// Moves this to the target unicity table.
void FontInfoTable::MoveTo(UnicityTable<FontInfo>* target) {
target->clear();
target->set_compare_callback(NewPermanentTessCallback(CompareFontInfo));
target->set_clear_callback(NewPermanentTessCallback(FontInfoDeleteCallback));
for (int i = 0; i < size(); ++i) {
// Bit copy the FontInfo and steal all the pointers.
target->push_back(get(i));
get(i).name = NULL;
get(i).spacing_vec = NULL;
}
}
bool UNICHARSET::load_from_file(tesseract::TFile *file, bool skip_fragments) {
TessResultCallback2<char *, char *, int> *fgets_cb =
NewPermanentTessCallback(file, &tesseract::TFile::FGets);
bool success = load_via_fgets(fgets_cb, skip_fragments);
delete fgets_cb;
return success;
}
void Dawg::iterate_words(const UNICHARSET &unicharset,
TessCallback1<const char *> *cb) const {
std::unique_ptr<TessCallback1<const WERD_CHOICE *>> shim(
NewPermanentTessCallback(CallWithUTF8, cb));
WERD_CHOICE word(&unicharset);
iterate_words_rec(word, 0, shim.get());
}
bool UNICHARSET::load_from_file(FILE *file, bool skip_fragments) {
LocalFilePointer lfp(file);
TessResultCallback2<char *, char *, int> *fgets_cb =
NewPermanentTessCallback(&lfp, &LocalFilePointer::fgets);
bool success = load_via_fgets(fgets_cb, skip_fragments);
delete fgets_cb;
return success;
}
bool UNICHARSET::load_from_inmemory_file(
const char *memory, int mem_size, bool skip_fragments) {
InMemoryFilePointer mem_fp(memory, mem_size);
TessResultCallback2<char *, char *, int> *fgets_cb =
NewPermanentTessCallback(&mem_fp, &InMemoryFilePointer::fgets);
bool success = load_via_fgets(fgets_cb, skip_fragments);
delete fgets_cb;
return success;
}
// Moves any non-empty FontSpacingInfo entries from other to this.
void FontInfoTable::MoveSpacingInfoFrom(FontInfoTable* other) {
set_compare_callback(NewPermanentTessCallback(CompareFontInfo));
set_clear_callback(NewPermanentTessCallback(FontInfoDeleteCallback));
for (int i = 0; i < other->size(); ++i) {
GenericVector<FontSpacingInfo*>* spacing_vec = other->get(i).spacing_vec;
if (spacing_vec != NULL) {
int target_index = get_index(other->get(i));
if (target_index < 0) {
// Bit copy the FontInfo and steal all the pointers.
push_back(other->get(i));
other->get(i).name = NULL;
} else {
delete [] get(target_index).spacing_vec;
get(target_index).spacing_vec = other->get(i).spacing_vec;
}
other->get(i).spacing_vec = NULL;
}
}
}
void Wordrec::InitBlamerForSegSearch(WERD_RES *word_res,
LMPainPoints *pain_points,
BlamerBundle *blamer_bundle,
STRING *blamer_debug) {
pain_points->Clear(); // Clear pain points heap.
TessResultCallback2<bool, int, int>* pp_cb = NewPermanentTessCallback(
pain_points, &LMPainPoints::GenerateForBlamer,
static_cast<double>(segsearch_max_char_wh_ratio), word_res);
blamer_bundle->InitForSegSearch(word_res->best_choice, word_res->ratings,
getDict().WildcardID(), wordrec_debug_blamer,
blamer_debug, pp_cb);
delete pp_cb;
}
// returns 0 if successful.
int WriteDawgAsWordlist(const UNICHARSET &unicharset,
const tesseract::Dawg *dawg,
const char *outfile_name) {
FILE *out = fopen(outfile_name, "wb");
if (out == NULL) {
tprintf("Could not open %s for writing.\n", outfile_name);
return 1;
}
WordOutputter outputter(out);
TessCallback1<const char *> *print_word_cb =
NewPermanentTessCallback(&outputter, &WordOutputter::output_word);
dawg->iterate_words(unicharset, print_word_cb);
delete print_word_cb;
return fclose(out);
}
// Set the universal_id member of each font to be unique among all
// instances of the same font loaded.
void Tesseract::SetupUniversalFontIds() {
// Note that we can get away with bitwise copying FontInfo in
// all_fonts, as it is a temporary structure and we avoid setting the
// delete callback.
UnicityTable<FontInfo> all_fonts;
all_fonts.set_compare_callback(NewPermanentTessCallback(CompareFontInfo));
// Create the universal ID table.
CollectFonts(get_fontinfo_table(), &all_fonts);
for (int i = 0; i < sub_langs_.size(); ++i) {
CollectFonts(sub_langs_[i]->get_fontinfo_table(), &all_fonts);
}
// Assign ids from the table to each font table.
AssignIds(all_fonts, &get_fontinfo_table());
for (int i = 0; i < sub_langs_.size(); ++i) {
AssignIds(all_fonts, &sub_langs_[i]->get_fontinfo_table());
}
font_table_size_ = all_fonts.size();
}
FontInfoTable::FontInfoTable() {
set_compare_callback(NewPermanentTessCallback(CompareFontInfo));
set_clear_callback(NewPermanentTessCallback(FontInfoDeleteCallback));
}
Classify::Classify()
: BOOL_MEMBER(prioritize_division, FALSE,
"Prioritize blob division over chopping", this->params()),
INT_MEMBER(tessedit_single_match, FALSE,
"Top choice only from CP", this->params()),
BOOL_MEMBER(classify_enable_learning, true,
"Enable adaptive classifier", this->params()),
INT_MEMBER(classify_debug_level, 0, "Classify debug level",
this->params()),
INT_MEMBER(classify_norm_method, character, "Normalization Method ...",
this->params()),
double_MEMBER(classify_char_norm_range, 0.2,
"Character Normalization Range ...", this->params()),
double_MEMBER(classify_min_norm_scale_x, 0.0, "Min char x-norm scale ...",
this->params()), /* PREV DEFAULT 0.1 */
double_MEMBER(classify_max_norm_scale_x, 0.325, "Max char x-norm scale ...",
this->params()), /* PREV DEFAULT 0.3 */
double_MEMBER(classify_min_norm_scale_y, 0.0, "Min char y-norm scale ...",
this->params()), /* PREV DEFAULT 0.1 */
double_MEMBER(classify_max_norm_scale_y, 0.325, "Max char y-norm scale ...",
this->params()), /* PREV DEFAULT 0.3 */
double_MEMBER(classify_max_rating_ratio, 1.5,
"Veto ratio between classifier ratings", this->params()),
double_MEMBER(classify_max_certainty_margin, 5.5,
"Veto difference between classifier certainties",
this->params()),
BOOL_MEMBER(tess_cn_matching, 0, "Character Normalized Matching",
this->params()),
BOOL_MEMBER(tess_bn_matching, 0, "Baseline Normalized Matching",
this->params()),
BOOL_MEMBER(classify_enable_adaptive_matcher, 1,
"Enable adaptive classifier",
this->params()),
BOOL_MEMBER(classify_use_pre_adapted_templates, 0,
"Use pre-adapted classifier templates", this->params()),
BOOL_MEMBER(classify_save_adapted_templates, 0,
"Save adapted templates to a file", this->params()),
BOOL_MEMBER(classify_enable_adaptive_debugger, 0, "Enable match debugger",
this->params()),
BOOL_MEMBER(classify_nonlinear_norm, 0,
"Non-linear stroke-density normalization", this->params()),
INT_MEMBER(matcher_debug_level, 0, "Matcher Debug Level", this->params()),
INT_MEMBER(matcher_debug_flags, 0, "Matcher Debug Flags", this->params()),
INT_MEMBER(classify_learning_debug_level, 0, "Learning Debug Level: ",
this->params()),
double_MEMBER(matcher_good_threshold, 0.125, "Good Match (0-1)",
this->params()),
double_MEMBER(matcher_great_threshold, 0.0, "Great Match (0-1)",
this->params()),
double_MEMBER(matcher_perfect_threshold, 0.02, "Perfect Match (0-1)",
this->params()),
double_MEMBER(matcher_bad_match_pad, 0.15, "Bad Match Pad (0-1)",
this->params()),
double_MEMBER(matcher_rating_margin, 0.1, "New template margin (0-1)",
this->params()),
double_MEMBER(matcher_avg_noise_size, 12.0, "Avg. noise blob length",
this->params()),
INT_MEMBER(matcher_permanent_classes_min, 1, "Min # of permanent classes",
this->params()),
INT_MEMBER(matcher_min_examples_for_prototyping, 3,
"Reliable Config Threshold", this->params()),
INT_MEMBER(matcher_sufficient_examples_for_prototyping, 5,
"Enable adaption even if the ambiguities have not been seen",
this->params()),
double_MEMBER(matcher_clustering_max_angle_delta, 0.015,
"Maximum angle delta for prototype clustering",
this->params()),
double_MEMBER(classify_misfit_junk_penalty, 0.0,
"Penalty to apply when a non-alnum is vertically out of "
"its expected textline position",
this->params()),
double_MEMBER(rating_scale, 1.5, "Rating scaling factor", this->params()),
double_MEMBER(certainty_scale, 20.0, "Certainty scaling factor",
this->params()),
double_MEMBER(tessedit_class_miss_scale, 0.00390625,
"Scale factor for features not used", this->params()),
double_MEMBER(classify_adapted_pruning_factor, 2.5,
"Prune poor adapted results this much worse than best result",
this->params()),
double_MEMBER(classify_adapted_pruning_threshold, -1.0,
"Threshold at which classify_adapted_pruning_factor starts",
this->params()),
INT_MEMBER(classify_adapt_proto_threshold, 230,
"Threshold for good protos during adaptive 0-255",
this->params()),
INT_MEMBER(classify_adapt_feature_threshold, 230,
"Threshold for good features during adaptive 0-255",
this->params()),
BOOL_MEMBER(disable_character_fragments, TRUE,
"Do not include character fragments in the"
" results of the classifier", this->params()),
double_MEMBER(classify_character_fragments_garbage_certainty_threshold,
-3.0, "Exclude fragments that do not look like whole"
" characters from training and adaption", this->params()),
BOOL_MEMBER(classify_debug_character_fragments, FALSE,
"Bring up graphical debugging windows for fragments training",
this->params()),
BOOL_MEMBER(matcher_debug_separate_windows, FALSE,
"Use two different windows for debugging the matching: "
"One for the protos and one for the features.", this->params()),
STRING_MEMBER(classify_learn_debug_str, "", "Class str to debug learning",
this->params()),
INT_MEMBER(classify_class_pruner_threshold, 229,
"Class Pruner Threshold 0-255", this->params()),
INT_MEMBER(classify_class_pruner_multiplier, 15,
"Class Pruner Multiplier 0-255: ", this->params()),
INT_MEMBER(classify_cp_cutoff_strength, 7,
"Class Pruner CutoffStrength: ", this->params()),
INT_MEMBER(classify_integer_matcher_multiplier, 10,
"Integer Matcher Multiplier 0-255: ", this->params()),
EnableLearning(true),
INT_MEMBER(il1_adaption_test, 0, "Dont adapt to i/I at beginning of word",
this->params()),
BOOL_MEMBER(classify_bln_numeric_mode, 0,
"Assume the input is numbers [0-9].", this->params()),
double_MEMBER(speckle_large_max_size, 0.30, "Max large speckle size",
this->params()),
double_MEMBER(speckle_rating_penalty, 10.0,
"Penalty to add to worst rating for noise", this->params()),
shape_table_(NULL),
dict_(this),
static_classifier_(NULL) {
fontinfo_table_.set_compare_callback(
NewPermanentTessCallback(CompareFontInfo));
fontinfo_table_.set_clear_callback(
NewPermanentTessCallback(FontInfoDeleteCallback));
fontset_table_.set_compare_callback(
NewPermanentTessCallback(CompareFontSet));
fontset_table_.set_clear_callback(
NewPermanentTessCallback(FontSetDeleteCallback));
AdaptedTemplates = NULL;
PreTrainedTemplates = NULL;
AllProtosOn = NULL;
AllConfigsOn = NULL;
AllConfigsOff = NULL;
TempProtoMask = NULL;
NormProtos = NULL;
NumAdaptationsFailed = 0;
learn_debug_win_ = NULL;
learn_fragmented_word_debug_win_ = NULL;
learn_fragments_debug_win_ = NULL;
CharNormCutoffs = new uinT16[MAX_NUM_CLASSES];
BaselineCutoffs = new uinT16[MAX_NUM_CLASSES];
}
// Deletes all samples with zero features marked by KillSample.
void TrainingSampleSet::DeleteDeadSamples() {
samples_.compact(
NewPermanentTessCallback(this, &TrainingSampleSet::DeleteableSample));
num_raw_samples_ = samples_.size();
// Samples must be re-organized now we have deleted a few.
}
// Apart from command-line flags, input is a collection of lstmf files, that
// were previously created using tesseract with the lstm.train config file.
// The program iterates over the inputs, feeding the data to the network,
// until the error rate reaches a specified target or max_iterations is reached.
int main(int argc, char **argv) {
ParseArguments(&argc, &argv);
// Purify the model name in case it is based on the network string.
if (FLAGS_model_output.empty()) {
tprintf("Must provide a --model_output!\n");
return 1;
}
STRING model_output = FLAGS_model_output.c_str();
for (int i = 0; i < model_output.length(); ++i) {
if (model_output[i] == '[' || model_output[i] == ']')
model_output[i] = '-';
if (model_output[i] == '(' || model_output[i] == ')')
model_output[i] = '_';
}
// Setup the trainer.
STRING checkpoint_file = FLAGS_model_output.c_str();
checkpoint_file += "_checkpoint";
STRING checkpoint_bak = checkpoint_file + ".bak";
tesseract::LSTMTrainer trainer(
NULL, NULL, NULL, NULL, FLAGS_model_output.c_str(),
checkpoint_file.c_str(), FLAGS_debug_interval,
static_cast<inT64>(FLAGS_max_image_MB) * 1048576);
// Reading something from an existing model doesn't require many flags,
// so do it now and exit.
if (FLAGS_stop_training || FLAGS_debug_network) {
if (!trainer.TryLoadingCheckpoint(FLAGS_continue_from.c_str())) {
tprintf("Failed to read continue from: %s\n",
FLAGS_continue_from.c_str());
return 1;
}
if (FLAGS_debug_network) {
trainer.DebugNetwork();
} else {
if (FLAGS_train_mode & tesseract::TF_INT_MODE)
trainer.ConvertToInt();
GenericVector<char> recognizer_data;
trainer.SaveRecognitionDump(&recognizer_data);
if (!tesseract::SaveDataToFile(recognizer_data,
FLAGS_model_output.c_str())) {
tprintf("Failed to write recognition model : %s\n",
FLAGS_model_output.c_str());
}
}
return 0;
}
// Get the list of files to process.
if (FLAGS_train_listfile.empty()) {
tprintf("Must supply a list of training filenames! --train_listfile\n");
return 1;
}
GenericVector<STRING> filenames;
if (!tesseract::LoadFileLinesToStrings(FLAGS_train_listfile.c_str(),
&filenames)) {
tprintf("Failed to load list of training filenames from %s\n",
FLAGS_train_listfile.c_str());
return 1;
}
UNICHARSET unicharset;
// Checkpoints always take priority if they are available.
if (trainer.TryLoadingCheckpoint(checkpoint_file.string()) ||
trainer.TryLoadingCheckpoint(checkpoint_bak.string())) {
tprintf("Successfully restored trainer from %s\n",
checkpoint_file.string());
} else {
if (!FLAGS_continue_from.empty()) {
// Load a past model file to improve upon.
if (!trainer.TryLoadingCheckpoint(FLAGS_continue_from.c_str())) {
tprintf("Failed to continue from: %s\n", FLAGS_continue_from.c_str());
return 1;
}
tprintf("Continuing from %s\n", FLAGS_continue_from.c_str());
trainer.InitIterations();
}
if (FLAGS_continue_from.empty() || FLAGS_append_index >= 0) {
// We need a unicharset to start from scratch or append.
string unicharset_str;
// Character coding to be used by the classifier.
if (!unicharset.load_from_file(FLAGS_U.c_str())) {
tprintf("Error: must provide a -U unicharset!\n");
return 1;
}
tesseract::SetupBasicProperties(true, &unicharset);
if (FLAGS_append_index >= 0) {
tprintf("Appending a new network to an old one!!");
if (FLAGS_continue_from.empty()) {
tprintf("Must set --continue_from for appending!\n");
return 1;
}
}
// We are initializing from scratch.
trainer.InitCharSet(unicharset, FLAGS_script_dir.c_str(),
FLAGS_train_mode);
if (!trainer.InitNetwork(FLAGS_net_spec.c_str(), FLAGS_append_index,
FLAGS_net_mode, FLAGS_weight_range,
FLAGS_learning_rate, FLAGS_momentum)) {
tprintf("Failed to create network from spec: %s\n",
FLAGS_net_spec.c_str());
return 1;
}
trainer.set_perfect_delay(FLAGS_perfect_sample_delay);
}
}
if (!trainer.LoadAllTrainingData(filenames)) {
tprintf("Load of images failed!!\n");
return 1;
}
bool best_dumped = true;
char* best_model_dump = NULL;
size_t best_model_size = 0;
STRING best_model_name;
tesseract::LSTMTester tester(static_cast<inT64>(FLAGS_max_image_MB) *
1048576);
tesseract::TestCallback tester_callback = nullptr;
if (!FLAGS_eval_listfile.empty()) {
if (!tester.LoadAllEvalData(FLAGS_eval_listfile.c_str())) {
tprintf("Failed to load eval data from: %s\n",
FLAGS_eval_listfile.c_str());
return 1;
}
tester_callback =
NewPermanentTessCallback(&tester, &tesseract::LSTMTester::RunEvalAsync);
}
do {
// Train a few.
int iteration = trainer.training_iteration();
for (int target_iteration = iteration + kNumPagesPerBatch;
iteration < target_iteration;
iteration = trainer.training_iteration()) {
trainer.TrainOnLine(&trainer, false);
}
STRING log_str;
trainer.MaintainCheckpoints(tester_callback, &log_str);
tprintf("%s\n", log_str.string());
} while (trainer.best_error_rate() > FLAGS_target_error_rate &&
(trainer.training_iteration() < FLAGS_max_iterations ||
FLAGS_max_iterations == 0));
delete tester_callback;
tprintf("Finished! Error rate = %g\n", trainer.best_error_rate());
return 0;
} /* main */