examples/neon_cnn.cpp

   1 /*
   2  * Copyright (c) 2016, 2017 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/runtime/NEON/NEFunctions.h"
  25
  26 #include "arm_compute/core/Types.h"
  27 #include "utils/Utils.h"
  28
  29 using namespace arm_compute;
  30 using namespace utils;
  31
  32 void main_cnn(int argc, const char **argv)
  33 {
  34     ARM_COMPUTE_UNUSED(argc);
  35     ARM_COMPUTE_UNUSED(argv);
  36
  37     // The src tensor should contain the input image
  38     Tensor src;
  39
  40     // The weights and biases tensors should be initialized with the values inferred with the training
  41     Tensor weights0;
  42     Tensor weights1;
  43     Tensor weights2;
  44     Tensor biases0;
  45     Tensor biases1;
  46     Tensor biases2;
  47
  48     Tensor out_conv0;
  49     Tensor out_conv1;
  50     Tensor out_act0;
  51     Tensor out_act1;
  52     Tensor out_act2;
  53     Tensor out_pool0;
  54     Tensor out_pool1;
  55     Tensor out_fc0;
  56     Tensor out_softmax;
  57
  58     NEConvolutionLayer    conv0;
  59     NEConvolutionLayer    conv1;
  60     NEPoolingLayer        pool0;
  61     NEPoolingLayer        pool1;
  62     NEFullyConnectedLayer fc0;
  63     NEActivationLayer     act0;
  64     NEActivationLayer     act1;
  65     NEActivationLayer     act2;
  66     NESoftmaxLayer        softmax;
  67
  68     /* [Initialize tensors] */
  69
  70     // Initialize src tensor
  71     constexpr unsigned int width_src_image  = 32;
  72     constexpr unsigned int height_src_image = 32;
  73     constexpr unsigned int ifm_src_img      = 1;
  74
  75     const TensorShape src_shape(width_src_image, height_src_image, ifm_src_img);
  76     src.allocator()->init(TensorInfo(src_shape, 1, DataType::F32));
  77
  78     // Initialize tensors of conv0
  79     constexpr unsigned int kernel_x_conv0 = 5;
  80     constexpr unsigned int kernel_y_conv0 = 5;
  81     constexpr unsigned int ofm_conv0      = 8;
  82
  83     const TensorShape weights_shape_conv0(kernel_x_conv0, kernel_y_conv0, src_shape.z(), ofm_conv0);
  84     const TensorShape biases_shape_conv0(weights_shape_conv0[3]);
  85     const TensorShape out_shape_conv0(src_shape.x(), src_shape.y(), weights_shape_conv0[3]);
  86
  87     weights0.allocator()->init(TensorInfo(weights_shape_conv0, 1, DataType::F32));
  88     biases0.allocator()->init(TensorInfo(biases_shape_conv0, 1, DataType::F32));
  89     out_conv0.allocator()->init(TensorInfo(out_shape_conv0, 1, DataType::F32));
  90
  91     // Initialize tensor of act0
  92     out_act0.allocator()->init(TensorInfo(out_shape_conv0, 1, DataType::F32));
  93
  94     // Initialize tensor of pool0
  95     TensorShape out_shape_pool0 = out_shape_conv0;
  96     out_shape_pool0.set(0, out_shape_pool0.x() / 2);
  97     out_shape_pool0.set(1, out_shape_pool0.y() / 2);
  98     out_pool0.allocator()->init(TensorInfo(out_shape_pool0, 1, DataType::F32));
  99
 100     // Initialize tensors of conv1
 101     constexpr unsigned int kernel_x_conv1 = 3;
 102     constexpr unsigned int kernel_y_conv1 = 3;
 103     constexpr unsigned int ofm_conv1      = 16;
 104
 105     const TensorShape weights_shape_conv1(kernel_x_conv1, kernel_y_conv1, out_shape_pool0.z(), ofm_conv1);
 106
 107     const TensorShape biases_shape_conv1(weights_shape_conv1[3]);
 108     const TensorShape out_shape_conv1(out_shape_pool0.x(), out_shape_pool0.y(), weights_shape_conv1[3]);
 109
 110     weights1.allocator()->init(TensorInfo(weights_shape_conv1, 1, DataType::F32));
 111     biases1.allocator()->init(TensorInfo(biases_shape_conv1, 1, DataType::F32));
 112     out_conv1.allocator()->init(TensorInfo(out_shape_conv1, 1, DataType::F32));
 113
 114     // Initialize tensor of act1
 115     out_act1.allocator()->init(TensorInfo(out_shape_conv1, 1, DataType::F32));
 116
 117     // Initialize tensor of pool1
 118     TensorShape out_shape_pool1 = out_shape_conv1;
 119     out_shape_pool1.set(0, out_shape_pool1.x() / 2);
 120     out_shape_pool1.set(1, out_shape_pool1.y() / 2);
 121     out_pool1.allocator()->init(TensorInfo(out_shape_pool1, 1, DataType::F32));
 122
 123     // Initialize tensor of fc0
 124     constexpr unsigned int num_labels = 128;
 125
 126     const TensorShape weights_shape_fc0(out_shape_pool1.x() * out_shape_pool1.y() * out_shape_pool1.z(), num_labels);
 127     const TensorShape biases_shape_fc0(num_labels);
 128     const TensorShape out_shape_fc0(num_labels);
 129
 130     weights2.allocator()->init(TensorInfo(weights_shape_fc0, 1, DataType::F32));
 131     biases2.allocator()->init(TensorInfo(biases_shape_fc0, 1, DataType::F32));
 132     out_fc0.allocator()->init(TensorInfo(out_shape_fc0, 1, DataType::F32));
 133
 134     // Initialize tensor of act2
 135     out_act2.allocator()->init(TensorInfo(out_shape_fc0, 1, DataType::F32));
 136
 137     // Initialize tensor of softmax
 138     const TensorShape out_shape_softmax(out_shape_fc0.x());
 139     out_softmax.allocator()->init(TensorInfo(out_shape_softmax, 1, DataType::F32));
 140
 141     /* -----------------------End: [Initialize tensors] */
 142
 143     /* [Configure functions] */
 144
 145     // in:32x32x1: 5x5 convolution, 8 output features maps (OFM)
 146     conv0.configure(&src, &weights0, &biases0, &out_conv0, PadStrideInfo());
 147
 148     // in:32x32x8, out:32x32x8, Activation function: relu
 149     act0.configure(&out_conv0, &out_act0, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 150
 151     // in:32x32x8, out:16x16x8 (2x2 pooling), Pool type function: Max
 152     pool0.configure(&out_act0, &out_pool0, PoolingLayerInfo(PoolingType::MAX, 2));
 153
 154     // in:16x16x8: 3x3 convolution, 16 output features maps (OFM)
 155     conv1.configure(&out_pool0, &weights1, &biases1, &out_conv1, PadStrideInfo());
 156
 157     // in:16x16x16, out:16x16x16, Activation function: relu
 158     act1.configure(&out_conv1, &out_act1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 159
 160     // in:16x16x16, out:8x8x16 (2x2 pooling), Pool type function: Average
 161     pool1.configure(&out_act1, &out_pool1, PoolingLayerInfo(PoolingType::AVG, 2));
 162
 163     // in:8x8x16, out:128
 164     fc0.configure(&out_pool1, &weights2, &biases2, &out_fc0);
 165
 166     // in:128, out:128, Activation function: relu
 167     act2.configure(&out_fc0, &out_act2, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 168
 169     // in:128, out:128
 170     softmax.configure(&out_act2, &out_softmax);
 171
 172     /* -----------------------End: [Configure functions] */
 173
 174     /* [Allocate tensors] */
 175
 176     // Now that the padding requirements are known we can allocate the images:
 177     src.allocator()->allocate();
 178     weights0.allocator()->allocate();
 179     weights1.allocator()->allocate();
 180     weights2.allocator()->allocate();
 181     biases0.allocator()->allocate();
 182     biases1.allocator()->allocate();
 183     biases2.allocator()->allocate();
 184     out_conv0.allocator()->allocate();
 185     out_conv1.allocator()->allocate();
 186     out_act0.allocator()->allocate();
 187     out_act1.allocator()->allocate();
 188     out_act2.allocator()->allocate();
 189     out_pool0.allocator()->allocate();
 190     out_pool1.allocator()->allocate();
 191     out_fc0.allocator()->allocate();
 192     out_softmax.allocator()->allocate();
 193
 194     /* -----------------------End: [Allocate tensors] */
 195
 196     /* [Initialize weights and biases tensors] */
 197
 198     // Once the tensors have been allocated, the src, weights and biases tensors can be initialized
 199     // ...
 200
 201     /* -----------------------[Initialize weights and biases tensors] */
 202
 203     /* [Execute the functions] */
 204
 205     conv0.run();
 206     act0.run();
 207     pool0.run();
 208     conv1.run();
 209     act1.run();
 210     pool1.run();
 211     fc0.run();
 212     act2.run();
 213     softmax.run();
 214
 215     /* -----------------------End: [Execute the functions] */
 216 }
 217
 218 /** Main program for cnn test
 219  *
 220  * The example implements the following CNN architecture:
 221  *
 222  * Input -> conv0:5x5 -> act0:relu -> pool:2x2 -> conv1:3x3 -> act1:relu -> pool:2x2 -> fc0 -> act2:relu -> softmax
 223  *
 224  * @param[in] argc Number of arguments
 225  * @param[in] argv Arguments
 226  */
 227 int main(int argc, const char **argv)
 228 {
 229     return utils::run_example(argc, argv, main_cnn);
 230 }