Gitiles
Code Review Sign In
review.mlplatform.org / ml / armnn / refs/tags/v22.05.01 / . / src / backends / backendsCommon / test / DepthwiseConvolution2dEndToEndTests.hpp
blob: 1f9b60a4f21fd62b7e3f0148a94a647966f215a3 [file] [log] [blame]
Cathal Corbett06902652022-04-14 17:55:11 +0100[diff] [blame]1//
2// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
3// SPDX-License-Identifier: MIT
4//
5#pragma once
6
7#include "EndToEndTestImpl.hpp"
8#include <armnnUtils/QuantizeHelper.hpp>
9
10#include <ResolveType.hpp>
11
12#include <CommonTestUtils.hpp>
13#include <armnnTestUtils/DataLayoutUtils.hpp>
14
15#include <map>
16#include <vector>
17
18namespace
19{
20
21armnn::INetworkPtr CreateDepthwiseConvolution2dNetwork(const armnn::DepthwiseConvolution2dDescriptor& descriptor,
22 const armnn::TensorInfo& inputInfo,
23 const armnn::TensorInfo& weightsInfo,
24 const armnn::TensorInfo& biasInfo,
25 const armnn::TensorInfo& outputInfo,
26 const armnn::ConstTensor& weights,
27 const armnn::ConstTensor& biases)
28{
29 using namespace armnn;
30
31 INetworkPtr network(INetwork::Create());
32 IConnectableLayer* input = network->AddInputLayer(0, "input");
33 armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weights, "Weights");
34 armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biases, "Bias");
35 IConnectableLayer* convolution2d = network->AddDepthwiseConvolution2dLayer(descriptor, "depthwiseConvolution2d");
36 IConnectableLayer* output = network->AddOutputLayer(0, "output");
37
38 Connect(input, convolution2d, inputInfo, 0, 0);
39 Connect(weightsLayer, convolution2d, weightsInfo, 0, 1);
40 Connect(biasLayer, convolution2d, biasInfo, 0, 2);
41 Connect(convolution2d, output, outputInfo, 0, 0);
42
43 return network;
44}
45
46} // anonymous namespace
47
48template<armnn::DataType ArmnnType, armnn::DataType ArmnnBType>
49void DepthwiseConvolution2dEndToEnd(const std::vector<armnn::BackendId>& backends,
50 armnn::DataLayout dataLayout)
51{
52 using namespace armnn;
53 using T = ResolveType<ArmnnType>;
54 using BT = ResolveType<ArmnnBType>;
55
56 const float qScale = IsQuantizedType<T>() ? 0.25f : 1.0f;
57 const int32_t qOffset = IsQuantizedType<T>() ? 50 : 0;
58
59 unsigned int depthMultiplier = 2;
60
61 unsigned int inputHeight = 8;
62 unsigned int inputWidth = 16;
63 unsigned int inputChannels = 2;
64 unsigned int inputBatchSize = 1;
65
66 unsigned int kernelHeight = 5;
67 unsigned int kernelWidth = 3;
68
69 unsigned int outputHeight = inputHeight - kernelHeight + 1 + 2;
70 unsigned int outputWidth = (inputWidth - kernelWidth + 1)/2;
71 unsigned int outputChannels = inputChannels * depthMultiplier;
72 unsigned int outputBatchSize = inputBatchSize;
73
74 TensorInfo inputInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, ArmnnType, qScale, qOffset, true);
75 TensorInfo outputInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, ArmnnType, qScale, qOffset);
76 TensorInfo weightsInfo({1, kernelHeight, kernelWidth, outputChannels}, ArmnnType, qScale, qOffset, true);
77 TensorInfo biasesInfo({outputChannels}, ArmnnBType, qScale * qScale, 0, true);
78
79 std::vector<float> inputData =
80 {
81 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
82 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
83 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
84 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
85 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
86 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
87 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
88 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f,
89 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
90 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
91 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
92 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
93 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
94 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
95 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
96 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f
97 };
98
99 std::vector<float> weightsData =
100 {
101 1.0f, 1.0f, 1.0f,
102 1.0f, -1.0f, 1.0f,
103 1.0f, 1.0f, 1.0f,
104 1.0f, 1.0f, 1.0f,
105 1.0f, 1.0f, 1.0f,
106
107 2.0f, 2.0f, 2.0f,
108 2.0f, 2.0f, 2.0f,
109 2.0f, 2.0f, 2.0f,
110 2.0f, 2.0f, 2.0f,
111 2.0f, 2.0f, 2.0f,
112
113 0.0f, 0.0f, 0.0f,
114 0.0f, -1.0f, 0.0f,
115 0.0f, 0.0f, 0.0f,
116 0.0f, 0.0f, 0.0f,
117 0.0f, 0.0f, 0.0f,
118
119 0.0f, 0.0f, 0.0f,
120 0.0f, 0.0f, 0.0f,
121 0.0f, 1.0f, 0.0f,
122 0.0f, 0.0f, 0.0f,
123 0.0f, 0.0f, 0.0f
124 };
125
126 std::vector<float> biasesData = { 0.0f, 2.0f, 1.0f, -1.0f };
127
128 std::vector<float> expectedOutputData =
129 {
130 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f,
131 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f,
132 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f,
133 2.5f, 2.5f, 2.5f, 2.5f, 2.5f, 2.5f, 2.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f,
134 4.5f, 4.5f, 4.5f, 4.5f, 4.5f, 4.5f, 4.5f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f,
135 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f,
136 1.0f, 3.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
137 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
138 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
139 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
140 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
141 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f
142 };
143
144 DepthwiseConvolution2dDescriptor descriptor;
145 descriptor.m_PadLeft = 0;
146 descriptor.m_PadRight = 0;
147 descriptor.m_PadTop = 1;
148 descriptor.m_PadBottom = 0;
149 descriptor.m_StrideX = 2;
150 descriptor.m_StrideY = 1;
151 descriptor.m_BiasEnabled = true;
152 descriptor.m_DataLayout = dataLayout;
153
154 // Permute input and output if NCDHW.
155 if (dataLayout == DataLayout::NCHW)
156 {
157 PermuteTensorNhwcToNchw(inputInfo, inputData);
158 PermuteTensorNhwcToNchw(outputInfo, expectedOutputData);
159 }
160
161 // Quantize data
162 std::vector<T> qInputData = armnnUtils::QuantizedVector<T>(inputData, qScale, qOffset);
163 std::vector<T> qWeightsData = armnnUtils::QuantizedVector<T>(weightsData, qScale, qOffset);
164 std::vector<T> qExpectedOutputData = armnnUtils::QuantizedVector<T>(expectedOutputData, qScale, qOffset);
165
166 std::vector<BT> qBiasesData = armnnUtils::QuantizedVector<BT>(biasesData, qScale * qScale, 0);
167
168 ConstTensor weights(weightsInfo, qWeightsData);
169 ConstTensor biases(biasesInfo, qBiasesData);
170
171 INetworkPtr network = CreateDepthwiseConvolution2dNetwork(descriptor,
172 inputInfo,
173 weightsInfo,
174 biasesInfo,
175 outputInfo,
176 weights,
177 biases);
178
179 EndToEndLayerTestImpl<ArmnnType, ArmnnType>(std::move(network),
180 { { 0, qInputData } },
181 { { 0, qExpectedOutputData } },
182 backends);
183}