tiny_dnn  1.0.0
A header only, dependency-free deep learning framework in C++11
partial_connected_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 "tiny_dnn/layers/layer.h"
30 
31 namespace tiny_dnn {
32 
33 template<typename Activation>
34 class partial_connected_layer : public feedforward_layer<Activation> {
35 public:
36  CNN_USE_LAYER_MEMBERS;
37 
38  typedef std::vector<std::pair<serial_size_t, serial_size_t> > io_connections;
39  typedef std::vector<std::pair<serial_size_t, serial_size_t> > wi_connections;
40  typedef std::vector<std::pair<serial_size_t, serial_size_t> > wo_connections;
42 
43  partial_connected_layer(serial_size_t in_dim,
44  serial_size_t out_dim,
45  size_t weight_dim,
46  size_t bias_dim,
47  float_t scale_factor = float_t(1))
48  : Base(std_input_order(bias_dim > 0)),
49  weight2io_(weight_dim),
50  out2wi_(out_dim),
51  in2wo_(in_dim),
52  bias2out_(bias_dim),
53  out2bias_(out_dim),
54  scale_factor_(scale_factor){}
55 
56  size_t param_size() const {
57  size_t total_param = 0;
58  for (auto w : weight2io_)
59  if (w.size() > 0) total_param++;
60  for (auto b : bias2out_)
61  if (b.size() > 0) total_param++;
62  return total_param;
63  }
64 
65  serial_size_t fan_in_size() const override {
66  return max_size(out2wi_);
67  }
68 
69  serial_size_t fan_out_size() const override {
70  return max_size(in2wo_);
71  }
72 
73  void connect_weight(serial_size_t input_index, serial_size_t output_index, serial_size_t weight_index) {
74  weight2io_[weight_index].emplace_back(input_index, output_index);
75  out2wi_[output_index].emplace_back(weight_index, input_index);
76  in2wo_[input_index].emplace_back(weight_index, output_index);
77  }
78 
79  void connect_bias(serial_size_t bias_index, serial_size_t output_index) {
80  out2bias_[output_index] = bias_index;
81  bias2out_[bias_index].push_back(output_index);
82  }
83 
84  void forward_propagation(const std::vector<tensor_t*>& in_data,
85  std::vector<tensor_t*>& out_data) override {
86  const tensor_t& in = *in_data[0];
87  const vec_t& W = (*in_data[1])[0];
88  const vec_t& b = (*in_data[2])[0];
89  tensor_t& a = *out_data[1];
90 
91  // @todo revise the parallelism strategy
92  for (serial_size_t sample = 0, sample_count = static_cast<serial_size_t>(in.size()); sample < sample_count; ++sample) {
93  vec_t& a_sample = a[sample];
94 
95  for_i(parallelize_, out2wi_.size(), [&](int i) {
96  const wi_connections& connections = out2wi_[i];
97 
98  float_t& a_element = a_sample[i];
99 
100  a_element = float_t(0);
101 
102  for (auto connection : connections)// 13.1%
103  a_element += W[connection.first] * in[sample][connection.second]; // 3.2%
104 
105  a_element *= scale_factor_;
106  a_element += b[out2bias_[i]];
107  });
108  }
109 
110  this->forward_activation(*out_data[0], *out_data[1]);
111  }
112 
113  void back_propagation(const std::vector<tensor_t*>& in_data,
114  const std::vector<tensor_t*>& out_data,
115  std::vector<tensor_t*>& out_grad,
116  std::vector<tensor_t*>& in_grad) override {
117  const tensor_t& prev_out = *in_data[0];
118  const vec_t& W = (*in_data[1])[0];
119  vec_t& dW = (*in_grad[1])[0];
120  vec_t& db = (*in_grad[2])[0];
121  tensor_t& prev_delta = *in_grad[0];
122  tensor_t& curr_delta = *out_grad[0];
123 
124  this->backward_activation(*out_grad[0], *out_data[0], curr_delta);
125 
126  // @todo revise the parallelism strategy
127  for (serial_size_t sample = 0, sample_count = static_cast<serial_size_t>(prev_out.size()); sample < sample_count; ++sample) {
128  for_(parallelize_, 0, in2wo_.size(), [&](const blocked_range& r) {
129  for (int i = r.begin(); i != r.end(); i++) {
130  const wo_connections& connections = in2wo_[i];
131  float_t delta = float_t(0);
132 
133  for (auto connection : connections)
134  delta += W[connection.first] * curr_delta[sample][connection.second]; // 40.6%
135 
136  prev_delta[sample][i] = delta * scale_factor_; // 2.1%
137  }
138  });
139 
140  for_(parallelize_, 0, weight2io_.size(), [&](const blocked_range& r) {
141  for (int i = r.begin(); i < r.end(); i++) {
142  const io_connections& connections = weight2io_[i];
143  float_t diff = float_t(0);
144 
145  for (auto connection : connections) // 11.9%
146  diff += prev_out[sample][connection.first] * curr_delta[sample][connection.second];
147 
148  dW[i] += diff * scale_factor_;
149  }
150  });
151 
152  for (size_t i = 0; i < bias2out_.size(); i++) {
153  const std::vector<serial_size_t>& outs = bias2out_[i];
154  float_t diff = float_t(0);
155 
156  for (auto o : outs)
157  diff += curr_delta[sample][o];
158 
159  db[i] += diff;
160  }
161  }
162  }
163 
164 protected:
165  std::vector<io_connections> weight2io_; // weight_id -> [(in_id, out_id)]
166  std::vector<wi_connections> out2wi_; // out_id -> [(weight_id, in_id)]
167  std::vector<wo_connections> in2wo_; // in_id -> [(weight_id, out_id)]
168  std::vector<std::vector<serial_size_t> > bias2out_;
169  std::vector<size_t> out2bias_;
170  float_t scale_factor_;
171 };
172 
173 } // namespace tiny_dnn
single-input, single-output network with activation function
Definition: feedforward_layer.h:37
bool parallelize_
Flag indicating whether the layer/node operations ara paralellized.
Definition: layer.h:696
Definition: partial_connected_layer.h:34
serial_size_t fan_in_size() const override
number of incoming connections for each output unit used only for weight/bias initialization methods ...
Definition: partial_connected_layer.h:65
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: partial_connected_layer.h:113
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: partial_connected_layer.h:69
void forward_propagation(const std::vector< tensor_t * > &in_data, std::vector< tensor_t * > &out_data) override
Definition: partial_connected_layer.h:84
Definition: parallel_for.h:70