tesseract  v4.0.0-17-g361f3264
Open Source OCR Engine
lstm.h
1 // File: lstm.h
3 // Description: Long-term-short-term-memory Recurrent neural network.
4 // Author: Ray Smith
5 // Created: Wed May 01 17:33:06 PST 2013
6 //
7 // (C) Copyright 2013, Google Inc.
8 // Licensed under the Apache License, Version 2.0 (the "License");
9 // you may not use this file except in compliance with the License.
10 // You may obtain a copy of the License at
11 // http://www.apache.org/licenses/LICENSE-2.0
12 // Unless required by applicable law or agreed to in writing, software
13 // distributed under the License is distributed on an "AS IS" BASIS,
14 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 // See the License for the specific language governing permissions and
16 // limitations under the License.
18 
19 #ifndef TESSERACT_LSTM_LSTM_H_
20 #define TESSERACT_LSTM_LSTM_H_
21 
22 #include "network.h"
23 #include "fullyconnected.h"
24 
25 namespace tesseract {
26 
27 // C++ Implementation of the LSTM class from lstm.py.
28 class LSTM : public Network {
29  public:
30  // Enum for the different weights in LSTM, to reduce some of the I/O and
31  // setup code to loops. The elements of the enum correspond to elements of an
32  // array of WeightMatrix or a corresponding array of NetworkIO.
33  enum WeightType {
34  CI, // Cell Inputs.
35  GI, // Gate at the input.
36  GF1, // Forget gate at the memory (1-d or looking back 1 timestep).
37  GO, // Gate at the output.
38  GFS, // Forget gate at the memory, looking back in the other dimension.
39 
40  WT_COUNT // Number of WeightTypes.
41  };
42 
43  // Constructor for NT_LSTM (regular 1 or 2-d LSTM), NT_LSTM_SOFTMAX (LSTM with
44  // additional softmax layer included and fed back into the input at the next
45  // timestep), or NT_LSTM_SOFTMAX_ENCODED (as LSTM_SOFTMAX, but the feedback
46  // is binary encoded instead of categorical) only.
47  // 2-d and bidi softmax LSTMs are not rejected, but are impossible to build
48  // in the conventional way because the output feedback both forwards and
49  // backwards in time does become impossible.
50  LSTM(const STRING& name, int num_inputs, int num_states, int num_outputs,
51  bool two_dimensional, NetworkType type);
52  virtual ~LSTM();
53 
54  // Returns the shape output from the network given an input shape (which may
55  // be partially unknown ie zero).
56  StaticShape OutputShape(const StaticShape& input_shape) const override;
57 
58  STRING spec() const override {
59  STRING spec;
60  if (type_ == NT_LSTM)
61  spec.add_str_int("Lfx", ns_);
62  else if (type_ == NT_LSTM_SUMMARY)
63  spec.add_str_int("Lfxs", ns_);
64  else if (type_ == NT_LSTM_SOFTMAX)
65  spec.add_str_int("LS", ns_);
66  else if (type_ == NT_LSTM_SOFTMAX_ENCODED)
67  spec.add_str_int("LE", ns_);
68  if (softmax_ != nullptr) spec += softmax_->spec();
69  return spec;
70  }
71 
72  // Suspends/Enables training by setting the training_ flag. Serialize and
73  // DeSerialize only operate on the run-time data if state is false.
74  void SetEnableTraining(TrainingState state) override;
75 
76  // Sets up the network for training. Initializes weights using weights of
77  // scale `range` picked according to the random number generator `randomizer`.
78  int InitWeights(float range, TRand* randomizer) override;
79  // Recursively searches the network for softmaxes with old_no outputs,
80  // and remaps their outputs according to code_map. See network.h for details.
81  int RemapOutputs(int old_no, const std::vector<int>& code_map) override;
82 
83  // Converts a float network to an int network.
84  void ConvertToInt() override;
85 
86  // Provides debug output on the weights.
87  void DebugWeights() override;
88 
89  // Writes to the given file. Returns false in case of error.
90  bool Serialize(TFile* fp) const override;
91  // Reads from the given file. Returns false in case of error.
92  bool DeSerialize(TFile* fp) override;
93 
94  // Runs forward propagation of activations on the input line.
95  // See Network for a detailed discussion of the arguments.
96  void Forward(bool debug, const NetworkIO& input,
97  const TransposedArray* input_transpose, NetworkScratch* scratch,
98  NetworkIO* output) override;
99 
100  // Runs backward propagation of errors on the deltas line.
101  // See Network for a detailed discussion of the arguments.
102  bool Backward(bool debug, const NetworkIO& fwd_deltas,
103  NetworkScratch* scratch, NetworkIO* back_deltas) override;
104  // Updates the weights using the given learning rate, momentum and adam_beta.
105  // num_samples is used in the adam computation iff use_adam_ is true.
106  void Update(float learning_rate, float momentum, float adam_beta,
107  int num_samples) override;
108  // Sums the products of weight updates in *this and other, splitting into
109  // positive (same direction) in *same and negative (different direction) in
110  // *changed.
111  void CountAlternators(const Network& other, double* same,
112  double* changed) const override;
113  // Prints the weights for debug purposes.
114  void PrintW();
115  // Prints the weight deltas for debug purposes.
116  void PrintDW();
117 
118  // Returns true of this is a 2-d lstm.
119  bool Is2D() const {
120  return is_2d_;
121  }
122 
123  private:
124  // Resizes forward data to cope with an input image of the given width.
125  void ResizeForward(const NetworkIO& input);
126 
127  private:
128  // Size of padded input to weight matrices = ni_ + no_ for 1-D operation
129  // and ni_ + 2 * no_ for 2-D operation. Note that there is a phantom 1 input
130  // for the bias that makes the weight matrices of size [na + 1][no].
131  int32_t na_;
132  // Number of internal states. Equal to no_ except for a softmax LSTM.
133  // ns_ is NOT serialized, but is calculated from gate_weights_.
134  int32_t ns_;
135  // Number of additional feedback states. The softmax types feed back
136  // additional output information on top of the ns_ internal states.
137  // In the case of a binary-coded (EMBEDDED) softmax, nf_ < no_.
138  int32_t nf_;
139  // Flag indicating 2-D operation.
140  bool is_2d_;
141 
142  // Gate weight arrays of size [na + 1, no].
144  // Used only if this is a softmax LSTM.
146  // Input padded with previous output of size [width, na].
148  // Internal state used during forward operation, of size [width, ns].
150  // State of the 2-d maxpool, generated during forward, used during backward.
152  // Internal state saved from forward, but used only during backward.
154  // Preserved input stride_map used for Backward when NT_LSTM_SQUASHED.
157 };
158 
159 } // namespace tesseract.
160 
161 
162 #endif // TESSERACT_LSTM_LSTM_H_
void add_str_int(const char *str, int number)
Definition: strngs.cpp:379
int input_width_
Definition: lstm.h:156
Definition: network.h:75
FullyConnected * softmax_
Definition: lstm.h:145
Definition: helpers.h:42
Definition: static_shape.h:38
Definition: network.h:76
WeightMatrix gate_weights_[WT_COUNT]
Definition: lstm.h:143
bool Is2D() const
Definition: lstm.h:119
bool is_2d_
Definition: lstm.h:140
Definition: networkscratch.h:36
int InitWeights(float range, TRand *randomizer) override
Definition: lstm.cpp:158
int32_t ns_
Definition: lstm.h:134
void ResizeForward(const NetworkIO &input)
Definition: lstm.cpp:753
Definition: lstm.h:40
Definition: lstm.h:36
Definition: network.h:60
Definition: serialis.h:77
void PrintDW()
Definition: lstm.cpp:727
STRING spec() const override
Definition: lstm.h:58
Definition: baseapi.cpp:94
GENERIC_2D_ARRAY< int8_t > which_fg_
Definition: lstm.h:151
const STRING & name() const
Definition: network.h:138
NetworkType
Definition: network.h:43
Definition: lstm.h:34
int32_t nf_
Definition: lstm.h:138
NetworkIO source_
Definition: lstm.h:147
Definition: lstm.h:38
NetworkIO node_values_[WT_COUNT]
Definition: lstm.h:153
Definition: weightmatrix.h:33
void Forward(bool debug, const NetworkIO &input, const TransposedArray *input_transpose, NetworkScratch *scratch, NetworkIO *output) override
Definition: lstm.cpp:250
Definition: network.h:105
void SetEnableTraining(TrainingState state) override
Definition: lstm.cpp:137
int32_t na_
Definition: lstm.h:131
Definition: weightmatrix.h:66
NetworkType type() const
Definition: network.h:112
bool Serialize(TFile *fp) const override
Definition: lstm.cpp:207
NetworkIO state_
Definition: lstm.h:149
bool DeSerialize(TFile *fp) override
Definition: lstm.cpp:220
int RemapOutputs(int old_no, const std::vector< int > &code_map) override
Definition: lstm.cpp:174
Definition: fullyconnected.h:28
void CountAlternators(const Network &other, double *same, double *changed) const override
Definition: lstm.cpp:687
Definition: strngs.h:45
Definition: lstm.h:35
void DebugWeights() override
Definition: lstm.cpp:194
TrainingState
Definition: network.h:92
Definition: lstm.h:28
StrideMap input_map_
Definition: lstm.h:155
StaticShape OutputShape(const StaticShape &input_shape) const override
Definition: lstm.cpp:127
LSTM(const STRING &name, int num_inputs, int num_states, int num_outputs, bool two_dimensional, NetworkType type)
Definition: lstm.cpp:99
void ConvertToInt() override
Definition: lstm.cpp:183
Definition: stridemap.h:43
Definition: networkio.h:39
STRING spec() const override
Definition: fullyconnected.h:37
NetworkType type_
Definition: network.h:299
void Update(float learning_rate, float momentum, float adam_beta, int num_samples) override
Definition: lstm.cpp:667
bool Backward(bool debug, const NetworkIO &fwd_deltas, NetworkScratch *scratch, NetworkIO *back_deltas) override
Definition: lstm.cpp:441
void PrintW()
Definition: lstm.cpp:701
Definition: lstm.h:37
virtual ~LSTM()
Definition: lstm.cpp:123
Definition: network.h:61
WeightType
Definition: lstm.h:33