Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/lstm/plumbing.cpp @ 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: plumbing.cpp
3 // Description: Base class for networks that organize other networks
4 // eg series or parallel.
5 // Author: Ray Smith
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.
17 ///////////////////////////////////////////////////////////////////////
18
19 #include "plumbing.h"
20
21 namespace tesseract {
22
23 // ni_ and no_ will be set by AddToStack.
24 Plumbing::Plumbing(const std::string &name) : Network(NT_PARALLEL, name, 0, 0) {}
25
26 // Suspends/Enables training by setting the training_ flag. Serialize and
27 // DeSerialize only operate on the run-time data if state is false.
28 void Plumbing::SetEnableTraining(TrainingState state) {
29 Network::SetEnableTraining(state);
30 for (auto &i : stack_) {
31 i->SetEnableTraining(state);
32 }
33 }
34
35 // Sets flags that control the action of the network. See NetworkFlags enum
36 // for bit values.
37 void Plumbing::SetNetworkFlags(uint32_t flags) {
38 Network::SetNetworkFlags(flags);
39 for (auto &i : stack_) {
40 i->SetNetworkFlags(flags);
41 }
42 }
43
44 // Sets up the network for training. Initializes weights using weights of
45 // scale `range` picked according to the random number generator `randomizer`.
46 // Note that randomizer is a borrowed pointer that should outlive the network
47 // and should not be deleted by any of the networks.
48 // Returns the number of weights initialized.
49 int Plumbing::InitWeights(float range, TRand *randomizer) {
50 num_weights_ = 0;
51 for (auto &i : stack_) {
52 num_weights_ += i->InitWeights(range, randomizer);
53 }
54 return num_weights_;
55 }
56
57 // Recursively searches the network for softmaxes with old_no outputs,
58 // and remaps their outputs according to code_map. See network.h for details.
59 int Plumbing::RemapOutputs(int old_no, const std::vector<int> &code_map) {
60 num_weights_ = 0;
61 for (auto &i : stack_) {
62 num_weights_ += i->RemapOutputs(old_no, code_map);
63 }
64 return num_weights_;
65 }
66
67 // Converts a float network to an int network.
68 void Plumbing::ConvertToInt() {
69 for (auto &i : stack_) {
70 i->ConvertToInt();
71 }
72 }
73
74 // Provides a pointer to a TRand for any networks that care to use it.
75 // Note that randomizer is a borrowed pointer that should outlive the network
76 // and should not be deleted by any of the networks.
77 void Plumbing::SetRandomizer(TRand *randomizer) {
78 for (auto &i : stack_) {
79 i->SetRandomizer(randomizer);
80 }
81 }
82
83 // Adds the given network to the stack.
84 void Plumbing::AddToStack(Network *network) {
85 if (stack_.empty()) {
86 ni_ = network->NumInputs();
87 no_ = network->NumOutputs();
88 } else if (type_ == NT_SERIES) {
89 // ni is input of first, no output of last, others match output to input.
90 ASSERT_HOST(no_ == network->NumInputs());
91 no_ = network->NumOutputs();
92 } else {
93 // All parallel types. Output is sum of outputs, inputs all match.
94 ASSERT_HOST(ni_ == network->NumInputs());
95 no_ += network->NumOutputs();
96 }
97 stack_.push_back(network);
98 }
99
100 // Sets needs_to_backprop_ to needs_backprop and calls on sub-network
101 // according to needs_backprop || any weights in this network.
102 bool Plumbing::SetupNeedsBackprop(bool needs_backprop) {
103 if (IsTraining()) {
104 needs_to_backprop_ = needs_backprop;
105 bool retval = needs_backprop;
106 for (auto &i : stack_) {
107 if (i->SetupNeedsBackprop(needs_backprop)) {
108 retval = true;
109 }
110 }
111 return retval;
112 }
113 // Frozen networks don't do backprop.
114 needs_to_backprop_ = false;
115 return false;
116 }
117
118 // Returns an integer reduction factor that the network applies to the
119 // time sequence. Assumes that any 2-d is already eliminated. Used for
120 // scaling bounding boxes of truth data.
121 // WARNING: if GlobalMinimax is used to vary the scale, this will return
122 // the last used scale factor. Call it before any forward, and it will return
123 // the minimum scale factor of the paths through the GlobalMinimax.
124 int Plumbing::XScaleFactor() const {
125 return stack_[0]->XScaleFactor();
126 }
127
128 // Provides the (minimum) x scale factor to the network (of interest only to
129 // input units) so they can determine how to scale bounding boxes.
130 void Plumbing::CacheXScaleFactor(int factor) {
131 for (auto &i : stack_) {
132 i->CacheXScaleFactor(factor);
133 }
134 }
135
136 // Provides debug output on the weights.
137 void Plumbing::DebugWeights() {
138 for (auto &i : stack_) {
139 i->DebugWeights();
140 }
141 }
142
143 // Returns a set of strings representing the layer-ids of all layers below.
144 void Plumbing::EnumerateLayers(const std::string *prefix, std::vector<std::string> &layers) const {
145 for (size_t i = 0; i < stack_.size(); ++i) {
146 std::string layer_name;
147 if (prefix) {
148 layer_name = *prefix;
149 }
150 layer_name += ":" + std::to_string(i);
151 if (stack_[i]->IsPlumbingType()) {
152 auto *plumbing = static_cast<Plumbing *>(stack_[i]);
153 plumbing->EnumerateLayers(&layer_name, layers);
154 } else {
155 layers.push_back(layer_name);
156 }
157 }
158 }
159
160 // Returns a pointer to the network layer corresponding to the given id.
161 Network *Plumbing::GetLayer(const char *id) const {
162 char *next_id;
163 int index = strtol(id, &next_id, 10);
164 if (index < 0 || static_cast<unsigned>(index) >= stack_.size()) {
165 return nullptr;
166 }
167 if (stack_[index]->IsPlumbingType()) {
168 auto *plumbing = static_cast<Plumbing *>(stack_[index]);
169 ASSERT_HOST(*next_id == ':');
170 return plumbing->GetLayer(next_id + 1);
171 }
172 return stack_[index];
173 }
174
175 // Returns a pointer to the learning rate for the given layer id.
176 float *Plumbing::LayerLearningRatePtr(const char *id) {
177 char *next_id;
178 int index = strtol(id, &next_id, 10);
179 if (index < 0 || static_cast<unsigned>(index) >= stack_.size()) {
180 return nullptr;
181 }
182 if (stack_[index]->IsPlumbingType()) {
183 auto *plumbing = static_cast<Plumbing *>(stack_[index]);
184 ASSERT_HOST(*next_id == ':');
185 return plumbing->LayerLearningRatePtr(next_id + 1);
186 }
187 if (static_cast<unsigned>(index) >= learning_rates_.size()) {
188 return nullptr;
189 }
190 return &learning_rates_[index];
191 }
192
193 // Writes to the given file. Returns false in case of error.
194 bool Plumbing::Serialize(TFile *fp) const {
195 if (!Network::Serialize(fp)) {
196 return false;
197 }
198 uint32_t size = stack_.size();
199 // Can't use PointerVector::Serialize here as we need a special DeSerialize.
200 if (!fp->Serialize(&size)) {
201 return false;
202 }
203 for (uint32_t i = 0; i < size; ++i) {
204 if (!stack_[i]->Serialize(fp)) {
205 return false;
206 }
207 }
208 if ((network_flags_ & NF_LAYER_SPECIFIC_LR) && !fp->Serialize(learning_rates_)) {
209 return false;
210 }
211 return true;
212 }
213
214 // Reads from the given file. Returns false in case of error.
215 bool Plumbing::DeSerialize(TFile *fp) {
216 for (auto data : stack_) {
217 delete data;
218 }
219 stack_.clear();
220 no_ = 0; // We will be modifying this as we AddToStack.
221 uint32_t size;
222 if (!fp->DeSerialize(&size)) {
223 return false;
224 }
225 for (uint32_t i = 0; i < size; ++i) {
226 Network *network = CreateFromFile(fp);
227 if (network == nullptr) {
228 return false;
229 }
230 AddToStack(network);
231 }
232 if ((network_flags_ & NF_LAYER_SPECIFIC_LR) && !fp->DeSerialize(learning_rates_)) {
233 return false;
234 }
235 return true;
236 }
237
238 // Updates the weights using the given learning rate, momentum and adam_beta.
239 // num_samples is used in the adam computation iff use_adam_ is true.
240 void Plumbing::Update(float learning_rate, float momentum, float adam_beta, int num_samples) {
241 for (size_t i = 0; i < stack_.size(); ++i) {
242 if (network_flags_ & NF_LAYER_SPECIFIC_LR) {
243 if (i < learning_rates_.size()) {
244 learning_rate = learning_rates_[i];
245 } else {
246 learning_rates_.push_back(learning_rate);
247 }
248 }
249 if (stack_[i]->IsTraining()) {
250 stack_[i]->Update(learning_rate, momentum, adam_beta, num_samples);
251 }
252 }
253 }
254
255 // Sums the products of weight updates in *this and other, splitting into
256 // positive (same direction) in *same and negative (different direction) in
257 // *changed.
258 void Plumbing::CountAlternators(const Network &other, TFloat *same, TFloat *changed) const {
259 ASSERT_HOST(other.type() == type_);
260 const auto *plumbing = static_cast<const Plumbing *>(&other);
261 ASSERT_HOST(plumbing->stack_.size() == stack_.size());
262 for (size_t i = 0; i < stack_.size(); ++i) {
263 stack_[i]->CountAlternators(*plumbing->stack_[i], same, changed);
264 }
265 }
266
267 } // namespace tesseract.