Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017-2018 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
Georgios Pinitas | d9eb275 | 2018-04-03 13:44:29 +0100 | [diff] [blame] | 24 | #include "arm_compute/graph.h" |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 25 | #include "support/ToolchainSupport.h" |
| 26 | #include "utils/GraphUtils.h" |
| 27 | #include "utils/Utils.h" |
| 28 | |
Georgios Pinitas | d9eb275 | 2018-04-03 13:44:29 +0100 | [diff] [blame] | 29 | #include <cstdlib> |
| 30 | |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 31 | using namespace arm_compute; |
Georgios Pinitas | d9eb275 | 2018-04-03 13:44:29 +0100 | [diff] [blame] | 32 | using namespace arm_compute::utils; |
| 33 | using namespace arm_compute::graph::frontend; |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 34 | using namespace arm_compute::graph_utils; |
| 35 | |
| 36 | /** Example demonstrating how to implement QASYMM8 MobileNet's network using the Compute Library's graph API |
| 37 | * |
| 38 | * @param[in] argc Number of arguments |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 39 | * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] npy_input, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) ) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 40 | */ |
Georgios Pinitas | d9eb275 | 2018-04-03 13:44:29 +0100 | [diff] [blame] | 41 | class GraphMobileNetQASYMM8Example : public Example |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 42 | { |
| 43 | public: |
| 44 | void do_setup(int argc, char **argv) override |
| 45 | { |
| 46 | std::string data_path; /* Path to the trainable data */ |
| 47 | std::string input; /* Image data */ |
| 48 | std::string label; /* Label data */ |
| 49 | |
| 50 | // Quantization info taken from the AndroidNN QASYMM8 MobileNet example |
| 51 | const QuantizationInfo in_quant_info = QuantizationInfo(0.0078125f, 128); |
| 52 | const QuantizationInfo mid_quant_info = QuantizationInfo(0.0784313753247f, 128); |
| 53 | |
| 54 | const std::vector<QuantizationInfo> conv_weights_quant_info = |
| 55 | { |
| 56 | QuantizationInfo(0.031778190285f, 156), // conv0 |
| 57 | QuantizationInfo(0.00604454148561f, 66) // conv14 |
| 58 | }; |
| 59 | |
| 60 | const std::vector<QuantizationInfo> depth_weights_quant_info = |
| 61 | { |
| 62 | QuantizationInfo(0.254282623529f, 129), // dwsc1 |
| 63 | QuantizationInfo(0.12828284502f, 172), // dwsc2 |
| 64 | QuantizationInfo(0.265911251307f, 83), // dwsc3 |
| 65 | QuantizationInfo(0.0985597148538f, 30), // dwsc4 |
| 66 | QuantizationInfo(0.0631204470992f, 54), // dwsc5 |
| 67 | QuantizationInfo(0.0137207424268f, 141), // dwsc6 |
| 68 | QuantizationInfo(0.0817828401923f, 125), // dwsc7 |
| 69 | QuantizationInfo(0.0393880493939f, 164), // dwsc8 |
| 70 | QuantizationInfo(0.211694166064f, 129), // dwsc9 |
| 71 | QuantizationInfo(0.158015936613f, 103), // dwsc10 |
| 72 | QuantizationInfo(0.0182712618262f, 137), // dwsc11 |
| 73 | QuantizationInfo(0.0127998134121f, 134), // dwsc12 |
| 74 | QuantizationInfo(0.299285322428f, 161) // dwsc13 |
| 75 | }; |
| 76 | |
| 77 | const std::vector<QuantizationInfo> point_weights_quant_info = |
| 78 | { |
| 79 | QuantizationInfo(0.0425766184926f, 129), // dwsc1 |
| 80 | QuantizationInfo(0.0250773020089f, 94), // dwsc2 |
| 81 | QuantizationInfo(0.015851572156f, 93), // dwsc3 |
| 82 | QuantizationInfo(0.0167811904103f, 98), // dwsc4 |
| 83 | QuantizationInfo(0.00951790809631f, 135), // dwsc5 |
| 84 | QuantizationInfo(0.00999817531556f, 128), // dwsc6 |
| 85 | QuantizationInfo(0.00590536883101f, 126), // dwsc7 |
| 86 | QuantizationInfo(0.00576109671965f, 133), // dwsc8 |
| 87 | QuantizationInfo(0.00830461271107f, 142), // dwsc9 |
| 88 | QuantizationInfo(0.0152327232063f, 72), // dwsc10 |
| 89 | QuantizationInfo(0.00741417845711f, 125), // dwsc11 |
| 90 | QuantizationInfo(0.0135628981516f, 142), // dwsc12 |
| 91 | QuantizationInfo(0.0338749065995f, 140) // dwsc13 |
| 92 | }; |
| 93 | |
Michele Di Giorgio | e3fba0a | 2018-02-14 14:18:01 +0000 | [diff] [blame] | 94 | // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 95 | const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; |
| 96 | Target target_hint = set_target_hint(target); |
| 97 | FastMathHint fast_math_hint = FastMathHint::DISABLED; |
Michele Di Giorgio | e3fba0a | 2018-02-14 14:18:01 +0000 | [diff] [blame] | 98 | |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 99 | // Parse arguments |
| 100 | if(argc < 2) |
| 101 | { |
| 102 | // Print help |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 103 | std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [npy_input] [labels] [fast_math_hint]\n\n"; |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 104 | std::cout << "No data folder provided: using random values\n\n"; |
| 105 | } |
| 106 | else if(argc == 2) |
| 107 | { |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 108 | std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [npy_input] [labels] [fast_math_hint]\n\n"; |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 109 | std::cout << "No input provided: using random values\n\n"; |
| 110 | } |
Michele Di Giorgio | e3fba0a | 2018-02-14 14:18:01 +0000 | [diff] [blame] | 111 | else if(argc == 4) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 112 | { |
Michele Di Giorgio | e3fba0a | 2018-02-14 14:18:01 +0000 | [diff] [blame] | 113 | data_path = argv[2]; |
| 114 | input = argv[3]; |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 115 | std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n"; |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 116 | std::cout << "No text file with labels provided: skipping output accessor\n\n"; |
| 117 | } |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 118 | else if(argc == 5) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 119 | { |
Michele Di Giorgio | e3fba0a | 2018-02-14 14:18:01 +0000 | [diff] [blame] | 120 | data_path = argv[2]; |
| 121 | input = argv[3]; |
| 122 | label = argv[4]; |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 123 | std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n"; |
| 124 | std::cout << "No fast math info provided: disabling fast math\n\n"; |
| 125 | } |
| 126 | else |
| 127 | { |
| 128 | data_path = argv[2]; |
| 129 | input = argv[3]; |
| 130 | label = argv[4]; |
| 131 | fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED; |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 132 | } |
| 133 | |
Michele Di Giorgio | e3fba0a | 2018-02-14 14:18:01 +0000 | [diff] [blame] | 134 | graph << target_hint |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 135 | << DepthwiseConvolutionMethod::OPTIMIZED_3x3 // FIXME(COMPMID-1073): Add heuristics to automatically call the optimized 3x3 method |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 136 | << fast_math_hint |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 137 | << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::QASYMM8, in_quant_info), |
| 138 | get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/" + input)) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 139 | << ConvolutionLayer( |
| 140 | 3U, 3U, 32U, |
| 141 | get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_weights.npy"), |
| 142 | get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_bias.npy"), |
| 143 | PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 144 | 1, conv_weights_quant_info.at(0), mid_quant_info) |
| 145 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)); |
| 146 | graph << get_dwsc_node(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), point_weights_quant_info.at(0)); |
| 147 | graph << get_dwsc_node(data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1), |
| 148 | point_weights_quant_info.at(1)); |
| 149 | graph << get_dwsc_node(data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2), |
| 150 | point_weights_quant_info.at(2)); |
| 151 | graph << get_dwsc_node(data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3), |
| 152 | point_weights_quant_info.at(3)); |
| 153 | graph << get_dwsc_node(data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4), |
| 154 | point_weights_quant_info.at(4)); |
| 155 | graph << get_dwsc_node(data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5), |
| 156 | point_weights_quant_info.at(5)); |
| 157 | graph << get_dwsc_node(data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6), |
| 158 | point_weights_quant_info.at(6)); |
| 159 | graph << get_dwsc_node(data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7), |
| 160 | point_weights_quant_info.at(7)); |
| 161 | graph << get_dwsc_node(data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8), |
| 162 | point_weights_quant_info.at(8)); |
| 163 | graph << get_dwsc_node(data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9), |
| 164 | point_weights_quant_info.at(9)); |
| 165 | graph << get_dwsc_node(data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10), |
| 166 | point_weights_quant_info.at(10)); |
| 167 | graph << get_dwsc_node(data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11), |
| 168 | point_weights_quant_info.at(11)); |
| 169 | graph << get_dwsc_node(data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12), |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 170 | point_weights_quant_info.at(12)) |
| 171 | << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)) |
| 172 | << ConvolutionLayer( |
| 173 | 1U, 1U, 1001U, |
| 174 | get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_weights.npy"), |
| 175 | get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_bias.npy"), |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 176 | PadStrideInfo(1U, 1U, 0U, 0U), 1, conv_weights_quant_info.at(1)) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 177 | << ReshapeLayer(TensorShape(1001U)) |
| 178 | << SoftmaxLayer() |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 179 | << OutputLayer(get_output_accessor(label, 5)); |
Gian Marco | c1b6e37 | 2018-02-21 18:03:26 +0000 | [diff] [blame] | 180 | |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 181 | // Finalize graph |
| 182 | GraphConfig config; |
Georgios Pinitas | 3d1489d | 2018-05-03 20:47:16 +0100 | [diff] [blame] | 183 | config.use_tuner = (target == 2); |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 184 | graph.finalize(target_hint, config); |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 185 | } |
| 186 | void do_run() override |
| 187 | { |
| 188 | // Run graph |
| 189 | graph.run(); |
| 190 | } |
| 191 | |
| 192 | private: |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 193 | Stream graph{ 0, "MobileNetV1_QASYMM8" }; |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 194 | |
| 195 | /** This function produces a depthwise separable convolution node (i.e. depthwise + pointwise layers) with ReLU6 activation after each layer. |
| 196 | * |
| 197 | * @param[in] data_path Path to trainable data folder |
| 198 | * @param[in] param_path Prefix of specific set of weights/biases data |
| 199 | * @param[in] conv_filt Filters depths for pointwise convolution |
| 200 | * @param[in] dwc_pad_stride_info PadStrideInfo for depthwise convolution |
| 201 | * @param[in] conv_pad_stride_info PadStrideInfo for pointwise convolution |
| 202 | * @param[in] depth_weights_quant_info QuantizationInfo for depthwise convolution's weights |
| 203 | * @param[in] point_weights_quant_info QuantizationInfo for pointwise convolution's weights |
| 204 | * |
| 205 | * @return The complete dwsc node |
| 206 | */ |
| 207 | BranchLayer get_dwsc_node(const std::string &data_path, std::string &¶m_path, |
| 208 | const unsigned int conv_filt, |
| 209 | PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info, |
| 210 | QuantizationInfo depth_weights_quant_info, QuantizationInfo point_weights_quant_info) |
| 211 | { |
| 212 | std::string total_path = "/cnn_data/mobilenet_qasymm8_model/" + param_path + "_"; |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 213 | SubStream sg(graph); |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 214 | |
| 215 | sg << DepthwiseConvolutionLayer( |
| 216 | 3U, 3U, |
| 217 | get_weights_accessor(data_path, total_path + "depthwise_weights.npy"), |
| 218 | get_weights_accessor(data_path, total_path + "depthwise_bias.npy"), |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 219 | dwc_pad_stride_info, depth_weights_quant_info) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 220 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) |
| 221 | << ConvolutionLayer( |
| 222 | 1U, 1U, conv_filt, |
| 223 | get_weights_accessor(data_path, total_path + "pointwise_weights.npy"), |
| 224 | get_weights_accessor(data_path, total_path + "pointwise_bias.npy"), |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 225 | conv_pad_stride_info, 1, point_weights_quant_info) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 226 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)); |
| 227 | |
| 228 | return BranchLayer(std::move(sg)); |
| 229 | } |
| 230 | }; |
| 231 | /** Main program for MobileNetQASYMM8 |
| 232 | * |
| 233 | * @param[in] argc Number of arguments |
Giorgio Arena | 59631a1 | 2018-05-02 13:59:04 +0100 | [diff] [blame] | 234 | * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] npy_input, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) ) |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 235 | */ |
| 236 | int main(int argc, char **argv) |
| 237 | { |
Georgios Pinitas | d9eb275 | 2018-04-03 13:44:29 +0100 | [diff] [blame] | 238 | return arm_compute::utils::run_example<GraphMobileNetQASYMM8Example>(argc, argv); |
Giorgio Arena | a66eaa2 | 2017-12-21 19:50:06 +0000 | [diff] [blame] | 239 | } |