tiny_dnn  1.0.0
A header only, dependency-free deep learning framework in C++11
fully_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/layers/layer.h"
29 
30 #include "tiny_dnn/core/kernels/fully_connected_op.h"
31 #include "tiny_dnn/core/kernels/fully_connected_grad_op.h"
32 
33 namespace tiny_dnn {
34 
38 template<typename Activation>
39 class fully_connected_layer : public feedforward_layer<Activation> {
40 public:
42  CNN_USE_LAYER_MEMBERS;
43 
49  fully_connected_layer(serial_size_t in_dim,
50  serial_size_t out_dim,
51  bool has_bias = true,
52  backend_t backend_type = core::default_engine())
53  : Base(std_input_order(has_bias)) {
54  set_params(in_dim, out_dim, has_bias);
55  init_backend(backend_type);
56  Base::set_backend_type(backend_type);
57  }
58 
59  // move constructor
61  : Base(std::move(other))
62  , params_(std::move(other.params_))
63  , kernel_fwd_(std::move(other.kernel_fwd_))
64  , kernel_back_(std::move(other.kernel_back_)) {
65  init_backend(std::move(other.engine()));
66  }
67 
68  serial_size_t fan_in_size() const override {
69  return params_.in_size_;
70  }
71 
72  serial_size_t fan_out_size() const override {
73  return params_.out_size_;
74  }
75 
76  std::vector<index3d<serial_size_t>> in_shape() const override {
77  if (params_.has_bias_) {
78  return { index3d<serial_size_t>(params_.in_size_, 1, 1),
79  index3d<serial_size_t>(params_.in_size_,
80  params_.out_size_, 1),
81  index3d<serial_size_t>(params_.out_size_, 1, 1) };
82  } else {
83  return { index3d<serial_size_t>(params_.in_size_, 1, 1),
84  index3d<serial_size_t>(params_.in_size_,
85  params_.out_size_, 1) };
86  }
87  }
88 
89  std::vector<index3d<serial_size_t>> out_shape() const override {
90  return { index3d<serial_size_t>(params_.out_size_, 1, 1),
91  index3d<serial_size_t>(params_.out_size_, 1, 1) };
92  }
93 
94  void forward_propagation(const std::vector<tensor_t*>& in_data,
95  std::vector<tensor_t*>& out_data) override {
96  // forward convolutional op context
97  auto ctx = OpKernelContext(in_data, out_data);
98  ctx.setParallelize(layer::parallelize());
99  ctx.setEngine(layer::engine());
100 
101  // launch convolutional kernel
102  kernel_fwd_->compute(ctx);
103 
104  // activations
105  this->forward_activation(*out_data[0], *out_data[1]);
106  }
107 
108  void back_propagation(const std::vector<tensor_t*>& in_data,
109  const std::vector<tensor_t*>& out_data,
110  std::vector<tensor_t*>& out_grad,
111  std::vector<tensor_t*>& in_grad) override {
112  // activations
113  // TODO(edgar/nyanp): refactor and move activations outside
114  this->backward_activation(*out_grad[0], *out_data[0], *out_grad[1]);
115 
116  // backward convolutional op context
117  auto ctx = OpKernelContext(in_data, out_data, out_grad, in_grad);
118  ctx.setParallelize(layer::parallelize());
119  ctx.setEngine(layer::engine());
120 
121  // launch convolutional kernel
122  kernel_back_->compute(ctx);
123  }
124 
125  std::string layer_type() const override { return "fully-connected"; }
126 
127  template <class Archive>
128  static void load_and_construct(Archive & ar, cereal::construct<fully_connected_layer> & construct) {
129  serial_size_t in_dim, out_dim;
130  bool has_bias;
131 
132  ar(cereal::make_nvp("in_size", in_dim),
133  cereal::make_nvp("out_size", out_dim),
134  cereal::make_nvp("has_bias", has_bias));
135  construct(in_dim, out_dim, has_bias);
136  }
137 
138  template <class Archive>
139  void serialize(Archive & ar) {
140  layer::serialize_prolog(ar);
141  ar(cereal::make_nvp("in_size", params_.in_size_),
142  cereal::make_nvp("out_size", params_.out_size_),
143  cereal::make_nvp("has_bias", params_.has_bias_));
144  }
145 
146 protected:
147 
148  void set_params(const serial_size_t in_size,
149  const serial_size_t out_size,
150  bool has_bias) {
151  params_.in_size_ = in_size;
152  params_.out_size_ = out_size;
153  params_.has_bias_ = has_bias;
154  }
155 
156  void init_backend(backend_t backend_type) {
157  core::OpKernelConstruction ctx =
158  core::OpKernelConstruction(layer::device(), &params_);
159 
160  if (backend_type == backend_t::internal ||
161  backend_type == backend_t::avx||
162  backend_type == backend_t::nnpack
163  ) {
164 
165  kernel_fwd_.reset(new FullyConnectedOp(ctx));
166  kernel_back_.reset(new FullyConnectedGradOp(ctx));
167 
168  return;
169  }
170  else {
171  throw nn_error("Not supported engine: " + to_string(backend_type));
172  }
173  }
174 
175  private:
176  /* The layer parameters */
177  fully_params params_;
178 
179  /* Forward and backward ops */
180  std::shared_ptr<core::OpKernel> kernel_fwd_;
181  std::shared_ptr<core::OpKernel> kernel_back_;
182 };
183 
184 } // namespace tiny_dnn
single-input, single-output network with activation function
Definition: feedforward_layer.h:37
compute fully-connected(matmul) operation
Definition: fully_connected_layer.h:39
std::vector< index3d< serial_size_t > > out_shape() const override
array of output shapes (width x height x depth)
Definition: fully_connected_layer.h:89
std::string layer_type() const override
name of layer, should be unique for each concrete class
Definition: fully_connected_layer.h:125
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: fully_connected_layer.h:108
fully_connected_layer(serial_size_t in_dim, serial_size_t out_dim, bool has_bias=true, backend_t backend_type=core::default_engine())
Definition: fully_connected_layer.h:49
void forward_propagation(const std::vector< tensor_t * > &in_data, std::vector< tensor_t * > &out_data) override
Definition: fully_connected_layer.h:94
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: fully_connected_layer.h:72
serial_size_t fan_in_size() const override
number of incoming connections for each output unit used only for weight/bias initialization methods ...
Definition: fully_connected_layer.h:68
std::vector< index3d< serial_size_t > > in_shape() const override
array of input shapes (width x height x depth)
Definition: fully_connected_layer.h:76
serial_size_t out_size() const
!
Definition: layer.h:181
serial_size_t in_size() const
!
Definition: layer.h:176