Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/lstm/weightmatrix.h @ 2:b50eed0cc0ef upstream

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
ADD: MuPDF v1.26.7: the MuPDF source as downloaded by a default build of PyMuPDF 1.26.4. The directory name has changed: no version number in the expanded directory now.
author Franz Glasner <fzglas.hg@dom66.de>
date Mon, 15 Sep 2025 11:43:07 +0200
parents
children
comparison
equal deleted inserted replaced
1:1d09e1dec1d9 2:b50eed0cc0ef
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.
16 ///////////////////////////////////////////////////////////////////////
17
18 #ifndef TESSERACT_LSTM_WEIGHTMATRIX_H_
19 #define TESSERACT_LSTM_WEIGHTMATRIX_H_
20
21 #include <memory>
22 #include <vector>
23 #include "intsimdmatrix.h"
24 #include "matrix.h"
25 #include "tesstypes.h"
26 #include "tprintf.h"
27
28 namespace tesseract {
29
30 // Convenience instantiation of GENERIC_2D_ARRAY<TFloat> with additional
31 // operations to write a strided vector, so the transposed form of the input
32 // is memory-contiguous.
33 class TransposedArray : public GENERIC_2D_ARRAY<TFloat> {
34 public:
35 // Copies the whole input transposed, converted to TFloat, into *this.
36 void Transpose(const GENERIC_2D_ARRAY<TFloat> &input);
37 // Writes a vector of data representing a timestep (gradients or sources).
38 // The data is assumed to be of size1 in size (the strided dimension).
39 ~TransposedArray() override;
40 void WriteStrided(int t, const float *data) {
41 int size1 = dim1();
42 for (int i = 0; i < size1; ++i) {
43 put(i, t, data[i]);
44 }
45 }
46 void WriteStrided(int t, const double *data) {
47 int size1 = dim1();
48 for (int i = 0; i < size1; ++i) {
49 put(i, t, data[i]);
50 }
51 }
52 // Prints the first and last num elements of the un-transposed array.
53 void PrintUnTransposed(int num) {
54 int num_features = dim1();
55 int width = dim2();
56 for (int y = 0; y < num_features; ++y) {
57 for (int t = 0; t < width; ++t) {
58 if (num == 0 || t < num || t + num >= width) {
59 tprintf(" %g", static_cast<double>((*this)(y, t)));
60 }
61 }
62 tprintf("\n");
63 }
64 }
65 }; // class TransposedArray
66
67 // Generic weight matrix for network layers. Can store the matrix as either
68 // an array of floats or int8_t. Provides functions to compute the forward and
69 // backward steps with the matrix and updates to the weights.
70 class WeightMatrix {
71 public:
72 WeightMatrix() : int_mode_(false), use_adam_(false) {}
73 // Sets up the network for training. Initializes weights using weights of
74 // scale `range` picked according to the random number generator `randomizer`.
75 // Note the order is outputs, inputs, as this is the order of indices to
76 // the matrix, so the adjacent elements are multiplied by the input during
77 // a forward operation.
78 int InitWeightsFloat(int no, int ni, bool use_adam, float weight_range, TRand *randomizer);
79 // Changes the number of outputs to the size of the given code_map, copying
80 // the old weight matrix entries for each output from code_map[output] where
81 // non-negative, and uses the mean (over all outputs) of the existing weights
82 // for all outputs with negative code_map entries. Returns the new number of
83 // weights.
84 int RemapOutputs(const std::vector<int> &code_map);
85
86 // Converts a float network to an int network. Each set of input weights that
87 // corresponds to a single output weight is converted independently:
88 // Compute the max absolute value of the weight set.
89 // Scale so the max absolute value becomes INT8_MAX.
90 // Round to integer.
91 // Store a multiplicative scale factor (as a float) that will reproduce
92 // the original value, subject to rounding errors.
93 void ConvertToInt();
94 // Returns the size rounded up to an internal factor used by the SIMD
95 // implementation for its input.
96 int RoundInputs(int size) const {
97 if (!int_mode_ || !IntSimdMatrix::intSimdMatrix) {
98 return size;
99 }
100 return IntSimdMatrix::intSimdMatrix->RoundInputs(size);
101 }
102
103 // Accessors.
104 bool is_int_mode() const {
105 return int_mode_;
106 }
107 int NumOutputs() const {
108 return int_mode_ ? wi_.dim1() : wf_.dim1();
109 }
110 // Provides one set of weights. Only used by peep weight maxpool.
111 const TFloat *GetWeights(int index) const {
112 return wf_[index];
113 }
114 // Provides access to the deltas (dw_).
115 TFloat GetDW(int i, int j) const {
116 return dw_(i, j);
117 }
118
119 // Allocates any needed memory for running Backward, and zeroes the deltas,
120 // thus eliminating any existing momentum.
121 void InitBackward();
122
123 // Writes to the given file. Returns false in case of error.
124 bool Serialize(bool training, TFile *fp) const;
125 // Reads from the given file. Returns false in case of error.
126 bool DeSerialize(bool training, TFile *fp);
127 // As DeSerialize, but reads an old (float) format WeightMatrix for
128 // backward compatibility.
129 bool DeSerializeOld(bool training, TFile *fp);
130
131 // Computes matrix.vector v = Wu.
132 // u is of size W.dim2() - 1 and the output v is of size W.dim1().
133 // u is imagined to have an extra element at the end with value 1, to
134 // implement the bias, but it doesn't actually have it.
135 // Asserts that the call matches what we have.
136 void MatrixDotVector(const TFloat *u, TFloat *v) const;
137 void MatrixDotVector(const int8_t *u, TFloat *v) const;
138 // MatrixDotVector for peep weights, MultiplyAccumulate adds the
139 // component-wise products of *this[0] and v to inout.
140 void MultiplyAccumulate(const TFloat *v, TFloat *inout);
141 // Computes vector.matrix v = uW.
142 // u is of size W.dim1() and the output v is of size W.dim2() - 1.
143 // The last result is discarded, as v is assumed to have an imaginary
144 // last value of 1, as with MatrixDotVector.
145 void VectorDotMatrix(const TFloat *u, TFloat *v) const;
146 // Fills dw_[i][j] with the dot product u[i][] . v[j][], using elements
147 // from u and v, starting with u[i][offset] and v[j][offset].
148 // Note that (matching MatrixDotVector) v[last][] is missing, presumed 1.0.
149 // Runs parallel if requested. Note that inputs must be transposed.
150 void SumOuterTransposed(const TransposedArray &u, const TransposedArray &v, bool parallel);
151 // Updates the weights using the given learning rate, momentum and adam_beta.
152 // num_samples is used in the Adam correction factor.
153 void Update(float learning_rate, float momentum, float adam_beta, int num_samples);
154 // Adds the dw_ in other to the dw_ is *this.
155 void AddDeltas(const WeightMatrix &other);
156 // Sums the products of weight updates in *this and other, splitting into
157 // positive (same direction) in *same and negative (different direction) in
158 // *changed.
159 void CountAlternators(const WeightMatrix &other, TFloat *same, TFloat *changed) const;
160
161 void Debug2D(const char *msg);
162
163 private:
164 // Choice between float and 8 bit int implementations.
165 GENERIC_2D_ARRAY<TFloat> wf_;
166 GENERIC_2D_ARRAY<int8_t> wi_;
167 // Transposed copy of wf_, used only for Backward, and set with each Update.
168 TransposedArray wf_t_;
169 // Which of wf_ and wi_ are we actually using.
170 bool int_mode_;
171 // True if we are running adam in this weight matrix.
172 bool use_adam_;
173 // If we are using wi_, then scales_ is a factor to restore the row product
174 // with a vector to the correct range.
175 std::vector<TFloat> scales_;
176 // Weight deltas. dw_ is the new delta, and updates_ the momentum-decaying
177 // amount to be added to wf_/wi_.
178 GENERIC_2D_ARRAY<TFloat> dw_;
179 GENERIC_2D_ARRAY<TFloat> updates_;
180 // Iff use_adam_, the sum of squares of dw_. The number of samples is
181 // given to Update(). Serialized iff use_adam_.
182 GENERIC_2D_ARRAY<TFloat> dw_sq_sum_;
183 // The weights matrix reorganized in whatever way suits this instance.
184 std::vector<int8_t> shaped_w_;
185 };
186
187 } // namespace tesseract.
188
189 #endif // TESSERACT_LSTM_WEIGHTMATRIX_H_