telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | |
| 6 | template<typename T, typename B> |
| 7 | LayerTestResult<T, 2> SimpleFullyConnectedTestImpl( |
| 8 | armnn::IWorkloadFactory& workloadFactory, |
| 9 | armnn::TensorInfo inputTensorInfo, |
| 10 | armnn::TensorInfo outputTensorInfo, |
| 11 | armnn::TensorInfo weightsDesc, |
| 12 | armnn::TensorInfo biasesDesc, |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 13 | boost::multi_array<T, 2>& weights, |
| 14 | boost::multi_array<B, 1>& bias, |
| 15 | boost::multi_array<T, 4>& input, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 16 | bool biasEnabled, |
| 17 | bool transposeWeights) |
| 18 | { |
| 19 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 20 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 21 | |
| 22 | armnn::FullyConnectedQueueDescriptor data; |
| 23 | armnn::WorkloadInfo info; |
| 24 | armnn::ScopedCpuTensorHandle weightsTensor(weightsDesc); |
| 25 | armnn::ScopedCpuTensorHandle biasTensor(biasesDesc); |
| 26 | |
| 27 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &weights[0][0]); |
| 28 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 29 | |
| 30 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 31 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 32 | data.m_Weight = &weightsTensor; |
| 33 | data.m_Bias = &biasTensor; |
| 34 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
| 35 | data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; |
| 36 | |
| 37 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateFullyConnected(data, info); |
| 38 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 39 | |
| 40 | inputHandle->Allocate(); |
| 41 | outputHandle->Allocate(); |
| 42 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 43 | |
Aron Virginas-Tar | 6057895 | 2018-10-31 11:04:01 +0000 | [diff] [blame^] | 44 | workloadFactory.Acquire(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 45 | workload->Execute(); |
Aron Virginas-Tar | 6057895 | 2018-10-31 11:04:01 +0000 | [diff] [blame^] | 46 | workloadFactory.Release(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 47 | |
| 48 | CopyDataFromITensorHandle(&result.output[0][0], outputHandle.get()); |
| 49 | |
| 50 | return result; |
| 51 | } |
| 52 | |
| 53 | LayerTestResult<float, 2> FullyConnectedFloat32Test(armnn::IWorkloadFactory& workloadFactory, bool biasEnabled, |
| 54 | bool transposeWeights) |
| 55 | { |
| 56 | unsigned int inputWidth = 1; |
| 57 | unsigned int inputHeight = 1; |
| 58 | unsigned int inputChannels = 5; |
| 59 | unsigned int inputNum = 2; |
| 60 | |
| 61 | unsigned int outputChannels = 3; |
| 62 | unsigned int outputNum = 2; |
| 63 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 64 | // Define the tensor descriptors. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 65 | armnn::TensorInfo inputTensorInfo; |
| 66 | armnn::TensorInfo outputTensorInfo; |
| 67 | armnn::TensorInfo weightsDesc; |
| 68 | armnn::TensorInfo biasesDesc; |
| 69 | |
| 70 | unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; |
| 71 | unsigned int outputShape[] = { outputNum, outputChannels }; |
| 72 | unsigned int weightsShape[] = { inputChannels, outputChannels }; |
| 73 | if (transposeWeights) |
| 74 | { |
| 75 | std::swap(weightsShape[0], weightsShape[1]); |
| 76 | } |
| 77 | unsigned int biasShape[] = { outputChannels }; |
| 78 | |
| 79 | inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32); |
| 80 | outputTensorInfo = armnn::TensorInfo(2, outputShape, armnn::DataType::Float32); |
| 81 | weightsDesc = armnn::TensorInfo(2, weightsShape, armnn::DataType::Float32); |
| 82 | biasesDesc = armnn::TensorInfo(1, biasShape, armnn::DataType::Float32); |
| 83 | |
| 84 | LayerTestResult<float, 2> result(outputTensorInfo); |
| 85 | |
| 86 | boost::multi_array<float, 4> input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>( |
| 87 | { |
| 88 | 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, |
| 89 | |
| 90 | 5.0f, 4.0f, 3.0f, 2.0f, 1.0f |
| 91 | }) |
| 92 | ); |
| 93 | |
| 94 | boost::multi_array<float, 2> weights = MakeTensor<float, 2>(weightsDesc, std::vector<float>( |
| 95 | { |
| 96 | .5f, 2.f, .5f, |
| 97 | .5f, 2.f, 1.f, |
| 98 | .5f, 2.f, 2.f, |
| 99 | .5f, 2.f, 3.f, |
| 100 | .5f, 2.f, 4.f |
| 101 | })); |
| 102 | |
| 103 | if (transposeWeights) |
| 104 | { |
| 105 | weights = MakeTensor<float, 2>(weightsDesc, std::vector<float>( |
| 106 | { |
| 107 | .5f, .5f, .5f, .5f, .5f, |
| 108 | 2.f, 2.f, 2.f, 2.f, 2.f, |
| 109 | .5f, 1.f, 2.f, 3.f, 4.f |
| 110 | })); |
| 111 | } |
| 112 | |
| 113 | |
| 114 | std::vector<float> biasValues({0.f, 0.f, 0.f}); |
| 115 | if (biasEnabled) |
| 116 | { |
| 117 | biasValues = std::vector<float>({10.f, 20.f, 30.f}); |
| 118 | } |
| 119 | boost::multi_array<float, 1> bias = MakeTensor<float, 1>(biasesDesc, biasValues); |
| 120 | |
| 121 | result = SimpleFullyConnectedTestImpl<float>( |
| 122 | workloadFactory, |
| 123 | inputTensorInfo, outputTensorInfo, |
| 124 | weightsDesc, biasesDesc, |
| 125 | weights, bias, input, |
| 126 | biasEnabled, transposeWeights |
| 127 | ); |
| 128 | |
| 129 | result.outputExpected = MakeTensor<float, 2>(outputTensorInfo, std::vector<float>( |
| 130 | { |
| 131 | 0.5f + 1.0f + 1.5f + 2.0f + 2.5f + biasValues[0], |
| 132 | 2.0f + 4.0f + 6.0f + 8.0f + 10.f + biasValues[1], |
| 133 | 0.5f + 2.0f + 6.0f + 12.f + 20.f + biasValues[2], |
| 134 | |
| 135 | 2.5f + 2.0f + 1.5f + 1.0f + 0.5f + biasValues[0], |
| 136 | 10.0f + 8.0f + 6.0f + 4.0f + 2.f + biasValues[1], |
| 137 | 2.5f + 4.0f + 6.0f + 6.f + 4.f + biasValues[2] |
| 138 | }) |
| 139 | ); |
| 140 | |
| 141 | return result; |
| 142 | } |
| 143 | |
| 144 | LayerTestResult<uint8_t, 2> FullyConnectedUint8Test(armnn::IWorkloadFactory& workloadFactory, bool biasEnabled) |
| 145 | { |
| 146 | constexpr static unsigned int inputWidth = 3u; |
| 147 | constexpr static unsigned int inputHeight = 2u; |
| 148 | constexpr static unsigned int inputChannels = 1u; |
| 149 | |
| 150 | constexpr static unsigned int inputSize = inputWidth * inputHeight * inputChannels; |
| 151 | |
| 152 | constexpr static unsigned int outputChannels = 2u; |
| 153 | |
| 154 | armnn::TensorInfo inputTensorInfo({ 1, inputChannels, inputHeight, inputWidth }, armnn::DataType::QuantisedAsymm8); |
| 155 | inputTensorInfo.SetQuantizationScale(0.1f); |
| 156 | inputTensorInfo.SetQuantizationOffset(63); |
| 157 | |
| 158 | armnn::TensorInfo outputTensorInfo({ 1, outputChannels }, armnn::DataType::QuantisedAsymm8); |
| 159 | outputTensorInfo.SetQuantizationScale(5.f); |
| 160 | outputTensorInfo.SetQuantizationOffset(biasEnabled ? -50 : 10); |
| 161 | |
| 162 | armnn::TensorInfo weightsDesc({ outputChannels, inputSize }, armnn::DataType::QuantisedAsymm8); |
| 163 | weightsDesc.SetQuantizationScale(0.2f); |
| 164 | weightsDesc.SetQuantizationOffset(93); |
| 165 | |
| 166 | armnn::TensorInfo biasesDesc({ outputChannels }, armnn::DataType::Signed32); |
| 167 | biasesDesc.SetQuantizationScale(inputTensorInfo.GetQuantizationScale() * weightsDesc.GetQuantizationScale()); |
| 168 | biasesDesc.SetQuantizationOffset(0); |
| 169 | |
| 170 | LayerTestResult<uint8_t, 2> result(outputTensorInfo); |
| 171 | |
| 172 | auto input = MakeTensor<uint8_t, 4>(inputTensorInfo, std::vector<uint8_t>{51, 124, 28, |
| 173 | 251, 8, 92}); |
| 174 | |
| 175 | auto weights = MakeTensor<uint8_t, 2>(weightsDesc, std::vector<uint8_t>{51, 193, 42, 53, 175, 34, |
| 176 | 210, 145, 23, 74, 34, 150}); |
| 177 | |
| 178 | // scale = 0.02 |
| 179 | // offset = 0 |
| 180 | auto bias = MakeTensor<int32_t, 1>(biasesDesc, std::vector<int32_t>{9250, 67500}); |
| 181 | |
| 182 | result = SimpleFullyConnectedTestImpl<uint8_t>( |
| 183 | workloadFactory, |
| 184 | inputTensorInfo, outputTensorInfo, |
| 185 | weightsDesc, biasesDesc, |
| 186 | weights, bias, input, |
| 187 | biasEnabled, true |
| 188 | ); |
| 189 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 190 | // Manually calculated. |
| 191 | // Note one of these values has been clamped to 0. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 192 | if (biasEnabled) |
| 193 | { |
| 194 | result.outputExpected = MakeTensor<uint8_t, 2>(outputTensorInfo, std::vector<uint8_t>{0, 242}); |
| 195 | } |
| 196 | else |
| 197 | { |
| 198 | result.outputExpected = MakeTensor<uint8_t, 2>(outputTensorInfo, std::vector<uint8_t>{0, 32}); |
| 199 | } |
| 200 | |
| 201 | return result; |
| 202 | } |
| 203 | |
| 204 | |
| 205 | |
| 206 | // |
| 207 | // ArmNN variant of the AndroidNN fully_connected_float_large test. |
| 208 | // |
| 209 | // Tests the fully connected layer with large values, optionally transposing weights. |
| 210 | // Note this is templated for consistency, but the nature of this tests makes it unlikely to be useful in Uint8 mode. |
| 211 | // |
| 212 | template<typename T> |
| 213 | LayerTestResult<T, 2> FullyConnectedLargeTestCommon(armnn::IWorkloadFactory& workloadFactory, |
| 214 | bool transposeWeights, |
| 215 | float qScale = 0.0f, |
| 216 | int32_t qOffset = 0) |
| 217 | { |
| 218 | unsigned int inputWidth = 1; |
| 219 | unsigned int inputHeight = 1; |
| 220 | unsigned int inputChannels = 5; |
| 221 | unsigned int inputNum = 1; |
| 222 | |
| 223 | unsigned int outputChannels = 1; |
| 224 | unsigned int outputNum = 1; |
| 225 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 226 | // Define the tensor descriptors. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 227 | armnn::TensorInfo inputTensorInfo; |
| 228 | armnn::TensorInfo outputTensorInfo; |
| 229 | armnn::TensorInfo weightsDesc; |
| 230 | armnn::TensorInfo biasesDesc; |
| 231 | |
| 232 | unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; |
| 233 | unsigned int outputShape[] = { outputNum, outputChannels }; |
| 234 | unsigned int weightsShape[] = { inputChannels, outputChannels }; |
| 235 | if (transposeWeights) |
| 236 | { |
| 237 | std::swap(weightsShape[0], weightsShape[1]); |
| 238 | } |
| 239 | |
| 240 | unsigned int biasShape[] = { outputChannels }; |
| 241 | |
| 242 | inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::GetDataType<T>()); |
| 243 | outputTensorInfo = armnn::TensorInfo(2, outputShape, armnn::GetDataType<T>()); |
| 244 | weightsDesc = armnn::TensorInfo(2, weightsShape, armnn::GetDataType<T>()); |
| 245 | biasesDesc = armnn::TensorInfo(1, biasShape, armnn::GetDataType<T>()); |
| 246 | |
| 247 | // Set quantization parameters if the requested type is a quantized type. |
| 248 | if(armnn::IsQuantizedType<T>()) |
| 249 | { |
| 250 | inputTensorInfo.SetQuantizationScale(qScale); |
| 251 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 252 | outputTensorInfo.SetQuantizationScale(qScale); |
| 253 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 254 | } |
| 255 | |
| 256 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 257 | |
| 258 | boost::multi_array<T, 4> input = MakeTensor<T, 4>(inputTensorInfo, |
| 259 | QuantizedVector<T>(qScale, qOffset, { |
| 260 | 1.0f, 10.0f, 100.0f, 1000.0f, 10000.0f, |
| 261 | }) |
| 262 | ); |
| 263 | |
| 264 | boost::multi_array<T, 2> weights = MakeTensor<T, 2>(weightsDesc, |
| 265 | QuantizedVector<T>(qScale, qOffset, { |
| 266 | 2.0f, 3.0f, 4.0f, 5.0f, 6.0f |
| 267 | }) |
| 268 | ); |
| 269 | |
| 270 | std::vector<T> biasValues({900000.f}); |
| 271 | boost::multi_array<T, 1> bias = MakeTensor<T, 1>(biasesDesc, biasValues); |
| 272 | |
| 273 | result = SimpleFullyConnectedTestImpl<T>( |
| 274 | workloadFactory, |
| 275 | inputTensorInfo, outputTensorInfo, |
| 276 | weightsDesc, biasesDesc, |
| 277 | weights, bias, input, |
| 278 | true, transposeWeights |
| 279 | ); |
| 280 | |
| 281 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, |
| 282 | QuantizedVector<T>(qScale, qOffset, { |
| 283 | 965432.0f, |
| 284 | }) |
| 285 | ); |
| 286 | |
| 287 | return result; |
| 288 | } |