tesseract 4.1.1
Loading...
Searching...
No Matches
fullyconnected.h
Go to the documentation of this file.
1
2// File: fullyconnected.h
3// Description: Simple feed-forward layer with various non-linearities.
4// Author: Ray Smith
5// Created: Wed Feb 26 14:46:06 PST 2014
6//
7// (C) Copyright 2014, 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_FULLYCONNECTED_H_
20#define TESSERACT_LSTM_FULLYCONNECTED_H_
21
22#include "network.h"
23#include "networkscratch.h"
24
25namespace tesseract {
26
27// C++ Implementation of the Softmax (output) class from lstm.py.
28class FullyConnected : public Network {
29 public:
30 FullyConnected(const STRING& name, int ni, int no, NetworkType type);
31 ~FullyConnected() override = default;
32
33 // Returns the shape output from the network given an input shape (which may
34 // be partially unknown ie zero).
35 StaticShape OutputShape(const StaticShape& input_shape) const override;
36
37 STRING spec() const override {
39 if (type_ == NT_TANH)
40 spec.add_str_int("Ft", no_);
41 else if (type_ == NT_LOGISTIC)
42 spec.add_str_int("Fs", no_);
43 else if (type_ == NT_RELU)
44 spec.add_str_int("Fr", no_);
45 else if (type_ == NT_LINEAR)
46 spec.add_str_int("Fl", no_);
47 else if (type_ == NT_POSCLIP)
48 spec.add_str_int("Fp", no_);
49 else if (type_ == NT_SYMCLIP)
50 spec.add_str_int("Fs", no_);
51 else if (type_ == NT_SOFTMAX)
52 spec.add_str_int("Fc", no_);
53 else
54 spec.add_str_int("Fm", no_);
55 return spec;
56 }
57
58 // Changes the type to the given type. Used to commute a softmax to a
59 // non-output type for adding on other networks.
61 type_ = type;
62 }
63
64 // Suspends/Enables training by setting the training_ flag. Serialize and
65 // DeSerialize only operate on the run-time data if state is false.
66 void SetEnableTraining(TrainingState state) override;
67
68 // Sets up the network for training. Initializes weights using weights of
69 // scale `range` picked according to the random number generator `randomizer`.
70 int InitWeights(float range, TRand* randomizer) override;
71 // Recursively searches the network for softmaxes with old_no outputs,
72 // and remaps their outputs according to code_map. See network.h for details.
73 int RemapOutputs(int old_no, const std::vector<int>& code_map) override;
74
75 // Converts a float network to an int network.
76 void ConvertToInt() override;
77
78 // Provides debug output on the weights.
79 void DebugWeights() override;
80
81 // Writes to the given file. Returns false in case of error.
82 bool Serialize(TFile* fp) const override;
83 // Reads from the given file. Returns false in case of error.
84 bool DeSerialize(TFile* fp) override;
85
86 // Runs forward propagation of activations on the input line.
87 // See Network for a detailed discussion of the arguments.
88 void Forward(bool debug, const NetworkIO& input,
89 const TransposedArray* input_transpose, NetworkScratch* scratch,
90 NetworkIO* output) override;
91 // Components of Forward so FullyConnected can be reused inside LSTM.
92 void SetupForward(const NetworkIO& input,
93 const TransposedArray* input_transpose);
94 void ForwardTimeStep(int t, double* output_line);
95 void ForwardTimeStep(const double* d_input, int t, double* output_line);
96 void ForwardTimeStep(const int8_t* i_input, int t, double* output_line);
97
98 // Runs backward propagation of errors on the deltas line.
99 // See Network for a detailed discussion of the arguments.
100 bool Backward(bool debug, const NetworkIO& fwd_deltas,
101 NetworkScratch* scratch, NetworkIO* back_deltas) override;
102 // Components of Backward so FullyConnected can be reused inside LSTM.
103 void BackwardTimeStep(const NetworkIO& fwd_deltas, int t, double* curr_errors,
104 TransposedArray* errors_t, double* backprop);
105 void FinishBackward(const TransposedArray& errors_t);
106
107 // Updates the weights using the given learning rate, momentum and adam_beta.
108 // num_samples is used in the adam computation iff use_adam_ is true.
109 void Update(float learning_rate, float momentum, float adam_beta,
110 int num_samples) override;
111 // Sums the products of weight updates in *this and other, splitting into
112 // positive (same direction) in *same and negative (different direction) in
113 // *changed.
114 void CountAlternators(const Network& other, double* same,
115 double* changed) const override;
116
117 protected:
118 // Weight arrays of size [no, ni + 1].
120 // Transposed copy of input used during training of size [ni, width].
122 // Pointer to transposed input stored elsewhere. If not null, this is used
123 // in preference to calculating the transpose and storing it in source_t_.
125 // Activations from forward pass of size [width, no].
127 // Memory of the integer mode input to forward as softmax always outputs
128 // float, so the information is otherwise lost.
130};
131
132} // namespace tesseract.
133
134
135
136#endif // TESSERACT_LSTM_FULLYCONNECTED_H_
TrainingState
Definition: network.h:92
NetworkType
Definition: network.h:43
@ NT_LINEAR
Definition: network.h:67
@ NT_RELU
Definition: network.h:66
@ NT_SOFTMAX
Definition: network.h:68
@ NT_LOGISTIC
Definition: network.h:62
@ NT_SYMCLIP
Definition: network.h:64
@ NT_POSCLIP
Definition: network.h:63
@ NT_TANH
Definition: network.h:65
Definition: strngs.h:45
void add_str_int(const char *str, int number)
Definition: strngs.cpp:377
void BackwardTimeStep(const NetworkIO &fwd_deltas, int t, double *curr_errors, TransposedArray *errors_t, double *backprop)
bool DeSerialize(TFile *fp) override
void FinishBackward(const TransposedArray &errors_t)
void SetupForward(const NetworkIO &input, const TransposedArray *input_transpose)
bool Backward(bool debug, const NetworkIO &fwd_deltas, NetworkScratch *scratch, NetworkIO *back_deltas) override
void CountAlternators(const Network &other, double *same, double *changed) const override
void SetEnableTraining(TrainingState state) override
const TransposedArray * external_source_
void Update(float learning_rate, float momentum, float adam_beta, int num_samples) override
int InitWeights(float range, TRand *randomizer) override
void Forward(bool debug, const NetworkIO &input, const TransposedArray *input_transpose, NetworkScratch *scratch, NetworkIO *output) override
void ChangeType(NetworkType type)
int RemapOutputs(int old_no, const std::vector< int > &code_map) override
STRING spec() const override
void ForwardTimeStep(int t, double *output_line)
void ConvertToInt() override
~FullyConnected() override=default
StaticShape OutputShape(const StaticShape &input_shape) const override
bool Serialize(TFile *fp) const override
NetworkType type_
Definition: network.h:293
const STRING & name() const
Definition: network.h:138
NetworkType type() const
Definition: network.h:112