/*---------------------------------------------------------------------------*/
int main (int argc, char **argv) {
/*
** Parameters:
** argc number of command line arguments
** argv array of command line arguments
** Globals: none
** Operation:
** This program reads in a text file consisting of feature
** samples from a training page in the following format:
**
** FontName UTF8-char-str xmin ymin xmax ymax page-number
** NumberOfFeatureTypes(N)
** FeatureTypeName1 NumberOfFeatures(M)
** Feature1
** ...
** FeatureM
** FeatureTypeName2 NumberOfFeatures(M)
** Feature1
** ...
** FeatureM
** ...
** FeatureTypeNameN NumberOfFeatures(M)
** Feature1
** ...
** FeatureM
** FontName CharName ...
**
** The result of this program is a binary inttemp file used by
** the OCR engine.
** Return: none
** Exceptions: none
** History: Fri Aug 18 08:56:17 1989, DSJ, Created.
** Mon May 18 1998, Christy Russson, Revistion started.
*/
ParseArguments(&argc, &argv);
ShapeTable* shape_table = NULL;
STRING file_prefix;
// Load the training data.
MasterTrainer* trainer = tesseract::LoadTrainingData(argc, argv,
false,
&shape_table,
&file_prefix);
if (trainer == NULL)
return 1; // Failed.
// Setup an index mapping from the shapes in the shape table to the classes
// that will be trained. In keeping with the original design, each shape
// with the same list of unichars becomes a different class and the configs
// represent the different combinations of fonts.
IndexMapBiDi config_map;
SetupConfigMap(shape_table, &config_map);
WriteShapeTable(file_prefix, *shape_table);
// If the shape_table is flat, then either we didn't run shape clustering, or
// it did nothing, so we just output the trainer's unicharset.
// Otherwise shape_set will hold a fake unicharset with an entry for each
// shape in the shape table, and we will output that instead.
UNICHARSET shape_set;
const UNICHARSET* unicharset = &trainer->unicharset();
// If we ran shapeclustering (and it worked) then at least one shape will
// have multiple unichars, so we have to build a fake unicharset.
if (shape_table->AnyMultipleUnichars()) {
unicharset = &shape_set;
// Now build a fake unicharset for the compact shape space to keep the
// output modules happy that we are doing things correctly.
int num_shapes = config_map.CompactSize();
for (int s = 0; s < num_shapes; ++s) {
char shape_label[kMaxShapeLabelLength + 1];
snprintf(shape_label, kMaxShapeLabelLength, "sh%04d", s);
shape_set.unichar_insert(shape_label);
}
}
// Now train each config separately.
int num_configs = shape_table->NumShapes();
LIST mf_classes = NIL_LIST;
for (int s = 0; s < num_configs; ++s) {
int unichar_id, font_id;
if (unicharset == &shape_set) {
// Using fake unichar_ids from the config_map/shape_set.
unichar_id = config_map.SparseToCompact(s);
} else {
// Get the real unichar_id from the shape table/unicharset.
shape_table->GetFirstUnicharAndFont(s, &unichar_id, &font_id);
}
const char* class_label = unicharset->id_to_unichar(unichar_id);
mf_classes = ClusterOneConfig(s, class_label, mf_classes, *shape_table,
trainer);
}
STRING inttemp_file = file_prefix;
inttemp_file += "inttemp";
STRING pffmtable_file = file_prefix;
pffmtable_file += "pffmtable";
CLASS_STRUCT* float_classes = SetUpForFloat2Int(*unicharset, mf_classes);
// Now write the inttemp and pffmtable.
trainer->WriteInttempAndPFFMTable(trainer->unicharset(), *unicharset,
*shape_table, float_classes,
inttemp_file.string(),
pffmtable_file.string());
delete [] float_classes;
FreeLabeledClassList(mf_classes);
delete trainer;
delete shape_table;
printf("Done!\n");
if (!FLAGS_test_ch.empty()) {
// If we are displaying debug window(s), wait for the user to look at them.
printf("Hit return to exit...\n");
while (getchar() != '\n');
}
return 0;
} /* main */
/**
* This program reads in a text file consisting of feature
* samples from a training page in the following format:
* @verbatim
FontName UTF8-char-str xmin ymin xmax ymax page-number
NumberOfFeatureTypes(N)
FeatureTypeName1 NumberOfFeatures(M)
Feature1
...
FeatureM
FeatureTypeName2 NumberOfFeatures(M)
Feature1
...
FeatureM
...
FeatureTypeNameN NumberOfFeatures(M)
Feature1
...
FeatureM
FontName CharName ...
@endverbatim
* The result of this program is a binary inttemp file used by
* the OCR engine.
* @param argc number of command line arguments
* @param argv array of command line arguments
* @return none
* @note Exceptions: none
* @note History: Fri Aug 18 08:56:17 1989, DSJ, Created.
* @note History: Mon May 18 1998, Christy Russson, Revistion started.
*/
int main (int argc, char **argv) {
ParseArguments(&argc, &argv);
ShapeTable* shape_table = NULL;
STRING file_prefix;
// Load the training data.
MasterTrainer* trainer = tesseract::LoadTrainingData(argc, argv,
false,
&shape_table,
&file_prefix);
if (trainer == NULL)
return 1; // Failed.
// Setup an index mapping from the shapes in the shape table to the classes
// that will be trained. In keeping with the original design, each shape
// with the same list of unichars becomes a different class and the configs
// represent the different combinations of fonts.
IndexMapBiDi config_map;
SetupConfigMap(shape_table, &config_map);
WriteShapeTable(file_prefix, *shape_table);
// If the shape_table is flat, then either we didn't run shape clustering, or
// it did nothing, so we just output the trainer's unicharset.
// Otherwise shape_set will hold a fake unicharset with an entry for each
// shape in the shape table, and we will output that instead.
UNICHARSET shape_set;
const UNICHARSET* unicharset = &trainer->unicharset();
// If we ran shapeclustering (and it worked) then at least one shape will
// have multiple unichars, so we have to build a fake unicharset.
if (shape_table->AnyMultipleUnichars()) {
unicharset = &shape_set;
// Now build a fake unicharset for the compact shape space to keep the
// output modules happy that we are doing things correctly.
int num_shapes = config_map.CompactSize();
for (int s = 0; s < num_shapes; ++s) {
char shape_label[kMaxShapeLabelLength + 1];
snprintf(shape_label, kMaxShapeLabelLength, "sh%04d", s);
shape_set.unichar_insert(shape_label);
}
}
// Now train each config separately.
int num_configs = shape_table->NumShapes();
LIST mf_classes = NIL_LIST;
for (int s = 0; s < num_configs; ++s) {
int unichar_id, font_id;
if (unicharset == &shape_set) {
// Using fake unichar_ids from the config_map/shape_set.
unichar_id = config_map.SparseToCompact(s);
} else {
// Get the real unichar_id from the shape table/unicharset.
shape_table->GetFirstUnicharAndFont(s, &unichar_id, &font_id);
}
const char* class_label = unicharset->id_to_unichar(unichar_id);
mf_classes = ClusterOneConfig(s, class_label, mf_classes, *shape_table,
trainer);
}
STRING inttemp_file = file_prefix;
inttemp_file += "inttemp";
STRING pffmtable_file = file_prefix;
pffmtable_file += "pffmtable";
CLASS_STRUCT* float_classes = SetUpForFloat2Int(*unicharset, mf_classes);
// Now write the inttemp and pffmtable.
trainer->WriteInttempAndPFFMTable(trainer->unicharset(), *unicharset,
*shape_table, float_classes,
inttemp_file.string(),
pffmtable_file.string());
for (int c = 0; c < unicharset->size(); ++c) {
FreeClassFields(&float_classes[c]);
}
delete [] float_classes;
FreeLabeledClassList(mf_classes);
delete trainer;
delete shape_table;
printf("Done!\n");
if (!FLAGS_test_ch.empty()) {
// If we are displaying debug window(s), wait for the user to look at them.
printf("Hit return to exit...\n");
while (getchar() != '\n');
}
return 0;
} /* main */
/**
* Creates a MasterTraininer and loads the training data into it:
* Initializes feature_defs and IntegerFX.
* Loads the shape_table if shape_table != nullptr.
* Loads initial unicharset from -U command-line option.
* If FLAGS_T is set, loads the majority of data from there, else:
* - Loads font info from -F option.
* - Loads xheights from -X option.
* - Loads samples from .tr files in remaining command-line args.
* - Deletes outliers and computes canonical samples.
* - If FLAGS_output_trainer is set, saves the trainer for future use.
* Computes canonical and cloud features.
* If shape_table is not nullptr, but failed to load, make a fake flat one,
* as shape clustering was not run.
*/
MasterTrainer* LoadTrainingData(int argc, const char* const * argv,
bool replication,
ShapeTable** shape_table,
STRING* file_prefix) {
InitFeatureDefs(&feature_defs);
InitIntegerFX();
*file_prefix = "";
if (!FLAGS_D.empty()) {
*file_prefix += FLAGS_D.c_str();
*file_prefix += "/";
}
// If we are shape clustering (nullptr shape_table) or we successfully load
// a shape_table written by a previous shape clustering, then
// shape_analysis will be true, meaning that the MasterTrainer will replace
// some members of the unicharset with their fragments.
bool shape_analysis = false;
if (shape_table != nullptr) {
*shape_table = LoadShapeTable(*file_prefix);
if (*shape_table != nullptr) shape_analysis = true;
} else {
shape_analysis = true;
}
MasterTrainer* trainer = new MasterTrainer(NM_CHAR_ANISOTROPIC,
shape_analysis,
replication,
FLAGS_debug_level);
IntFeatureSpace fs;
fs.Init(kBoostXYBuckets, kBoostXYBuckets, kBoostDirBuckets);
if (FLAGS_T.empty()) {
trainer->LoadUnicharset(FLAGS_U.c_str());
// Get basic font information from font_properties.
if (!FLAGS_F.empty()) {
if (!trainer->LoadFontInfo(FLAGS_F.c_str())) {
delete trainer;
return nullptr;
}
}
if (!FLAGS_X.empty()) {
if (!trainer->LoadXHeights(FLAGS_X.c_str())) {
delete trainer;
return nullptr;
}
}
trainer->SetFeatureSpace(fs);
const char* page_name;
// Load training data from .tr files on the command line.
while ((page_name = GetNextFilename(argc, argv)) != nullptr) {
tprintf("Reading %s ...\n", page_name);
trainer->ReadTrainingSamples(page_name, feature_defs, false);
// If there is a file with [lang].[fontname].exp[num].fontinfo present,
// read font spacing information in to fontinfo_table.
int pagename_len = strlen(page_name);
char *fontinfo_file_name = new char[pagename_len + 7];
strncpy(fontinfo_file_name, page_name, pagename_len - 2); // remove "tr"
strcpy(fontinfo_file_name + pagename_len - 2, "fontinfo"); // +"fontinfo"
trainer->AddSpacingInfo(fontinfo_file_name);
delete[] fontinfo_file_name;
// Load the images into memory if required by the classifier.
if (FLAGS_load_images) {
STRING image_name = page_name;
// Chop off the tr and replace with tif. Extension must be tif!
image_name.truncate_at(image_name.length() - 2);
image_name += "tif";
trainer->LoadPageImages(image_name.string());
}
}
trainer->PostLoadCleanup();
// Write the master trainer if required.
if (!FLAGS_output_trainer.empty()) {
FILE* fp = fopen(FLAGS_output_trainer.c_str(), "wb");
if (fp == nullptr) {
tprintf("Can't create saved trainer data!\n");
} else {
trainer->Serialize(fp);
fclose(fp);
}
}
} else {
bool success = false;
tprintf("Loading master trainer from file:%s\n",
FLAGS_T.c_str());
FILE* fp = fopen(FLAGS_T.c_str(), "rb");
if (fp == nullptr) {
tprintf("Can't read file %s to initialize master trainer\n",
FLAGS_T.c_str());
} else {
success = trainer->DeSerialize(false, fp);
fclose(fp);
}
if (!success) {
tprintf("Deserialize of master trainer failed!\n");
delete trainer;
return nullptr;
}
trainer->SetFeatureSpace(fs);
}
trainer->PreTrainingSetup();
if (!FLAGS_O.empty() &&
!trainer->unicharset().save_to_file(FLAGS_O.c_str())) {
fprintf(stderr, "Failed to save unicharset to file %s\n", FLAGS_O.c_str());
delete trainer;
return nullptr;
}
if (shape_table != nullptr) {
// If we previously failed to load a shapetable, then shape clustering
// wasn't run so make a flat one now.
if (*shape_table == nullptr) {
*shape_table = new ShapeTable;
trainer->SetupFlatShapeTable(*shape_table);
tprintf("Flat shape table summary: %s\n",
(*shape_table)->SummaryStr().string());
}
(*shape_table)->set_unicharset(trainer->unicharset());
}
return trainer;
}