tiny_dnn  1.0.0
A header only, dependency-free deep learning framework in C++11
lrn_layer.h
1 /*
2  Copyright (c) 2013, Taiga Nomi
3  All rights reserved.
4 
5  Redistribution and use in source and binary forms, with or without
6  modification, are permitted provided that the following conditions are met:
7  * Redistributions of source code must retain the above copyright
8  notice, this list of conditions and the following disclaimer.
9  * Redistributions in binary form must reproduce the above copyright
10  notice, this list of conditions and the following disclaimer in the
11  documentation and/or other materials provided with the distribution.
12  * Neither the name of the <organization> nor the
13  names of its contributors may be used to endorse or promote products
14  derived from this software without specific prior written permission.
15 
16  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY
17  EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18  WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19  DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
20  DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21  (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
22  LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
23  ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
25  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 */
27 #pragma once
28 #include "tiny_dnn/util/util.h"
29 #include <algorithm>
30 
31 namespace tiny_dnn {
32 
33  enum class norm_region {
34  across_channels,
35  within_channels
36  };
37 
41 template<typename Activation>
42 class lrn_layer : public feedforward_layer<Activation> {
43 public:
44  CNN_USE_LAYER_MEMBERS;
45 
47 
48  lrn_layer(const shape3d& in_shape,
49  serial_size_t local_size,
50  float_t alpha = 1.0,
51  float_t beta = 5.0,
52  norm_region region = norm_region::across_channels)
53  : Base({ vector_type::data }),
54  in_shape_(in_shape),
55  size_(local_size),
56  alpha_(alpha),
57  beta_(beta),
58  region_(region),
59  in_square_(in_shape_.area()) {
60  }
61 
70  serial_size_t local_size,
71  float_t alpha = 1.0,
72  float_t beta = 5.0,
73  norm_region region = norm_region::across_channels)
74  : lrn_layer(prev->out_data_shape()[0], local_size, alpha, beta, region) {}
75 
84  lrn_layer(serial_size_t in_width,
85  serial_size_t in_height,
86  serial_size_t local_size,
87  serial_size_t in_channels,
88  float_t alpha = 1.0,
89  float_t beta = 5.0,
90  norm_region region = norm_region::across_channels)
91  : lrn_layer(shape3d{in_width, in_height, in_channels}, local_size, alpha, beta, region){}
92 
93  serial_size_t fan_in_size() const override {
94  return size_;
95  }
96 
97  serial_size_t fan_out_size() const override {
98  return size_;
99  }
100 
101  std::vector<shape3d> in_shape() const override {
102  return { in_shape_ };
103  }
104 
105  std::vector<shape3d> out_shape() const override {
106  return { in_shape_, in_shape_ };
107  }
108 
109  std::string layer_type() const override { return "lrn"; }
110 
111  void forward_propagation(const std::vector<tensor_t*>& in_data,
112  std::vector<tensor_t*>& out_data) override {
113 
114  // @todo revise the parallelism strategy
115  for (size_t sample = 0, sample_count = in_data[0]->size(); sample < sample_count; ++sample) {
116  vec_t& in = (*in_data[0])[sample];
117  vec_t& out = (*out_data[0])[sample];
118  vec_t& a = (*out_data[1])[sample];
119 
120  if (region_ == norm_region::across_channels) {
121  forward_across(in, a);
122  }
123  else {
124  forward_within(in, a);
125  }
126 
127  for_i(parallelize_, out.size(), [&](int i) {
128  out[i] = h_.f(a, i);
129  });
130  }
131  }
132 
133  void back_propagation(const std::vector<tensor_t*>& in_data,
134  const std::vector<tensor_t*>& out_data,
135  std::vector<tensor_t*>& out_grad,
136  std::vector<tensor_t*>& in_grad) override {
137  CNN_UNREFERENCED_PARAMETER(in_data);
138  CNN_UNREFERENCED_PARAMETER(out_data);
139  CNN_UNREFERENCED_PARAMETER(out_grad);
140  CNN_UNREFERENCED_PARAMETER(in_grad);
141  throw nn_error("not implemented");
142  }
143 
144  template <class Archive>
145  static void load_and_construct(Archive & ar, cereal::construct<lrn_layer> & construct) {
147  serial_size_t size;
148  float_t alpha, beta;
149  norm_region region;
150 
151  ar(cereal::make_nvp("in_shape", in_shape),
152  cereal::make_nvp("size", size),
153  cereal::make_nvp("alpha", alpha),
154  cereal::make_nvp("beta", beta),
155  cereal::make_nvp("region", region));
156  construct(in_shape, size, alpha, beta, region);
157  }
158 
159  template <class Archive>
160  void serialize(Archive & ar) {
161  layer::serialize_prolog(ar);
162  ar(cereal::make_nvp("in_shape", in_shape_),
163  cereal::make_nvp("size", size_),
164  cereal::make_nvp("alpha", alpha_),
165  cereal::make_nvp("beta", beta_),
166  cereal::make_nvp("region", region_));
167  }
168 
169 private:
170  void forward_across(const vec_t& in, vec_t& out) {
171  std::fill(in_square_.begin(), in_square_.end(), float_t(0));
172 
173  for (serial_size_t i = 0; i < size_ / 2; i++) {
174  serial_size_t idx = in_shape_.get_index(0, 0, i);
175  add_square_sum(&in[idx], in_shape_.area(), &in_square_[0]);
176  }
177 
178  serial_size_t head = size_ / 2;
179  long tail = static_cast<long>(head) - static_cast<long>(size_);
180  serial_size_t channels = in_shape_.depth_;
181  const serial_size_t wxh = in_shape_.area();
182  const float_t alpha_div_size = alpha_ / size_;
183 
184  for (serial_size_t i = 0; i < channels; i++, head++, tail++) {
185  if (head < channels)
186  add_square_sum(&in[in_shape_.get_index(0, 0, head)], wxh, &in_square_[0]);
187 
188  if (tail >= 0)
189  sub_square_sum(&in[in_shape_.get_index(0, 0, tail)], wxh, &in_square_[0]);
190 
191  float_t *dst = &out[in_shape_.get_index(0, 0, i)];
192  const float_t *src = &in[in_shape_.get_index(0, 0, i)];
193  for (serial_size_t j = 0; j < wxh; j++)
194  dst[j] = src[j] * std::pow(float_t(1) + alpha_div_size * in_square_[j], -beta_);
195  }
196  }
197 
198  void forward_within(const vec_t& in, vec_t& out) {
199  CNN_UNREFERENCED_PARAMETER(in);
200  CNN_UNREFERENCED_PARAMETER(out);
201  throw nn_error("not implemented");
202  }
203 
204  void add_square_sum(const float_t *src, serial_size_t size, float_t *dst) {
205  for (serial_size_t i = 0; i < size; i++)
206  dst[i] += src[i] * src[i];
207  }
208 
209  void sub_square_sum(const float_t *src, serial_size_t size, float_t *dst) {
210  for (serial_size_t i = 0; i < size; i++)
211  dst[i] -= src[i] * src[i];
212  }
213 
214  shape3d in_shape_;
215 
216  serial_size_t size_;
217  float_t alpha_, beta_;
218  norm_region region_;
219 
220  vec_t in_square_;
221 };
222 
223 } // namespace tiny_dnn
single-input, single-output network with activation function
Definition: feedforward_layer.h:37
base class of all kind of NN layers
Definition: layer.h:62
bool parallelize_
Flag indicating whether the layer/node operations ara paralellized.
Definition: layer.h:696
serial_size_t in_channels() const
number of outgoing edges in this layer
Definition: layer.h:146
local response normalization
Definition: lrn_layer.h:42
serial_size_t fan_out_size() const override
number of outgoing connections for each input unit used only for weight/bias initialization methods w...
Definition: lrn_layer.h:97
serial_size_t fan_in_size() const override
number of incoming connections for each output unit used only for weight/bias initialization methods ...
Definition: lrn_layer.h:93
std::vector< shape3d > in_shape() const override
array of input shapes (width x height x depth)
Definition: lrn_layer.h:101
lrn_layer(serial_size_t in_width, serial_size_t in_height, serial_size_t local_size, serial_size_t in_channels, float_t alpha=1.0, float_t beta=5.0, norm_region region=norm_region::across_channels)
Definition: lrn_layer.h:84
lrn_layer(layer *prev, serial_size_t local_size, float_t alpha=1.0, float_t beta=5.0, norm_region region=norm_region::across_channels)
Definition: lrn_layer.h:69
void back_propagation(const std::vector< tensor_t * > &in_data, const std::vector< tensor_t * > &out_data, std::vector< tensor_t * > &out_grad, std::vector< tensor_t * > &in_grad) override
return delta of previous layer (delta=\frac{dE}{da}, a=wx in fully-connected layer)
Definition: lrn_layer.h:133
std::string layer_type() const override
name of layer, should be unique for each concrete class
Definition: lrn_layer.h:109
std::vector< shape3d > out_shape() const override
array of output shapes (width x height x depth)
Definition: lrn_layer.h:105
void forward_propagation(const std::vector< tensor_t * > &in_data, std::vector< tensor_t * > &out_data) override
Definition: lrn_layer.h:111
error exception class for tiny-dnn
Definition: nn_error.h:37