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 &¶m_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 &¶m_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