Matthew Sloyan | eb5f810 | 2021-10-05 17:31:42 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2021 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ParserFlatbuffersFixture.hpp" |
| 7 | #include <sstream> |
| 8 | |
Matthew Sloyan | 4d217c0 | 2021-10-07 11:48:58 +0100 | [diff] [blame] | 9 | // Conv3D support was added in TF 2.5, so for backwards compatibility a hash define is needed. |
| 10 | #if defined(ARMNN_POST_TFLITE_2_3) |
Matthew Sloyan | eb5f810 | 2021-10-05 17:31:42 +0100 | [diff] [blame] | 11 | TEST_SUITE("TensorflowLiteParser_Conv3D") |
| 12 | { |
| 13 | struct SimpleConv3DFixture : public ParserFlatbuffersFixture |
| 14 | { |
| 15 | explicit SimpleConv3DFixture() |
| 16 | { |
| 17 | m_JsonString = R"( |
| 18 | { |
| 19 | "version": 3, |
| 20 | "operator_codes": [ { "builtin_code": "CONV_3D" } ], |
| 21 | "subgraphs": [ { |
| 22 | "tensors": [ |
| 23 | { |
| 24 | "shape": [ 1, 2, 3, 3, 1 ], |
| 25 | "type": "UINT8", |
| 26 | "buffer": 0, |
| 27 | "name": "inputTensor", |
| 28 | "quantization": { |
| 29 | "min": [ 0.0 ], |
| 30 | "max": [ 255.0 ], |
| 31 | "scale": [ 1.0 ], |
| 32 | "zero_point": [ 0 ], |
| 33 | } |
| 34 | }, |
| 35 | { |
| 36 | "shape": [ 1, 1, 1, 1, 1 ], |
| 37 | "type": "UINT8", |
| 38 | "buffer": 1, |
| 39 | "name": "outputTensor", |
| 40 | "quantization": { |
| 41 | "min": [ 0.0 ], |
| 42 | "max": [ 511.0 ], |
| 43 | "scale": [ 2.0 ], |
| 44 | "zero_point": [ 0 ], |
| 45 | } |
| 46 | }, |
| 47 | { |
| 48 | "shape": [ 2, 3, 3, 1, 1 ], |
| 49 | "type": "UINT8", |
| 50 | "buffer": 2, |
| 51 | "name": "filterTensor", |
| 52 | "quantization": { |
| 53 | "min": [ 0.0 ], |
| 54 | "max": [ 255.0 ], |
| 55 | "scale": [ 1.0 ], |
| 56 | "zero_point": [ 0 ], |
| 57 | } |
| 58 | } |
| 59 | ], |
| 60 | "inputs": [ 0 ], |
| 61 | "outputs": [ 1 ], |
| 62 | "operators": [ |
| 63 | { |
| 64 | "opcode_index": 0, |
| 65 | "inputs": [ 0, 2 ], |
| 66 | "outputs": [ 1 ], |
| 67 | "builtin_options_type": "Conv3DOptions", |
| 68 | "builtin_options": { |
| 69 | "padding": "VALID", |
| 70 | "stride_d": 1, |
| 71 | "stride_w": 1, |
| 72 | "stride_h": 1, |
| 73 | "fused_activation_function": "NONE" |
| 74 | }, |
| 75 | "custom_options_format": "FLEXBUFFERS" |
| 76 | } |
| 77 | ], |
| 78 | } ], |
| 79 | "buffers" : [ |
| 80 | { }, |
| 81 | { }, |
| 82 | { "data": [ 2,1,0, 6,2,1, 4,1,2, |
| 83 | 1,2,1, 2,0,2, 2,1,1 ], }, |
| 84 | { }, |
| 85 | ] |
| 86 | } |
| 87 | )"; |
| 88 | SetupSingleInputSingleOutput("inputTensor", "outputTensor"); |
| 89 | } |
| 90 | }; |
| 91 | |
| 92 | TEST_CASE_FIXTURE(SimpleConv3DFixture, "ParseSimpleConv3D") |
| 93 | { |
| 94 | RunTest<5, armnn::DataType::QAsymmU8>( |
| 95 | 0, |
| 96 | { |
| 97 | 1, 2, 3, |
| 98 | 4, 5, 6, |
| 99 | 7, 8, 9, |
| 100 | |
| 101 | 10, 11, 12, |
| 102 | 13, 14, 15, |
| 103 | 16, 17, 18, |
| 104 | }, |
| 105 | // Due to the output scaling we need to half the values. |
| 106 | { |
| 107 | (1*2 + 2*1 + 3*0 + |
| 108 | 4*6 + 5*2 + 6*1 + |
| 109 | 7*4 + 8*1 + 9*2 + |
| 110 | |
| 111 | 10*1 + 11*2 + 12*1 + |
| 112 | 13*2 + 14*0 + 15*2 + |
| 113 | 16*2 + 17*1 + 18*1) /2 |
| 114 | }); |
| 115 | } |
| 116 | struct Conv3DWithBiasesFixture : public ParserFlatbuffersFixture |
| 117 | { |
| 118 | explicit Conv3DWithBiasesFixture(const std::string& inputShape, |
| 119 | const std::string& outputShape, |
| 120 | const std::string& filterShape, |
| 121 | const std::string& filterData, |
| 122 | const std::string& biasShape, |
| 123 | const std::string& biasData, |
| 124 | const std::string& strides, |
| 125 | const std::string& activation="NONE", |
| 126 | const std::string& filterScale="1.0", |
| 127 | const std::string& filterZeroPoint="0", |
| 128 | const std::string& outputScale="1.0", |
| 129 | const std::string& outputZeroPoint="0") |
| 130 | { |
| 131 | m_JsonString = R"( |
| 132 | { |
| 133 | "version": 3, |
| 134 | "operator_codes": [ { "builtin_code": "CONV_3D" } ], |
| 135 | "subgraphs": [ { |
| 136 | "tensors": [ |
| 137 | { |
| 138 | "shape": )" + inputShape + R"(, |
| 139 | "type": "UINT8", |
| 140 | "buffer": 0, |
| 141 | "name": "inputTensor", |
| 142 | "quantization": { |
| 143 | "min": [ 0.0 ], |
| 144 | "max": [ 255.0 ], |
| 145 | "scale": [ 1.0 ], |
| 146 | "zero_point": [ 0 ], |
| 147 | } |
| 148 | }, |
| 149 | { |
| 150 | "shape": )" + outputShape + R"(, |
| 151 | "type": "UINT8", |
| 152 | "buffer": 1, |
| 153 | "name": "outputTensor", |
| 154 | "quantization": { |
| 155 | "min": [ 0.0 ], |
| 156 | "max": [ 511.0 ], |
| 157 | "scale": [ )" + outputScale + R"( ], |
| 158 | "zero_point": [ )" + outputZeroPoint + R"( ], |
| 159 | } |
| 160 | }, |
| 161 | { |
| 162 | "shape": )" + filterShape + R"( , |
| 163 | "type": "UINT8", |
| 164 | "buffer": 2, |
| 165 | "name": "filterTensor", |
| 166 | "quantization": { |
| 167 | "min": [ 0.0 ], |
| 168 | "max": [ 255.0 ], |
| 169 | "scale": [ )" + filterScale + R"( ], |
| 170 | "zero_point": [ )" + filterZeroPoint + R"( ], |
| 171 | } |
| 172 | }, |
| 173 | { |
| 174 | "shape": )" + biasShape + R"( , |
| 175 | "type": "INT32", |
| 176 | "buffer": 3, |
| 177 | "name": "biasTensor", |
| 178 | "quantization": { |
| 179 | "min": [ 0.0 ], |
| 180 | "max": [ 255.0 ], |
| 181 | "scale": [ 1.0 ], |
| 182 | "zero_point": [ 0 ], |
| 183 | } |
| 184 | } |
| 185 | ], |
| 186 | "inputs": [ 0 ], |
| 187 | "outputs": [ 1 ], |
| 188 | "operators": [ |
| 189 | { |
| 190 | "opcode_index": 0, |
| 191 | "inputs": [ 0, 2, 3 ], |
| 192 | "outputs": [ 1 ], |
| 193 | "builtin_options_type": "Conv3DOptions", |
| 194 | "builtin_options": { |
| 195 | "padding": "SAME", |
| 196 | "stride_d": )" + strides + R"(, |
| 197 | "stride_w": )" + strides + R"(, |
| 198 | "stride_h": )" + strides + R"(, |
| 199 | "fused_activation_function": )" + activation + R"( |
| 200 | }, |
| 201 | "custom_options_format": "FLEXBUFFERS" |
| 202 | } |
| 203 | ], |
| 204 | } ], |
| 205 | "buffers" : [ |
| 206 | { }, |
| 207 | { }, |
| 208 | { "data": )" + filterData + R"(, }, |
| 209 | { "data": )" + biasData + R"(, }, |
| 210 | ] |
| 211 | } |
| 212 | )"; |
| 213 | SetupSingleInputSingleOutput("inputTensor", "outputTensor"); |
| 214 | } |
| 215 | }; |
| 216 | |
| 217 | struct SimpleConv3DWithBiasesFixture : Conv3DWithBiasesFixture |
| 218 | { |
| 219 | SimpleConv3DWithBiasesFixture() |
| 220 | : Conv3DWithBiasesFixture("[ 1, 2, 2, 2, 1 ]", // inputShape |
| 221 | "[ 1, 2, 2, 2, 1 ]", // outputShape |
| 222 | "[ 2, 2, 2, 1, 1 ]", // filterShape |
| 223 | "[ 2,1, 1,0, 0,1, 1,1 ]", // filterData |
| 224 | "[ 1 ]", // biasShape |
| 225 | "[ 5, 0, 0, 0 ]", // biasData |
| 226 | "1") // stride d, w and h |
| 227 | {} |
| 228 | }; |
| 229 | |
| 230 | TEST_CASE_FIXTURE(SimpleConv3DWithBiasesFixture, "ParseConv3DWithBias") |
| 231 | { |
| 232 | RunTest<5, |
| 233 | armnn::DataType::QAsymmU8>(0, |
| 234 | { 1, 2, 3, 4, 5, 6, 7, 8 }, |
| 235 | { 33, 21, 23, 13, 28, 25, 27, 21 }); |
| 236 | } |
| 237 | |
| 238 | TEST_CASE_FIXTURE(SimpleConv3DWithBiasesFixture, "ParseDynamicConv3DWithBias") |
| 239 | { |
| 240 | RunTest<5, |
| 241 | armnn::DataType::QAsymmU8, |
| 242 | armnn::DataType::QAsymmU8>(0, |
| 243 | { { "inputTensor", { 2, 4, 6, 8, 10, 12, 14, 16 } } }, |
| 244 | { { "outputTensor", { 61, 37, 41, 21, 51, 45, 49, 37 } } }, |
| 245 | true); |
| 246 | } |
| 247 | |
| 248 | struct Relu6Conv3DWithBiasesFixture : Conv3DWithBiasesFixture |
| 249 | { |
| 250 | Relu6Conv3DWithBiasesFixture() |
| 251 | : Conv3DWithBiasesFixture("[ 1, 2, 2, 2, 1 ]", // inputShape |
| 252 | "[ 1, 2, 2, 2, 1 ]", // outputShape |
| 253 | "[ 2, 2, 2, 1, 1 ]", // filterShape |
| 254 | "[ 2,1, 1,0, 0,1, 1,1 ]", // filterData |
| 255 | "[ 1 ]", // biasShape |
| 256 | "[ 0, 0, 0, 0 ]", // biasData |
| 257 | "1", // stride d, w, and h |
| 258 | "RELU6", // activation |
| 259 | "1.0", // filter scale |
| 260 | "0", // filter zero point |
| 261 | "2.0", // output scale |
| 262 | "0") // output zero point |
| 263 | {} |
| 264 | }; |
| 265 | |
| 266 | TEST_CASE_FIXTURE(Relu6Conv3DWithBiasesFixture, "ParseConv3DAndRelu6WithBias") |
| 267 | { |
| 268 | uint8_t relu6Min = 6 / 2; // Divide by output scale |
| 269 | |
| 270 | RunTest<5, armnn::DataType::QAsymmU8>( |
| 271 | 0, |
| 272 | { |
| 273 | 1, 2, 3, 4, 5, 6, 7, 8 |
| 274 | }, |
| 275 | // RELU6 cuts output values at +6 |
| 276 | { |
| 277 | std::min(relu6Min, static_cast<uint8_t>((1*2 + 2*1 + 3*1 + 4*0 + 5*0 + 6*1 + 7*1 + 8*1)/2)), |
| 278 | std::min(relu6Min, static_cast<uint8_t>((2*2 + 0*1 + 0*1 + 0*0 + 0*0 + 0*1 + 8*1 + 0*1)/2)), |
| 279 | std::min(relu6Min, static_cast<uint8_t>((3*2 + 0*1 + 0*1 + 0*0 + 0*0 + 8*1 + 0*1 + 0*1)/2)), |
| 280 | std::min(relu6Min, static_cast<uint8_t>((4*2 + 0*1 + 0*1 + 0*0 + 8*0 + 0*1 + 0*1 + 0*1)/2)), |
| 281 | std::min(relu6Min, static_cast<uint8_t>((5*2 + 0*1 + 0*1 + 8*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)), |
| 282 | std::min(relu6Min, static_cast<uint8_t>((6*2 + 0*1 + 8*1 + 0*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)), |
| 283 | std::min(relu6Min, static_cast<uint8_t>((7*2 + 8*1 + 0*1 + 0*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)), |
| 284 | std::min(relu6Min, static_cast<uint8_t>((8*2 + 0*1 + 0*1 + 0*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)) |
| 285 | }); |
| 286 | } |
| 287 | |
| 288 | } |
Matthew Sloyan | 4d217c0 | 2021-10-07 11:48:58 +0100 | [diff] [blame] | 289 | #endif |