tesseract 4.1.1
Loading...
Searching...
No Matches
weightmatrix.h
Go to the documentation of this file.
1
2// File: weightmatrix.h
3// Description: Hides distinction between float/int implementations.
4// Author: Ray Smith
5//
6// (C) Copyright 2014, Google Inc.
7// Licensed under the Apache License, Version 2.0 (the "License");
8// you may not use this file except in compliance with the License.
9// You may obtain a copy of the License at
10// http://www.apache.org/licenses/LICENSE-2.0
11// Unless required by applicable law or agreed to in writing, software
12// distributed under the License is distributed on an "AS IS" BASIS,
13// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14// See the License for the specific language governing permissions and
15// limitations under the License.
17
18#ifndef TESSERACT_LSTM_WEIGHTMATRIX_H_
19#define TESSERACT_LSTM_WEIGHTMATRIX_H_
20
21#include <memory>
22#include "genericvector.h"
23#include "intsimdmatrix.h"
24#include "matrix.h"
25#include "tprintf.h"
26
27namespace tesseract {
28
29// Convenience instantiation of GENERIC_2D_ARRAY<double> with additional
30// operations to write a strided vector, so the transposed form of the input
31// is memory-contiguous.
32class TransposedArray : public GENERIC_2D_ARRAY<double> {
33 public:
34 // Copies the whole input transposed, converted to double, into *this.
35 void Transpose(const GENERIC_2D_ARRAY<double>& input);
36 // Writes a vector of data representing a timestep (gradients or sources).
37 // The data is assumed to be of size1 in size (the strided dimension).
38 ~TransposedArray() override;
39 void WriteStrided(int t, const float* data) {
40 int size1 = dim1();
41 for (int i = 0; i < size1; ++i) put(i, t, data[i]);
42 }
43 void WriteStrided(int t, const double* data) {
44 int size1 = dim1();
45 for (int i = 0; i < size1; ++i) put(i, t, data[i]);
46 }
47 // Prints the first and last num elements of the un-transposed array.
48 void PrintUnTransposed(int num) {
49 int num_features = dim1();
50 int width = dim2();
51 for (int y = 0; y < num_features; ++y) {
52 for (int t = 0; t < width; ++t) {
53 if (num == 0 || t < num || t + num >= width) {
54 tprintf(" %g", (*this)(y, t));
55 }
56 }
57 tprintf("\n");
58 }
59 }
60}; // class TransposedArray
61
62// Generic weight matrix for network layers. Can store the matrix as either
63// an array of floats or int8_t. Provides functions to compute the forward and
64// backward steps with the matrix and updates to the weights.
66 public:
67 WeightMatrix() : int_mode_(false), use_adam_(false) {}
68 // Sets up the network for training. Initializes weights using weights of
69 // scale `range` picked according to the random number generator `randomizer`.
70 // Note the order is outputs, inputs, as this is the order of indices to
71 // the matrix, so the adjacent elements are multiplied by the input during
72 // a forward operation.
73 int InitWeightsFloat(int no, int ni, bool use_adam, float weight_range,
74 TRand* randomizer);
75 // Changes the number of outputs to the size of the given code_map, copying
76 // the old weight matrix entries for each output from code_map[output] where
77 // non-negative, and uses the mean (over all outputs) of the existing weights
78 // for all outputs with negative code_map entries. Returns the new number of
79 // weights.
80 int RemapOutputs(const std::vector<int>& code_map);
81
82 // Converts a float network to an int network. Each set of input weights that
83 // corresponds to a single output weight is converted independently:
84 // Compute the max absolute value of the weight set.
85 // Scale so the max absolute value becomes INT8_MAX.
86 // Round to integer.
87 // Store a multiplicative scale factor (as a float) that will reproduce
88 // the original value, subject to rounding errors.
89 void ConvertToInt();
90 // Returns the size rounded up to an internal factor used by the SIMD
91 // implementation for its input.
92 int RoundInputs(int size) const {
93 if (!int_mode_ || !IntSimdMatrix::intSimdMatrix) return size;
95 }
96
97 // Accessors.
98 bool is_int_mode() const {
99 return int_mode_;
100 }
101 int NumOutputs() const { return int_mode_ ? wi_.dim1() : wf_.dim1(); }
102 // Provides one set of weights. Only used by peep weight maxpool.
103 const double* GetWeights(int index) const { return wf_[index]; }
104 // Provides access to the deltas (dw_).
105 double GetDW(int i, int j) const { return dw_(i, j); }
106
107 // Allocates any needed memory for running Backward, and zeroes the deltas,
108 // thus eliminating any existing momentum.
109 void InitBackward();
110
111 // Writes to the given file. Returns false in case of error.
112 bool Serialize(bool training, TFile* fp) const;
113 // Reads from the given file. Returns false in case of error.
114 bool DeSerialize(bool training, TFile* fp);
115 // As DeSerialize, but reads an old (float) format WeightMatrix for
116 // backward compatibility.
117 bool DeSerializeOld(bool training, TFile* fp);
118
119 // Computes matrix.vector v = Wu.
120 // u is of size W.dim2() - 1 and the output v is of size W.dim1().
121 // u is imagined to have an extra element at the end with value 1, to
122 // implement the bias, but it doesn't actually have it.
123 // Asserts that the call matches what we have.
124 void MatrixDotVector(const double* u, double* v) const;
125 void MatrixDotVector(const int8_t* u, double* v) const;
126 // MatrixDotVector for peep weights, MultiplyAccumulate adds the
127 // component-wise products of *this[0] and v to inout.
128 void MultiplyAccumulate(const double* v, double* inout);
129 // Computes vector.matrix v = uW.
130 // u is of size W.dim1() and the output v is of size W.dim2() - 1.
131 // The last result is discarded, as v is assumed to have an imaginary
132 // last value of 1, as with MatrixDotVector.
133 void VectorDotMatrix(const double* u, double* v) const;
134 // Fills dw_[i][j] with the dot product u[i][] . v[j][], using elements
135 // from u and v, starting with u[i][offset] and v[j][offset].
136 // Note that (matching MatrixDotVector) v[last][] is missing, presumed 1.0.
137 // Runs parallel if requested. Note that inputs must be transposed.
138 void SumOuterTransposed(const TransposedArray& u, const TransposedArray& v,
139 bool parallel);
140 // Updates the weights using the given learning rate, momentum and adam_beta.
141 // num_samples is used in the Adam correction factor.
142 void Update(double learning_rate, double momentum, double adam_beta,
143 int num_samples);
144 // Adds the dw_ in other to the dw_ is *this.
145 void AddDeltas(const WeightMatrix& other);
146 // Sums the products of weight updates in *this and other, splitting into
147 // positive (same direction) in *same and negative (different direction) in
148 // *changed.
149 void CountAlternators(const WeightMatrix& other, double* same,
150 double* changed) const;
151
152 void Debug2D(const char* msg);
153
154 // Utility function converts an array of float to the corresponding array
155 // of double.
156 static void FloatToDouble(const GENERIC_2D_ARRAY<float>& wf,
158
159 private:
160 // Choice between float and 8 bit int implementations.
163 // Transposed copy of wf_, used only for Backward, and set with each Update.
164 TransposedArray wf_t_;
165 // Which of wf_ and wi_ are we actually using.
166 bool int_mode_;
167 // True if we are running adam in this weight matrix.
168 bool use_adam_;
169 // If we are using wi_, then scales_ is a factor to restore the row product
170 // with a vector to the correct range.
171 GenericVector<double> scales_;
172 // Weight deltas. dw_ is the new delta, and updates_ the momentum-decaying
173 // amount to be added to wf_/wi_.
176 // Iff use_adam_, the sum of squares of dw_. The number of samples is
177 // given to Update(). Serialized iff use_adam_.
178 GENERIC_2D_ARRAY<double> dw_sq_sum_;
179 // The weights matrix reorganized in whatever way suits this instance.
180 std::vector<int8_t> shaped_w_;
181};
182
183} // namespace tesseract.
184
185#endif // TESSERACT_LSTM_WEIGHTMATRIX_H_
DLLSYM void tprintf(const char *format,...)
Definition: tprintf.cpp:35
void put(ICOORD pos, const double &thing)
Definition: matrix.h:223
int RoundInputs(int size) const
Definition: intsimdmatrix.h:69
static const IntSimdMatrix * intSimdMatrix
void WriteStrided(int t, const double *data)
Definition: weightmatrix.h:43
void WriteStrided(int t, const float *data)
Definition: weightmatrix.h:39
void Transpose(const GENERIC_2D_ARRAY< double > &input)
void PrintUnTransposed(int num)
Definition: weightmatrix.h:48
void SumOuterTransposed(const TransposedArray &u, const TransposedArray &v, bool parallel)
bool Serialize(bool training, TFile *fp) const
bool DeSerializeOld(bool training, TFile *fp)
const double * GetWeights(int index) const
Definition: weightmatrix.h:103
void Update(double learning_rate, double momentum, double adam_beta, int num_samples)
int InitWeightsFloat(int no, int ni, bool use_adam, float weight_range, TRand *randomizer)
void Debug2D(const char *msg)
bool is_int_mode() const
Definition: weightmatrix.h:98
static void FloatToDouble(const GENERIC_2D_ARRAY< float > &wf, GENERIC_2D_ARRAY< double > *wd)
void CountAlternators(const WeightMatrix &other, double *same, double *changed) const
void AddDeltas(const WeightMatrix &other)
int RemapOutputs(const std::vector< int > &code_map)
void MultiplyAccumulate(const double *v, double *inout)
void MatrixDotVector(const double *u, double *v) const
double GetDW(int i, int j) const
Definition: weightmatrix.h:105
bool DeSerialize(bool training, TFile *fp)
int RoundInputs(int size) const
Definition: weightmatrix.h:92
void VectorDotMatrix(const double *u, double *v) const