telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [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 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 4 | // |
| 5 | |
| 6 | #include <boost/test/unit_test.hpp> |
| 7 | #include "ParserFlatbuffersFixture.hpp" |
| 8 | #include "../TfLiteParser.hpp" |
| 9 | #include <sstream> |
| 10 | |
| 11 | BOOST_AUTO_TEST_SUITE(TensorflowLiteParser) |
| 12 | |
| 13 | struct SimpleConv2DFixture : public ParserFlatbuffersFixture |
| 14 | { |
| 15 | explicit SimpleConv2DFixture() |
| 16 | { |
| 17 | m_JsonString = R"( |
| 18 | { |
| 19 | "version": 3, |
| 20 | "operator_codes": [ { "builtin_code": "CONV_2D" } ], |
| 21 | "subgraphs": [ { |
| 22 | "tensors": [ |
| 23 | { |
| 24 | "shape": [ 1, 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 ], |
| 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": [ 1, 3, 3, 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": "Conv2DOptions", |
| 68 | "builtin_options": { |
| 69 | "padding": "VALID", |
| 70 | "stride_w": 1, |
| 71 | "stride_h": 1, |
| 72 | "fused_activation_function": "NONE" |
| 73 | }, |
| 74 | "custom_options_format": "FLEXBUFFERS" |
| 75 | } |
| 76 | ], |
| 77 | } ], |
| 78 | "buffers" : [ |
| 79 | { }, |
| 80 | { }, |
| 81 | { "data": [ 2,1,0, 6,2,1, 4,1,2 ], }, |
| 82 | { }, |
| 83 | ] |
| 84 | } |
| 85 | )"; |
| 86 | SetupSingleInputSingleOutput("inputTensor", "outputTensor"); |
| 87 | } |
| 88 | }; |
| 89 | |
| 90 | BOOST_FIXTURE_TEST_CASE( ParseSimpleConv2D, SimpleConv2DFixture ) |
| 91 | { |
| 92 | RunTest<4, uint8_t>( |
| 93 | 0, |
| 94 | { |
| 95 | 1, 2, 3, |
| 96 | 4, 5, 6, |
| 97 | 7, 8, 9, |
| 98 | }, |
| 99 | // because of the output scaling we need to take half of the values |
| 100 | { |
| 101 | (1*2 + 2*1 + 3*0 + |
| 102 | 4*6 + 5*2 + 6*1 + |
| 103 | 7*4 + 8*1 + 9*2) /2 |
| 104 | }); |
| 105 | } |
| 106 | |
| 107 | struct Conv2DWithBiasesFixture : public ParserFlatbuffersFixture |
| 108 | { |
| 109 | explicit Conv2DWithBiasesFixture(const std::string & inputShape, |
| 110 | const std::string & outputShape, |
| 111 | const std::string & filterShape, |
| 112 | const std::string & filterData, |
| 113 | const std::string & biasShape, |
| 114 | const std::string & biasData, |
| 115 | const std::string & strides, |
| 116 | const std::string & activation="NONE", |
| 117 | const std::string & filterScale="1.0", |
| 118 | const std::string & filterZeroPoint="0", |
| 119 | const std::string & outputScale="2.0", |
| 120 | const std::string & outputZeroPoint="0") |
| 121 | { |
| 122 | m_JsonString = R"( |
| 123 | { |
| 124 | "version": 3, |
| 125 | "operator_codes": [ { "builtin_code": "CONV_2D" } ], |
| 126 | "subgraphs": [ { |
| 127 | "tensors": [ |
| 128 | { |
| 129 | "shape": )" + inputShape + R"(, |
| 130 | "type": "UINT8", |
| 131 | "buffer": 0, |
| 132 | "name": "inputTensor", |
| 133 | "quantization": { |
| 134 | "min": [ 0.0 ], |
| 135 | "max": [ 255.0 ], |
| 136 | "scale": [ 1.0 ], |
| 137 | "zero_point": [ 0 ], |
| 138 | } |
| 139 | }, |
| 140 | { |
| 141 | "shape": )" + outputShape + R"(, |
| 142 | "type": "UINT8", |
| 143 | "buffer": 1, |
| 144 | "name": "outputTensor", |
| 145 | "quantization": { |
| 146 | "min": [ 0.0 ], |
| 147 | "max": [ 511.0 ], |
| 148 | "scale": [ )" + outputScale + R"( ], |
| 149 | "zero_point": [ )" + outputZeroPoint + R"( ], |
| 150 | } |
| 151 | }, |
| 152 | { |
| 153 | "shape": )" + filterShape + R"( , |
| 154 | "type": "UINT8", |
| 155 | "buffer": 2, |
| 156 | "name": "filterTensor", |
| 157 | "quantization": { |
| 158 | "min": [ 0.0 ], |
| 159 | "max": [ 255.0 ], |
| 160 | "scale": [ )" + filterScale + R"( ], |
| 161 | "zero_point": [ )" + filterZeroPoint + R"( ], |
| 162 | } |
| 163 | }, |
| 164 | { |
| 165 | "shape": )" + biasShape + R"( , |
| 166 | "type": "INT32", |
| 167 | "buffer": 3, |
| 168 | "name": "biasTensor", |
| 169 | "quantization": { |
| 170 | "min": [ 0.0 ], |
| 171 | "max": [ 255.0 ], |
| 172 | "scale": [ 1.0 ], |
| 173 | "zero_point": [ 0 ], |
| 174 | } |
| 175 | } |
| 176 | ], |
| 177 | "inputs": [ 0 ], |
| 178 | "outputs": [ 1 ], |
| 179 | "operators": [ |
| 180 | { |
| 181 | "opcode_index": 0, |
| 182 | "inputs": [ 0, 2, 3 ], |
| 183 | "outputs": [ 1 ], |
| 184 | "builtin_options_type": "Conv2DOptions", |
| 185 | "builtin_options": { |
| 186 | "padding": "SAME", |
| 187 | "stride_w": )" + strides + R"(, |
| 188 | "stride_h": )" + strides + R"(, |
| 189 | "fused_activation_function": )" + activation + R"( |
| 190 | }, |
| 191 | "custom_options_format": "FLEXBUFFERS" |
| 192 | } |
| 193 | ], |
| 194 | } ], |
| 195 | "buffers" : [ |
| 196 | { }, |
| 197 | { }, |
| 198 | { "data": )" + filterData + R"(, }, |
| 199 | { "data": )" + biasData + R"(, }, |
| 200 | ] |
| 201 | } |
| 202 | )"; |
| 203 | SetupSingleInputSingleOutput("inputTensor", "outputTensor"); |
| 204 | } |
| 205 | }; |
| 206 | |
| 207 | struct SimpleConv2DWithBiasesFixture : Conv2DWithBiasesFixture |
| 208 | { |
| 209 | SimpleConv2DWithBiasesFixture() |
| 210 | : Conv2DWithBiasesFixture("[ 1, 2, 2, 1 ]", // inputShape |
| 211 | "[ 1, 2, 2, 1 ]", // outputShape |
| 212 | "[ 1, 2, 2, 1 ]", // filterShape |
| 213 | "[ 2,1, 0,6 ]", // filterData |
| 214 | "[ 1 ]", // biasShape |
| 215 | "[ 10, 0, 0, 0 ]", // biasData |
| 216 | "1") // stride w and h |
| 217 | {} |
| 218 | }; |
| 219 | |
| 220 | BOOST_FIXTURE_TEST_CASE( ParseConv2DWithBias, SimpleConv2DWithBiasesFixture ) |
| 221 | { |
| 222 | RunTest<4, uint8_t>( |
| 223 | 0, |
| 224 | { |
| 225 | 1, 2, |
| 226 | 3, 4, |
| 227 | }, |
| 228 | // because of the output scaling we need to take half of the values |
| 229 | { |
| 230 | (1*2 + 2*1 + 3*0 + 4*6 + 10)/2, |
| 231 | (2*2 + 0*1 + 4*0 + 0*6 + 10)/2, |
| 232 | (3*2 + 4*1 + 0*0 + 0*6 + 10)/2, |
| 233 | (4*2 + 0*1 + 0*0 + 0*6 + 10)/2 |
| 234 | }); |
| 235 | } |
| 236 | |
| 237 | struct Conv2DShapeTestFixture : Conv2DWithBiasesFixture |
| 238 | { |
| 239 | static std::string GenerateInts(unsigned int n) |
| 240 | { |
| 241 | std::stringstream ss; |
| 242 | ss << " [ "; |
| 243 | for( unsigned int i=0; i<n; ++i ) { |
| 244 | if (i > 0 ) |
| 245 | { |
| 246 | ss << " , "; |
| 247 | } |
| 248 | ss << " " << (i%256); |
| 249 | } |
| 250 | ss << " ] "; |
| 251 | return ss.str(); |
| 252 | } |
| 253 | |
| 254 | Conv2DShapeTestFixture() |
| 255 | : Conv2DWithBiasesFixture("[ 1, 224, 224, 3 ]", // inputShape |
| 256 | "[ 1, 112, 112, 32 ]", // outputShape |
| 257 | "[ 32, 3, 3, 3 ]", // filterShape |
| 258 | GenerateInts(32*3*3*3), // filterData |
| 259 | "[ 32 ]", // biasShape |
| 260 | GenerateInts(32*4), // biasData |
| 261 | "2") // stride w and h |
| 262 | {} |
| 263 | }; |
| 264 | |
| 265 | BOOST_FIXTURE_TEST_CASE( ParseConv2D_112x112_out, Conv2DShapeTestFixture ) |
| 266 | { |
| 267 | } |
| 268 | |
| 269 | struct ReluConv2DWithBiasesFixture : Conv2DWithBiasesFixture |
| 270 | { |
| 271 | ReluConv2DWithBiasesFixture() |
| 272 | : Conv2DWithBiasesFixture("[ 1, 2, 2, 1 ]", // inputShape |
| 273 | "[ 1, 2, 2, 1 ]", // outputShape |
| 274 | "[ 1, 2, 2, 1 ]", // filterShape |
| 275 | "[ 2,1, 0,6 ]", // filterData |
| 276 | "[ 1 ]", // biasShape |
| 277 | "[ 16, 0, 0, 0 ]", // biasData |
| 278 | "1", // stride w and h |
| 279 | "RELU", // activation |
| 280 | "1.0", // filter scale |
| 281 | "4", // filter zero point |
| 282 | "2.0", // output scale |
| 283 | "20") // output zero point |
| 284 | {} |
| 285 | }; |
| 286 | |
| 287 | BOOST_FIXTURE_TEST_CASE( ParseConv2DAndReluWithBias, ReluConv2DWithBiasesFixture ) |
| 288 | { |
| 289 | uint8_t bias = 16; |
| 290 | uint8_t outZero = 20; |
| 291 | uint8_t fz = 4; // filter zero point |
| 292 | |
| 293 | RunTest<4, uint8_t>( |
| 294 | 0, |
| 295 | { |
| 296 | 1, 2, |
| 297 | 4, 8, |
| 298 | }, |
| 299 | // factors to consider: |
| 300 | // - the filter zero point is non zero, hence the (x-fz) |
| 301 | // - the output scale is 2 hence the /2 |
| 302 | // - output zero point is non zero, hence the +outZero |
| 303 | // - RELU cuts negative values and then we add the output zero point |
| 304 | { |
| 305 | std::max(outZero, static_cast<uint8_t>((1*(2-fz) + 2*(1-fz) + 4*(0-fz) + 8*(6-fz) + bias)/2 + outZero)), |
| 306 | std::max(outZero, static_cast<uint8_t>((2*(2-fz) + 0*(1-fz) + 8*(0-fz) + 0*(6-fz) + bias)/2 + outZero)), |
| 307 | std::max(outZero, static_cast<uint8_t>((4*(2-fz) + 8*(1-fz) + 0*(0-fz) + 0*(6-fz) + bias)/2 + outZero)), |
| 308 | std::max(outZero, static_cast<uint8_t>((8*(2-fz) + 0*(1-fz) + 0*(0-fz) + 0*(6-fz) + bias)/2 + outZero)) |
| 309 | }); |
| 310 | } |
| 311 | |
| 312 | struct Relu6Conv2DWithBiasesFixture : Conv2DWithBiasesFixture |
| 313 | { |
| 314 | Relu6Conv2DWithBiasesFixture() |
| 315 | : Conv2DWithBiasesFixture("[ 1, 2, 2, 1 ]", // inputShape |
| 316 | "[ 1, 2, 2, 1 ]", // outputShape |
| 317 | "[ 1, 2, 2, 1 ]", // filterShape |
| 318 | "[ 2,1, 0,6 ]", // filterData |
| 319 | "[ 1 ]", // biasShape |
| 320 | "[ 0, 0, 0, 0 ]", // biasData |
| 321 | "1", // stride w and h |
| 322 | "RELU6", // activation |
| 323 | "1.0", // filter scale |
| 324 | "0", // filter zero point |
| 325 | "2.0", // output scale |
| 326 | "0") // output zero point |
| 327 | {} |
| 328 | }; |
| 329 | |
| 330 | BOOST_FIXTURE_TEST_CASE( ParseConv2DAndRelu6WithBias, Relu6Conv2DWithBiasesFixture ) |
| 331 | { |
| 332 | uint8_t relu6Min = 6 / 2; // divide by output scale |
| 333 | |
| 334 | RunTest<4, uint8_t>( |
| 335 | 0, |
| 336 | { |
| 337 | 1, 2, |
| 338 | 4, 1, |
| 339 | }, |
| 340 | // factors to consider: |
| 341 | // - the output scale is 2 hence the /2 |
| 342 | // - RELU6 cuts output values at +6 |
| 343 | { |
| 344 | std::min(relu6Min, static_cast<uint8_t>((1*2 + 2*1 + 4*0 + 1*6)/2)), |
| 345 | std::min(relu6Min, static_cast<uint8_t>((2*2 + 0*1 + 1*0 + 0*6)/2)), |
| 346 | std::min(relu6Min, static_cast<uint8_t>((4*2 + 1*1 + 0*0 + 0*6)/2)), |
| 347 | std::min(relu6Min, static_cast<uint8_t>((1*2 + 0*1 + 0*0 + 0*6)/2)) |
| 348 | }); |
| 349 | } |
| 350 | |
| 351 | BOOST_AUTO_TEST_SUITE_END() |