xref: /aosp_15_r20/external/ComputeLibrary/examples/graph_mobilenet_v2.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1  /*
2   * Copyright (c) 2018-2021 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   */
24  #include "arm_compute/graph.h"
25  #include "support/ToolchainSupport.h"
26  #include "utils/CommonGraphOptions.h"
27  #include "utils/GraphUtils.h"
28  #include "utils/Utils.h"
29  
30  using namespace arm_compute;
31  using namespace arm_compute::utils;
32  using namespace arm_compute::graph::frontend;
33  using namespace arm_compute::graph_utils;
34  
35  /** Example demonstrating how to implement MobileNetV2's network using the Compute Library's graph API */
36  class GraphMobilenetV2Example : public Example
37  {
38  public:
GraphMobilenetV2Example()39      GraphMobilenetV2Example()
40          : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetV2")
41      {
42      }
43      GraphMobilenetV2Example(const GraphMobilenetV2Example &) = delete;
44      GraphMobilenetV2Example &operator=(const GraphMobilenetV2Example &) = delete;
45      ~GraphMobilenetV2Example() override                                 = default;
46  
do_setup(int argc,char ** argv)47      bool do_setup(int argc, char **argv) override
48      {
49          // Parse arguments
50          cmd_parser.parse(argc, argv);
51          cmd_parser.validate();
52  
53          // Consume common parameters
54          common_params = consume_common_graph_parameters(common_opts);
55  
56          // Return when help menu is requested
57          if(common_params.help)
58          {
59              cmd_parser.print_help(argv[0]);
60              return false;
61          }
62  
63          // Print parameter values
64          std::cout << common_params << std::endl;
65  
66          // Create input descriptor
67          const TensorShape tensor_shape     = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, common_params.data_layout);
68          TensorDescriptor  input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout);
69  
70          // Set graph hints
71          graph << common_params.target
72                << common_params.fast_math_hint;
73  
74          // Create core graph
75          if(arm_compute::is_data_type_float(common_params.data_type))
76          {
77              create_graph_float(input_descriptor);
78          }
79          else
80          {
81              create_graph_qasymm8(input_descriptor);
82          }
83          // Create common tail
84          graph << ReshapeLayer(TensorShape(1001U)).set_name("Predictions/Reshape")
85                << SoftmaxLayer().set_name("Predictions/Softmax")
86                << OutputLayer(get_output_accessor(common_params, 5));
87  
88          // Finalize graph
89          GraphConfig config;
90          config.num_threads = common_params.threads;
91          config.use_tuner   = common_params.enable_tuner;
92          config.tuner_mode  = common_params.tuner_mode;
93          config.tuner_file  = common_params.tuner_file;
94          config.mlgo_file   = common_params.mlgo_file;
95  
96          graph.finalize(common_params.target, config);
97  
98          return true;
99      }
100  
do_run()101      void do_run() override
102      {
103          // Run graph
104          graph.run();
105      }
106  
107  private:
108      CommandLineParser  cmd_parser;
109      CommonGraphOptions common_opts;
110      CommonGraphParams  common_params;
111      Stream             graph;
112  
113  private:
114      enum class IsResidual
115      {
116          Yes,
117          No
118      };
119  
120      enum class HasExpand
121      {
122          Yes,
123          No
124      };
125  
126  private:
create_graph_float(TensorDescriptor & input_descriptor)127      void create_graph_float(TensorDescriptor &input_descriptor)
128      {
129          // Create model path
130          const std::string model_path = "/cnn_data/mobilenet_v2_1.0_224_model/";
131  
132          // Create a preprocessor object
133          std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>();
134  
135          // Get trainable parameters data path
136          std::string data_path = common_params.data_path;
137  
138          // Add model path to data path
139          if(!data_path.empty())
140          {
141              data_path += model_path;
142          }
143  
144          graph << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false))
145                << ConvolutionLayer(3U, 3U, 32U,
146                                    get_weights_accessor(data_path, "Conv_weights.npy", DataLayout::NCHW),
147                                    std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
148                                    PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL))
149                .set_name("Conv")
150                << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv_BatchNorm_moving_mean.npy"),
151                                           get_weights_accessor(data_path, "Conv_BatchNorm_moving_variance.npy"),
152                                           get_weights_accessor(data_path, "Conv_BatchNorm_gamma.npy"),
153                                           get_weights_accessor(data_path, "Conv_BatchNorm_beta.npy"),
154                                           0.0010000000474974513f)
155                .set_name("Conv/BatchNorm")
156                << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
157                .set_name("Conv/Relu6");
158  
159          get_expanded_conv_float(data_path, "expanded_conv", 32U, 16U, PadStrideInfo(1, 1, 1, 1));
160          get_expanded_conv_float(data_path, "expanded_conv_1", 16U, 24U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes);
161          get_expanded_conv_float(data_path, "expanded_conv_2", 24U, 24U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
162          get_expanded_conv_float(data_path, "expanded_conv_3", 24U, 32U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes);
163          get_expanded_conv_float(data_path, "expanded_conv_4", 32U, 32U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
164          get_expanded_conv_float(data_path, "expanded_conv_5", 32U, 32U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
165          get_expanded_conv_float(data_path, "expanded_conv_6", 32U, 64U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes);
166          get_expanded_conv_float(data_path, "expanded_conv_7", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
167          get_expanded_conv_float(data_path, "expanded_conv_8", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
168          get_expanded_conv_float(data_path, "expanded_conv_9", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
169          get_expanded_conv_float(data_path, "expanded_conv_10", 64U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes);
170          get_expanded_conv_float(data_path, "expanded_conv_11", 96U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
171          get_expanded_conv_float(data_path, "expanded_conv_12", 96U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
172          get_expanded_conv_float(data_path, "expanded_conv_13", 96U, 160U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes);
173          get_expanded_conv_float(data_path, "expanded_conv_14", 160U, 160U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
174          get_expanded_conv_float(data_path, "expanded_conv_15", 160U, 160U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes);
175          get_expanded_conv_float(data_path, "expanded_conv_16", 160U, 320U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes);
176  
177          graph << ConvolutionLayer(1U, 1U, 1280U,
178                                    get_weights_accessor(data_path, "Conv_1_weights.npy", DataLayout::NCHW),
179                                    std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
180                                    PadStrideInfo(1, 1, 0, 0))
181                .set_name("Conv_1")
182                << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv_1_BatchNorm_moving_mean.npy"),
183                                           get_weights_accessor(data_path, "Conv_1_BatchNorm_moving_variance.npy"),
184                                           get_weights_accessor(data_path, "Conv_1_BatchNorm_gamma.npy"),
185                                           get_weights_accessor(data_path, "Conv_1_BatchNorm_beta.npy"),
186                                           0.0010000000474974513f)
187                .set_name("Conv_1/BatchNorm")
188                << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
189                .set_name("Conv_1/Relu6")
190                << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool")
191                << ConvolutionLayer(1U, 1U, 1001U,
192                                    get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW),
193                                    get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
194                                    PadStrideInfo(1, 1, 0, 0))
195                .set_name("Logits/Conv2d_1c_1x1");
196      }
197  
get_expanded_conv_float(const std::string & data_path,std::string && param_path,unsigned int input_channels,unsigned int output_channels,PadStrideInfo dwc_pad_stride_info,HasExpand has_expand=HasExpand::No,IsResidual is_residual=IsResidual::No,unsigned int expansion_size=6)198      void get_expanded_conv_float(const std::string &data_path, std::string &&param_path,
199                                   unsigned int input_channels, unsigned int output_channels,
200                                   PadStrideInfo dwc_pad_stride_info,
201                                   HasExpand has_expand = HasExpand::No, IsResidual is_residual = IsResidual::No,
202                                   unsigned int expansion_size = 6)
203      {
204          std::string total_path = param_path + "_";
205          SubStream   left(graph);
206  
207          // Add expand node
208          if(has_expand == HasExpand::Yes)
209          {
210              left << ConvolutionLayer(1U, 1U, input_channels * expansion_size,
211                                       get_weights_accessor(data_path, total_path + "expand_weights.npy", DataLayout::NCHW),
212                                       std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
213                   .set_name(param_path + "/expand/Conv2D")
214                   << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "expand_BatchNorm_moving_mean.npy"),
215                                              get_weights_accessor(data_path, total_path + "expand_BatchNorm_moving_variance.npy"),
216                                              get_weights_accessor(data_path, total_path + "expand_BatchNorm_gamma.npy"),
217                                              get_weights_accessor(data_path, total_path + "expand_BatchNorm_beta.npy"),
218                                              0.0010000000474974513f)
219                   .set_name(param_path + "/expand/BatchNorm")
220                   << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
221                   .set_name(param_path + "/expand/Relu6");
222          }
223  
224          // Add depthwise node
225          left << DepthwiseConvolutionLayer(3U, 3U,
226                                            get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW),
227                                            std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
228                                            dwc_pad_stride_info)
229               .set_name(param_path + "/depthwise/depthwise")
230               << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"),
231                                          get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"),
232                                          get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"),
233                                          get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"),
234                                          0.0010000000474974513f)
235               .set_name(param_path + "/depthwise/BatchNorm")
236               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
237               .set_name(param_path + "/depthwise/Relu6");
238  
239          // Add project node
240          left << ConvolutionLayer(1U, 1U, output_channels,
241                                   get_weights_accessor(data_path, total_path + "project_weights.npy", DataLayout::NCHW),
242                                   std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
243               .set_name(param_path + "/project/Conv2D")
244               << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "project_BatchNorm_moving_mean.npy"),
245                                          get_weights_accessor(data_path, total_path + "project_BatchNorm_moving_variance.npy"),
246                                          get_weights_accessor(data_path, total_path + "project_BatchNorm_gamma.npy"),
247                                          get_weights_accessor(data_path, total_path + "project_BatchNorm_beta.npy"),
248                                          0.0010000000474974513)
249               .set_name(param_path + "/project/BatchNorm");
250  
251          if(is_residual == IsResidual::Yes)
252          {
253              // Add residual node
254              SubStream right(graph);
255              graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(param_path + "/add");
256          }
257          else
258          {
259              graph.forward_tail(left.tail_node());
260          }
261      }
262  
create_graph_qasymm8(TensorDescriptor & input_descriptor)263      void create_graph_qasymm8(TensorDescriptor &input_descriptor)
264      {
265          // Create model path
266          const std::string model_path = "/cnn_data/mobilenet_v2_1.0_224_quantized_model/";
267  
268          // Get trainable parameters data path
269          std::string data_path = common_params.data_path;
270  
271          // Add model path to data path
272          if(!data_path.empty())
273          {
274              data_path += model_path;
275          }
276  
277          const QuantizationInfo in_quant_info  = QuantizationInfo(0.0078125f, 128);
278          const QuantizationInfo mid_quant_info = QuantizationInfo(0.023528477177023888f, 128);
279  
280          const std::vector<QuantizationInfo> conv_weights_quant_info =
281          {
282              QuantizationInfo(0.03396892547607422f, 122),  // Conv
283              QuantizationInfo(0.005167067516595125f, 125), // Conv1
284              QuantizationInfo(0.0016910821432247758f, 113) // Conv2d_1c_1x1
285          };
286  
287          // Pointwise expand convolution quantization info
288          const std::vector<QuantizationInfo> pwc_q =
289          {
290              QuantizationInfo(0.254282623529f, 129),        // expand_0 (Dummy)
291              QuantizationInfo(0.009758507832884789f, 127),  // expand_1
292              QuantizationInfo(0.0036556976847350597f, 144), // expand_2
293              QuantizationInfo(0.0029988749884068966f, 104), // expand_3
294              QuantizationInfo(0.0019244228024035692f, 128), // expand_4
295              QuantizationInfo(0.0013649158645421267f, 135), // expand_5
296              QuantizationInfo(0.0019170437008142471f, 127), // expand_6
297              QuantizationInfo(0.0015538912266492844f, 125), // expand_7
298              QuantizationInfo(0.0014702979242429137f, 134), // expand_8
299              QuantizationInfo(0.0013733493397012353f, 127), // expand_9
300              QuantizationInfo(0.0016282502328976989f, 131), // expand_10
301              QuantizationInfo(0.0016309921629726887f, 134), // expand_11
302              QuantizationInfo(0.0018258779309689999f, 138), // expand_12
303              QuantizationInfo(0.0013828007504343987f, 123), // expand_13
304              QuantizationInfo(0.0020222084131091833f, 135), // expand_14
305              QuantizationInfo(0.04281935095787048f, 102),   // expand_15
306              QuantizationInfo(0.002046825597062707f, 135)   // expand_16
307          };
308          // Depthwise expand convolution quantization info
309          const std::vector<QuantizationInfo> dwc_q =
310          {
311              QuantizationInfo(0.3436955213546753f, 165),   // expand_0
312              QuantizationInfo(0.020969120785593987f, 109), // expand_1
313              QuantizationInfo(0.16981913149356842f, 52),   // expand_2
314              QuantizationInfo(0.017202870920300484f, 143), // expand_3
315              QuantizationInfo(0.06525065749883652f, 118),  // expand_4
316              QuantizationInfo(0.07909784466028214f, 95),   // expand_5
317              QuantizationInfo(0.010087885893881321f, 127), // expand_6
318              QuantizationInfo(0.06092711538076401f, 110),  // expand_7
319              QuantizationInfo(0.052407849580049515f, 133), // expand_8
320              QuantizationInfo(0.04077887907624245f, 155),  // expand_9
321              QuantizationInfo(0.031107846647500992f, 143), // expand_10
322              QuantizationInfo(0.07080810517072678f, 66),   // expand_11
323              QuantizationInfo(0.07448793947696686f, 159),  // expand_12
324              QuantizationInfo(0.01525793131440878f, 92),   // expand_13
325              QuantizationInfo(0.04166752099990845f, 147),  // expand_14
326              QuantizationInfo(0.04281935095787048f, 102),  // expand_15
327              QuantizationInfo(0.16456253826618195, 201)    // expand_16
328          };
329          // Project convolution quantization info
330          const std::vector<QuantizationInfo> prwc_q =
331          {
332              QuantizationInfo(0.03737175464630127f, 140),  // expand_0
333              QuantizationInfo(0.0225360207259655f, 156),   // expand_1
334              QuantizationInfo(0.02740888111293316f, 122),  // expand_2
335              QuantizationInfo(0.016844693571329117f, 111), // expand_3
336              QuantizationInfo(0.019062912091612816f, 146), // expand_4
337              QuantizationInfo(0.018293123692274094f, 128), // expand_5
338              QuantizationInfo(0.014601286500692368f, 147), // expand_6
339              QuantizationInfo(0.016782939434051514f, 124), // expand_7
340              QuantizationInfo(0.012898261658847332f, 125), // expand_8
341              QuantizationInfo(0.019561484456062317f, 144), // expand_9
342              QuantizationInfo(0.007436311338096857f, 129), // expand_10
343              QuantizationInfo(0.00838223285973072f, 136),  // expand_11
344              QuantizationInfo(0.023982593789696693f, 154), // expand_12
345              QuantizationInfo(0.009447949007153511f, 140), // expand_13
346              QuantizationInfo(0.00789870135486126f, 139),  // expand_14
347              QuantizationInfo(0.03697410225868225f, 131),  // expand_15
348              QuantizationInfo(0.008009289391338825f, 111)  // expand_16
349          };
350  
351          graph << InputLayer(input_descriptor.set_quantization_info(in_quant_info),
352                              get_weights_accessor(data_path, common_params.image))
353                << ConvolutionLayer(
354                    3U, 3U, 32U,
355                    get_weights_accessor(data_path, "Conv_weights.npy"),
356                    get_weights_accessor(data_path, "Conv_bias.npy"),
357                    PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR),
358                    1, conv_weights_quant_info.at(0), mid_quant_info)
359                .set_name("Conv")
360                << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("Conv/Relu6")
361                << DepthwiseConvolutionLayer(3U, 3U,
362                                             get_weights_accessor(data_path, "expanded_conv_depthwise_depthwise_weights.npy"),
363                                             get_weights_accessor(data_path, "expanded_conv_depthwise_depthwise_biases.npy"),
364                                             PadStrideInfo(1, 1, 1, 1), 1, dwc_q.at(0))
365                .set_name("expanded_conv/depthwise/depthwise")
366                << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("expanded_conv/depthwise/Relu6")
367                << ConvolutionLayer(1U, 1U, 16U,
368                                    get_weights_accessor(data_path, "expanded_conv_project_weights.npy"),
369                                    get_weights_accessor(data_path, "expanded_conv_project_biases.npy"),
370                                    PadStrideInfo(1, 1, 0, 0), 1, prwc_q.at(0))
371                .set_name("expanded_conv/project/Conv2D");
372  
373          get_expanded_conv_qasymm8(data_path, "expanded_conv_1", IsResidual::No, 96U, 24U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL),
374                                    pwc_q.at(1), dwc_q.at(1), prwc_q.at(1));
375          get_expanded_conv_qasymm8(data_path, "expanded_conv_2", IsResidual::Yes, 144U, 24U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(2), dwc_q.at(2), prwc_q.at(2));
376          get_expanded_conv_qasymm8(data_path, "expanded_conv_3", IsResidual::No, 144U, 32U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL),
377                                    pwc_q.at(3), dwc_q.at(3), prwc_q.at(3));
378          get_expanded_conv_qasymm8(data_path, "expanded_conv_4", IsResidual::Yes, 192U, 32U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(4), dwc_q.at(4), prwc_q.at(4));
379          get_expanded_conv_qasymm8(data_path, "expanded_conv_5", IsResidual::Yes, 192U, 32U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(5), dwc_q.at(5), prwc_q.at(5));
380          get_expanded_conv_qasymm8(data_path, "expanded_conv_6", IsResidual::No, 192U, 64U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL),
381                                    pwc_q.at(6), dwc_q.at(6), prwc_q.at(6));
382          get_expanded_conv_qasymm8(data_path, "expanded_conv_7", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(7), dwc_q.at(7), prwc_q.at(7));
383          get_expanded_conv_qasymm8(data_path, "expanded_conv_8", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(8), dwc_q.at(8), prwc_q.at(8));
384          get_expanded_conv_qasymm8(data_path, "expanded_conv_9", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(9), dwc_q.at(9), prwc_q.at(9));
385          get_expanded_conv_qasymm8(data_path, "expanded_conv_10", IsResidual::No, 384U, 96U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(10), dwc_q.at(10), prwc_q.at(10));
386          get_expanded_conv_qasymm8(data_path, "expanded_conv_11", IsResidual::Yes, 576U, 96U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(11), dwc_q.at(11), prwc_q.at(11));
387          get_expanded_conv_qasymm8(data_path, "expanded_conv_12", IsResidual::Yes, 576U, 96U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(12), dwc_q.at(12), prwc_q.at(12));
388          get_expanded_conv_qasymm8(data_path, "expanded_conv_13", IsResidual::No, 576U, 160U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL),
389                                    pwc_q.at(13), dwc_q.at(13), prwc_q.at(13));
390          get_expanded_conv_qasymm8(data_path, "expanded_conv_14", IsResidual::Yes, 960U, 160U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(14), dwc_q.at(14), prwc_q.at(14));
391          get_expanded_conv_qasymm8(data_path, "expanded_conv_15", IsResidual::Yes, 960U, 160U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(15), dwc_q.at(15), prwc_q.at(15));
392          get_expanded_conv_qasymm8(data_path, "expanded_conv_16", IsResidual::No, 960U, 320U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(16), dwc_q.at(16), prwc_q.at(16));
393  
394          graph << ConvolutionLayer(1U, 1U, 1280U,
395                                    get_weights_accessor(data_path, "Conv_1_weights.npy"),
396                                    get_weights_accessor(data_path, "Conv_1_biases.npy"),
397                                    PadStrideInfo(1, 1, 0, 0), 1, conv_weights_quant_info.at(1))
398                .set_name("Conv_1")
399                << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("Conv_1/Relu6")
400                << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool")
401                << ConvolutionLayer(1U, 1U, 1001U,
402                                    get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"),
403                                    get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
404                                    PadStrideInfo(1, 1, 0, 0), 1, conv_weights_quant_info.at(2))
405                .set_name("Logits/Conv2d_1c_1x1");
406      }
407  
get_expanded_conv_qasymm8(const std::string & data_path,std::string && param_path,IsResidual is_residual,unsigned int input_channels,unsigned int output_channels,PadStrideInfo dwc_pad_stride_info,const QuantizationInfo & pwi,const QuantizationInfo & dwi,const QuantizationInfo & pji)408      void get_expanded_conv_qasymm8(const std::string &data_path, std::string &&param_path, IsResidual is_residual,
409                                     unsigned int input_channels, unsigned int output_channels,
410                                     PadStrideInfo           dwc_pad_stride_info,
411                                     const QuantizationInfo &pwi, const QuantizationInfo &dwi, const QuantizationInfo &pji)
412      {
413          std::string total_path = param_path + "_";
414  
415          SubStream left(graph);
416          left << ConvolutionLayer(1U, 1U, input_channels,
417                                   get_weights_accessor(data_path, total_path + "project_weights.npy"),
418                                   get_weights_accessor(data_path, total_path + "project_biases.npy"),
419                                   PadStrideInfo(1, 1, 0, 0), 1, pwi)
420               .set_name(param_path + "/Conv2D")
421               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name(param_path + "/Conv2D/Relu6")
422               << DepthwiseConvolutionLayer(3U, 3U,
423                                            get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy"),
424                                            get_weights_accessor(data_path, total_path + "depthwise_depthwise_biases.npy"),
425                                            dwc_pad_stride_info, 1, dwi)
426               .set_name(param_path + "/depthwise/depthwise")
427               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name(param_path + "/depthwise/Relu6")
428               << ConvolutionLayer(1U, 1U, output_channels,
429                                   get_weights_accessor(data_path, total_path + "project_weights.npy"),
430                                   get_weights_accessor(data_path, total_path + "project_biases.npy"),
431                                   PadStrideInfo(1, 1, 0, 0), 1, pji)
432               .set_name(param_path + "/project/Conv2D");
433  
434          if(is_residual == IsResidual::Yes)
435          {
436              // Add residual node
437              SubStream right(graph);
438              graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(param_path + "/add");
439          }
440          else
441          {
442              graph.forward_tail(left.tail_node());
443          }
444      }
445  };
446  
447  /** Main program for MobileNetV2
448   *
449   * Model is based on:
450   *      https://arxiv.org/abs/1801.04381
451   *      "MobileNetV2: Inverted Residuals and Linear Bottlenecks"
452   *      Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen
453   *
454   * Provenance: https://storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_1.0_224.tgz
455   *
456   * @note To list all the possible arguments execute the binary appended with the --help option
457   *
458   * @param[in] argc Number of arguments
459   * @param[in] argv Arguments
460   */
main(int argc,char ** argv)461  int main(int argc, char **argv)
462  {
463      return arm_compute::utils::run_example<GraphMobilenetV2Example>(argc, argv);
464  }
465