {{"outputTensor", { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 }}}); + {{"inputTensor", { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 }}}, + {{"outputTensor", { 1, 7, 4, 10, 2, 8, 5, 11, 3, 9, 6, 12 }}}); BOOST_TEST((m_Parser->GetNetworkOutputBindingInfo(0, "outputTensor").second.GetShape() - == armnn::TensorShape({2,3,2}))); + == armnn::TensorShape({3,2,2}))); } BOOST_AUTO_TEST_SUITE_END() \ No newline at end of file -- cgit v1.2.1