Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2019 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "QuantizedLstmEndToEndTestImpl.hpp" |
| 7 | |
| 8 | #include "CommonTestUtils.hpp" |
| 9 | #include "EndToEndTestImpl.hpp" |
| 10 | |
| 11 | #include <ResolveType.hpp> |
| 12 | |
| 13 | #include <armnn/INetwork.hpp> |
Matthew Bentham | 246bd46 | 2020-01-20 16:16:06 +0000 | [diff] [blame] | 14 | #include <armnn/QuantizedLstmParams.hpp> |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 15 | |
| 16 | #include <test/TensorHelpers.hpp> |
| 17 | |
| 18 | #include <boost/test/unit_test.hpp> |
| 19 | |
| 20 | #include <type_traits> |
| 21 | |
| 22 | namespace |
| 23 | { |
| 24 | |
| 25 | using MultiArray = const boost::multi_array<uint8_t, 2>&; |
| 26 | |
| 27 | armnn::INetworkPtr CreateQuantizedLstmNetwork(MultiArray input, |
| 28 | MultiArray expectedOutput) |
| 29 | { |
| 30 | auto batchSize = boost::numeric_cast<unsigned int>(input.shape()[0]); |
| 31 | auto inputSize = boost::numeric_cast<unsigned int>(input.shape()[1]); |
| 32 | auto outputSize = boost::numeric_cast<unsigned int>(expectedOutput.shape()[1]); |
| 33 | |
| 34 | float inputOutputScale = 0.0078125f; |
| 35 | int32_t inputOutputOffset = 128; |
| 36 | |
| 37 | float weightsScale = 0.00408021f; |
| 38 | int32_t weightsOffset = 100; |
| 39 | |
| 40 | float biasScale = 3.1876640625e-05f; |
| 41 | int32_t biasOffset = 0; |
| 42 | |
| 43 | float cellStateScale = 0.00048828125f; |
| 44 | int32_t cellStateOffset = 0; |
| 45 | |
| 46 | armnn::TensorInfo inputWeightsInfo({outputSize, inputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 47 | armnn::DataType::QAsymmU8, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 48 | weightsScale, |
| 49 | weightsOffset); |
| 50 | |
| 51 | armnn::TensorInfo recurrentWeightsInfo({outputSize, outputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 52 | armnn::DataType::QAsymmU8, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 53 | weightsScale, |
| 54 | weightsOffset); |
| 55 | |
| 56 | armnn::TensorInfo biasInfo({outputSize}, armnn::DataType::Signed32, biasScale, biasOffset); |
| 57 | |
| 58 | armnn::QuantizedLstmInputParams data; |
| 59 | |
| 60 | const std::vector<uint8_t> inputToInputWeightsVector = {146, 250, 235, 171, 10, 218, 171, 108}; |
| 61 | armnn::ConstTensor inputToInputWeightsTensor(inputWeightsInfo, inputToInputWeightsVector.data()); |
| 62 | |
| 63 | const std::vector<uint8_t> inputToForgetWeightsVector = {24, 50, 132, 179, 158, 110, 3, 169}; |
| 64 | armnn::ConstTensor inputToForgetWeightsTensor(inputWeightsInfo, inputToForgetWeightsVector.data()); |
| 65 | |
| 66 | const std::vector<uint8_t> inputToCellWeightsTensorVector = {133, 34, 29, 49, 206, 109, 54, 183}; |
| 67 | armnn::ConstTensor inputToCellWeightsTensor(inputWeightsInfo, inputToCellWeightsTensorVector.data()); |
| 68 | |
| 69 | const std::vector<uint8_t> inputToOutputWeightsTensorVector = {195, 187, 11, 99, 109, 10, 218, 48}; |
| 70 | armnn::ConstTensor inputToOutputWeightsTensor(inputWeightsInfo, inputToOutputWeightsTensorVector.data()); |
| 71 | |
| 72 | const std::vector<uint8_t> recurrentToInputWeightsTensorVector = |
| 73 | {254, 206, 77, 168, 71, 20, 215, 6, 223, 7, 118, 225, 59, 130, 174, 26}; |
| 74 | armnn::ConstTensor recurrentToInputWeightsTensor(recurrentWeightsInfo, recurrentToInputWeightsTensorVector.data()); |
| 75 | |
| 76 | const std::vector<uint8_t> recurrentToForgetWeightsTensorVector = |
| 77 | {137, 240, 103, 52, 68, 51, 237, 112, 0, 220, 89, 23, 69, 4, 207, 253}; |
| 78 | armnn::ConstTensor recurrentToForgetWeightsTensor(recurrentWeightsInfo, |
| 79 | recurrentToForgetWeightsTensorVector.data()); |
| 80 | |
| 81 | const std::vector<uint8_t> recurrentToCellWeightsTensorVector = |
| 82 | {172, 60, 205, 65, 14, 0, 140, 168, 240, 223, 133, 56, 142, 64, 246, 216}; |
| 83 | armnn::ConstTensor recurrentToCellWeightsTensor(recurrentWeightsInfo, recurrentToCellWeightsTensorVector.data()); |
| 84 | |
| 85 | const std::vector<uint8_t> recurrentToOutputWeightsTensorVector = |
| 86 | {106, 214, 67, 23, 59, 158, 45, 3, 119, 132, 49, 205, 129, 218, 11, 98}; |
| 87 | armnn::ConstTensor recurrentToOutputWeightsTensor(recurrentWeightsInfo, |
| 88 | recurrentToOutputWeightsTensorVector.data()); |
| 89 | |
| 90 | const std::vector<int32_t> inputGateBiasTensorVector = {-7876, 13488, -726, 32839}; |
| 91 | armnn::ConstTensor inputGateBiasTensor(biasInfo, inputGateBiasTensorVector.data()); |
| 92 | |
| 93 | const std::vector<int32_t> forgetGateBiasTensorVector = {9206, -46884, -11693, -38724}; |
| 94 | armnn::ConstTensor forgetGateBiasTensor(biasInfo, forgetGateBiasTensorVector.data()); |
| 95 | |
| 96 | const std::vector<int32_t> cellBiasTensorVector = {39481, 48624, 48976, -21419}; |
| 97 | armnn::ConstTensor cellBiasTensor(biasInfo, cellBiasTensorVector.data()); |
| 98 | |
| 99 | const std::vector<int32_t> outputGateBiasTensorVector = {-58999, -17050, -41852, -40538}; |
| 100 | armnn::ConstTensor outputGateBiasTensor(biasInfo, outputGateBiasTensorVector.data()); |
| 101 | |
| 102 | data.m_InputToInputWeights = &inputToInputWeightsTensor; |
| 103 | data.m_InputToForgetWeights = &inputToForgetWeightsTensor; |
| 104 | data.m_InputToCellWeights = &inputToCellWeightsTensor; |
| 105 | data.m_InputToOutputWeights = &inputToOutputWeightsTensor; |
| 106 | data.m_RecurrentToInputWeights = &recurrentToInputWeightsTensor; |
| 107 | data.m_RecurrentToForgetWeights = &recurrentToForgetWeightsTensor; |
| 108 | data.m_RecurrentToCellWeights = &recurrentToCellWeightsTensor; |
| 109 | data.m_RecurrentToOutputWeights = &recurrentToOutputWeightsTensor; |
| 110 | data.m_InputGateBias = &inputGateBiasTensor; |
| 111 | data.m_ForgetGateBias = &forgetGateBiasTensor; |
| 112 | data.m_CellBias = &cellBiasTensor; |
| 113 | data.m_OutputGateBias = &outputGateBiasTensor; |
| 114 | |
| 115 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 116 | |
| 117 | armnn::IConnectableLayer* const inputLayer = net->AddInputLayer(0); |
| 118 | armnn::IConnectableLayer* const cellStateIn = net->AddInputLayer(1); |
| 119 | armnn::IConnectableLayer* const outputStateIn = net->AddInputLayer(2); |
| 120 | armnn::IConnectableLayer* const quantizedLstmLayer = net->AddQuantizedLstmLayer(data, "quantizedLstm"); |
| 121 | armnn::IConnectableLayer* const cellStateOut = net->AddOutputLayer(0); |
| 122 | armnn::IConnectableLayer* const outputStateOut = net->AddOutputLayer(1); |
| 123 | |
| 124 | armnn::TensorInfo inputTensorInfo({batchSize , inputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 125 | armnn::DataType::QAsymmU8, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 126 | inputOutputScale, |
| 127 | inputOutputOffset); |
| 128 | |
| 129 | armnn::TensorInfo cellStateInTensorInfo({batchSize , outputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 130 | armnn::DataType::QSymmS16, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 131 | cellStateScale, |
| 132 | cellStateOffset); |
| 133 | |
| 134 | armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 135 | armnn::DataType::QAsymmU8, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 136 | inputOutputScale, |
| 137 | inputOutputOffset); |
| 138 | |
| 139 | armnn::TensorInfo cellStateOutTensorInfo({batchSize, outputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 140 | armnn::DataType::QSymmS16, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 141 | cellStateScale, |
| 142 | cellStateOffset); |
| 143 | |
| 144 | armnn::TensorInfo outputTensorInfo({batchSize, outputSize}, |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 145 | armnn::DataType::QAsymmU8, |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 146 | inputOutputScale, |
| 147 | inputOutputOffset); |
| 148 | |
| 149 | // connect up |
| 150 | // inputs |
| 151 | Connect(inputLayer, quantizedLstmLayer, inputTensorInfo, 0, 0); |
| 152 | Connect(cellStateIn, quantizedLstmLayer, cellStateInTensorInfo, 0, 1); |
| 153 | Connect(outputStateIn, quantizedLstmLayer, outputStateInTensorInfo, 0, 2); |
| 154 | |
| 155 | // outputs |
| 156 | Connect(quantizedLstmLayer, cellStateOut, cellStateOutTensorInfo, 0, 0); |
| 157 | Connect(quantizedLstmLayer, outputStateOut, outputTensorInfo, 1, 0); |
| 158 | |
| 159 | return net; |
| 160 | } |
| 161 | |
| 162 | // Checks if two values of an arithmetic type are close enough to each other |
| 163 | // with regard to a given tolerance value. |
| 164 | template<typename T> |
| 165 | typename std::enable_if<std::is_arithmetic<T>::value, bool>::type |
| 166 | IsCloseEnough(T value1, T value2, T tolerance) |
| 167 | { |
| 168 | if (tolerance < 0) |
| 169 | { |
| 170 | throw armnn::InvalidArgumentException("Tolerance cannot be < 0"); |
| 171 | } |
| 172 | |
| 173 | T diff = value1 >= value2 ? static_cast<T>(value1 - value2) : static_cast<T>(value2 - value1); |
| 174 | return diff <= tolerance; |
| 175 | } |
| 176 | |
| 177 | } // anonymous namespace |
| 178 | |
| 179 | void QuantizedLstmEndToEnd(const std::vector<armnn::BackendId>& backends) |
| 180 | { |
| 181 | std::vector<uint8_t> inputVector = {166, 179, 50, 150}; |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 182 | armnn::TensorInfo inputDesc({2, 2}, armnn::DataType::QAsymmU8); |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 183 | boost::multi_array<uint8_t, 2> input = MakeTensor<uint8_t, 2>(inputDesc, inputVector); |
| 184 | |
| 185 | std::vector<int16_t> cellStateInVector = {876, 1034, 955, -909, 761, 1029, 796, -1036}; |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 186 | armnn::TensorInfo cellStateInDesc({2, 4}, armnn::DataType::QSymmS16); |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 187 | boost::multi_array<int16_t, 2> cellStateIn = MakeTensor<int16_t, 2>(cellStateInDesc, cellStateInVector); |
| 188 | |
| 189 | std::vector<uint8_t> outputStateInVector = {136, 150, 140, 115, 135, 152, 138, 112}; |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 190 | armnn::TensorInfo outputStateInDesc({2, 4}, armnn::DataType::QAsymmU8); |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 191 | boost::multi_array<uint8_t, 2> outputStateIn = MakeTensor<uint8_t, 2>(outputStateInDesc, outputStateInVector); |
| 192 | |
| 193 | std::vector<int16_t> cellStateOutVector = {1485, 1177, 1373, -1023, 1019, 1355, 1097, -1235}; |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 194 | armnn::TensorInfo cellStateOutVectorDesc({2, 4}, armnn::DataType::QSymmS16); |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 195 | boost::multi_array<int16_t, 2> cellStateOut = MakeTensor<int16_t, 2>(cellStateOutVectorDesc, cellStateOutVector); |
| 196 | |
| 197 | std::vector<uint8_t> outputStateOutVector = {140, 151, 146, 112, 136, 156, 142, 112}; |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 198 | armnn::TensorInfo outputDesc({2, 4}, armnn::DataType::QAsymmU8); |
Aron Virginas-Tar | 46ff1ca | 2019-09-12 11:03:09 +0100 | [diff] [blame] | 199 | boost::multi_array<uint8_t, 2> outputStateOut = MakeTensor<uint8_t, 2>(outputDesc, outputStateOutVector); |
| 200 | |
| 201 | // Builds up the structure of the network |
| 202 | armnn::INetworkPtr net = CreateQuantizedLstmNetwork(input, outputStateOut); |
| 203 | |
| 204 | BOOST_TEST_CHECKPOINT("create a network"); |
| 205 | |
| 206 | IRuntime::CreationOptions options; |
| 207 | IRuntimePtr runtime(IRuntime::Create(options)); |
| 208 | |
| 209 | // optimize the network |
| 210 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 211 | |
| 212 | // Loads it into the runtime. |
| 213 | NetworkId netId; |
| 214 | runtime->LoadNetwork(netId, std::move(optNet)); |
| 215 | |
| 216 | InputTensors inputTensors; |
| 217 | inputTensors.reserve(3); |
| 218 | |
| 219 | // input |
| 220 | inputTensors.push_back({0, ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputVector.data())}); |
| 221 | inputTensors.push_back({1, ConstTensor(runtime->GetInputTensorInfo(netId, 1), cellStateInVector.data())}); |
| 222 | inputTensors.push_back({2, ConstTensor(runtime->GetInputTensorInfo(netId, 2), outputStateInVector.data())}); |
| 223 | |
| 224 | OutputTensors outputTensors; |
| 225 | outputTensors.reserve(2); |
| 226 | |
| 227 | //output |
| 228 | std::vector<int16_t > cellStateOutResult(cellStateOutVector.size()); |
| 229 | std::vector<uint8_t > outputStateOutResult(outputStateOutVector.size()); |
| 230 | outputTensors.push_back({0, Tensor(runtime->GetOutputTensorInfo(netId, 0), cellStateOutResult.data())}); |
| 231 | outputTensors.push_back({1, Tensor(runtime->GetOutputTensorInfo(netId, 1), outputStateOutResult.data())}); |
| 232 | |
| 233 | // Does the inference. |
| 234 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 235 | |
| 236 | // Checks the results |
| 237 | constexpr int16_t toleranceInt16 = 2; |
| 238 | for (unsigned int i = 0u; i < cellStateOutResult.size(); ++i) |
| 239 | { |
| 240 | BOOST_CHECK(IsCloseEnough(cellStateOutVector[i], cellStateOutResult[i], toleranceInt16)); |
| 241 | } |
| 242 | |
| 243 | constexpr uint8_t toleranceUint8 = 1; |
| 244 | for (unsigned int i = 0u; i < outputStateOutResult.size(); ++i) |
| 245 | { |
| 246 | BOOST_TEST(IsCloseEnough(outputStateOutVector[i], outputStateOutResult[i], toleranceUint8)); |
| 247 | } |
| 248 | } |