blob: e85108702d051eaae24b62920fc0efb17513b887 [file] [log] [blame]
Georgios Pinitas37561862017-10-19 10:51:03 +01001/*
Gian Marco36a0a462018-01-12 10:21:40 +00002 * Copyright (c) 2017-2018 ARM Limited.
Georgios Pinitas37561862017-10-19 10:51:03 +01003 *
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 Pinitas37561862017-10-19 10:51:03 +010024#include "arm_compute/graph/Graph.h"
25#include "arm_compute/graph/Nodes.h"
26#include "arm_compute/graph/SubGraph.h"
Georgios Pinitas37561862017-10-19 10:51:03 +010027#include "support/ToolchainSupport.h"
28#include "utils/GraphUtils.h"
29#include "utils/Utils.h"
30
31#include <cstdlib>
Georgios Pinitas37561862017-10-19 10:51:03 +010032#include <tuple>
33
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +000034using namespace arm_compute::utils;
Georgios Pinitas37561862017-10-19 10:51:03 +010035using namespace arm_compute::graph;
36using namespace arm_compute::graph_utils;
Isabella Gottardi4398bec2017-10-19 16:10:59 +010037using namespace arm_compute::logging;
Georgios Pinitas37561862017-10-19 10:51:03 +010038
Gian Marcobfa3b522017-12-12 10:08:38 +000039namespace
40{
Gian Marcobfa3b522017-12-12 10:08:38 +000041} // namespace
Georgios Pinitas37561862017-10-19 10:51:03 +010042
43/** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API
44 *
45 * @param[in] argc Number of arguments
Gian Marcobfa3b522017-12-12 10:08:38 +000046 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
Georgios Pinitas37561862017-10-19 10:51:03 +010047 */
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +000048class GraphSqueezenetExample : public Example
Georgios Pinitas37561862017-10-19 10:51:03 +010049{
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +000050public:
51 void do_setup(int argc, char **argv) override
52 {
53 std::string data_path; /* Path to the trainable data */
54 std::string image; /* Image data */
55 std::string label; /* Label data */
Isabella Gottardi97988a42017-11-03 14:39:44 +000056
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +000057 constexpr float mean_r = 122.68f; /* Mean value to subtract from red channel */
58 constexpr float mean_g = 116.67f; /* Mean value to subtract from green channel */
59 constexpr float mean_b = 104.01f; /* Mean value to subtract from blue channel */
Georgios Pinitas37561862017-10-19 10:51:03 +010060
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +000061 // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
Gian Marco36a0a462018-01-12 10:21:40 +000062 TargetHint target_hint = set_target_hint(argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0);
Gian Marcobfa3b522017-12-12 10:08:38 +000063
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +000064 // Parse arguments
65 if(argc < 2)
66 {
67 // Print help
68 std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
69 std::cout << "No data folder provided: using random values\n\n";
70 }
71 else if(argc == 2)
72 {
73 std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
74 std::cout << "No data folder provided: using random values\n\n";
75 }
76 else if(argc == 3)
77 {
78 data_path = argv[2];
79 std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
80 std::cout << "No image provided: using random values\n\n";
81 }
82 else if(argc == 4)
83 {
84 data_path = argv[2];
85 image = argv[3];
86 std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
87 std::cout << "No text file with labels provided: skipping output accessor\n\n";
88 }
89 else
90 {
91 data_path = argv[2];
92 image = argv[3];
93 label = argv[4];
94 }
95
96 graph << target_hint
97 << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
98 get_input_accessor(image, mean_r, mean_g, mean_b))
99 << ConvolutionLayer(
100 7U, 7U, 96U,
101 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"),
102 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"),
103 PadStrideInfo(2, 2, 0, 0))
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000104 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
105 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
106 << ConvolutionLayer(
107 1U, 1U, 16U,
108 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy"),
109 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"),
110 PadStrideInfo(1, 1, 0, 0))
111 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
112 << get_expand_fire_node(data_path, "fire2", 64U, 64U)
113 << ConvolutionLayer(
114 1U, 1U, 16U,
115 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy"),
116 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"),
117 PadStrideInfo(1, 1, 0, 0))
118 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
119 << get_expand_fire_node(data_path, "fire3", 64U, 64U)
120 << ConvolutionLayer(
121 1U, 1U, 32U,
122 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy"),
123 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"),
124 PadStrideInfo(1, 1, 0, 0))
125 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
126 << get_expand_fire_node(data_path, "fire4", 128U, 128U)
127 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
128 << ConvolutionLayer(
129 1U, 1U, 32U,
130 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy"),
131 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"),
132 PadStrideInfo(1, 1, 0, 0))
133 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
134 << get_expand_fire_node(data_path, "fire5", 128U, 128U)
135 << ConvolutionLayer(
136 1U, 1U, 48U,
137 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy"),
138 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"),
139 PadStrideInfo(1, 1, 0, 0))
140 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
141 << get_expand_fire_node(data_path, "fire6", 192U, 192U)
142 << ConvolutionLayer(
143 1U, 1U, 48U,
144 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy"),
145 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"),
146 PadStrideInfo(1, 1, 0, 0))
147 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
148 << get_expand_fire_node(data_path, "fire7", 192U, 192U)
149 << ConvolutionLayer(
150 1U, 1U, 64U,
151 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy"),
152 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"),
153 PadStrideInfo(1, 1, 0, 0))
154 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
155 << get_expand_fire_node(data_path, "fire8", 256U, 256U)
156 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
157 << ConvolutionLayer(
158 1U, 1U, 64U,
159 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy"),
160 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"),
161 PadStrideInfo(1, 1, 0, 0))
162 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
163 << get_expand_fire_node(data_path, "fire9", 256U, 256U)
164 << ConvolutionLayer(
165 1U, 1U, 1000U,
166 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy"),
167 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"),
168 PadStrideInfo(1, 1, 0, 0))
169 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
170 << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
171 << FlattenLayer()
172 << SoftmaxLayer()
173 << Tensor(get_output_accessor(label, 5));
Georgios Pinitas37561862017-10-19 10:51:03 +0100174 }
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000175 void do_run() override
Georgios Pinitas37561862017-10-19 10:51:03 +0100176 {
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000177 // Run graph
178 graph.run();
Georgios Pinitas37561862017-10-19 10:51:03 +0100179 }
180
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000181private:
182 Graph graph{};
Georgios Pinitas37561862017-10-19 10:51:03 +0100183
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000184 BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&param_path, unsigned int expand1_filt, unsigned int expand3_filt)
185 {
186 std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_";
187 SubGraph i_a;
188 i_a << ConvolutionLayer(
189 1U, 1U, expand1_filt,
190 get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
191 get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
192 PadStrideInfo(1, 1, 0, 0))
193 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
Georgios Pinitas37561862017-10-19 10:51:03 +0100194
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000195 SubGraph i_b;
196 i_b << ConvolutionLayer(
197 3U, 3U, expand3_filt,
198 get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
199 get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
200 PadStrideInfo(1, 1, 1, 1))
201 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
202
203 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
204 }
205};
Georgios Pinitas37561862017-10-19 10:51:03 +0100206
207/** Main program for Squeezenet v1.0
208 *
209 * @param[in] argc Number of arguments
Gian Marcobfa3b522017-12-12 10:08:38 +0000210 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
Georgios Pinitas37561862017-10-19 10:51:03 +0100211 */
Anthony Barbier6db0ff52018-01-05 10:59:12 +0000212int main(int argc, char **argv)
Georgios Pinitas37561862017-10-19 10:51:03 +0100213{
Michalis Spyrou2b5f0f22018-01-10 14:08:50 +0000214 return arm_compute::utils::run_example<GraphSqueezenetExample>(argc, argv);
Anthony Barbier6db0ff52018-01-05 10:59:12 +0000215}