Tesseract  3.02
tesseract-ocr/classify/mastertrainer.cpp
Go to the documentation of this file.
00001 // Copyright 2010 Google Inc. All Rights Reserved.
00002 // Author: rays@google.com (Ray Smith)
00004 // File:        mastertrainer.cpp
00005 // Description: Trainer to build the MasterClassifier.
00006 // Author:      Ray Smith
00007 // Created:     Wed Nov 03 18:10:01 PDT 2010
00008 //
00009 // (C) Copyright 2010, Google Inc.
00010 // Licensed under the Apache License, Version 2.0 (the "License");
00011 // you may not use this file except in compliance with the License.
00012 // You may obtain a copy of the License at
00013 // http://www.apache.org/licenses/LICENSE-2.0
00014 // Unless required by applicable law or agreed to in writing, software
00015 // distributed under the License is distributed on an "AS IS" BASIS,
00016 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
00017 // See the License for the specific language governing permissions and
00018 // limitations under the License.
00019 //
00021 
00022 #include "mastertrainer.h"
00023 #include <math.h>
00024 #include <time.h>
00025 #include "allheaders.h"
00026 #include "boxread.h"
00027 #include "classify.h"
00028 #include "errorcounter.h"
00029 #include "featdefs.h"
00030 #include "sampleiterator.h"
00031 #include "shapeclassifier.h"
00032 #include "shapetable.h"
00033 #include "svmnode.h"
00034 
00035 namespace tesseract {
00036 
00037 // Constants controlling clustering. With a low kMinClusteredShapes and a high
00038 // kMaxUnicharsPerCluster, then kFontMergeDistance is the only limiting factor.
00039 // Min number of shapes in the output.
00040 const int kMinClusteredShapes = 1;
00041 // Max number of unichars in any individual cluster.
00042 const int kMaxUnicharsPerCluster = 2000;
00043 // Mean font distance below which to merge fonts and unichars.
00044 const float kFontMergeDistance = 0.025;
00045 
00046 MasterTrainer::MasterTrainer(NormalizationMode norm_mode,
00047                              bool shape_analysis,
00048                              bool replicate_samples,
00049                              int debug_level)
00050   : norm_mode_(norm_mode), samples_(fontinfo_table_),
00051     junk_samples_(fontinfo_table_), verify_samples_(fontinfo_table_),
00052     charsetsize_(0),
00053     enable_shape_anaylsis_(shape_analysis),
00054     enable_replication_(replicate_samples),
00055     fragments_(NULL), prev_unichar_id_(-1), debug_level_(debug_level) {
00056   fontinfo_table_.set_compare_callback(
00057       NewPermanentTessCallback(CompareFontInfo));
00058   fontinfo_table_.set_clear_callback(
00059       NewPermanentTessCallback(FontInfoDeleteCallback));
00060 }
00061 
00062 MasterTrainer::~MasterTrainer() {
00063   delete [] fragments_;
00064   for (int p = 0; p < page_images_.size(); ++p)
00065     pixDestroy(&page_images_[p]);
00066 }
00067 
00068 // WARNING! Serialize/DeSerialize are only partial, providing
00069 // enough data to get the samples back and display them.
00070 // Writes to the given file. Returns false in case of error.
00071 bool MasterTrainer::Serialize(FILE* fp) const {
00072   if (fwrite(&norm_mode_, sizeof(norm_mode_), 1, fp) != 1) return false;
00073   if (!unicharset_.save_to_file(fp)) return false;
00074   if (!feature_space_.Serialize(fp)) return false;
00075   if (!samples_.Serialize(fp)) return false;
00076   if (!junk_samples_.Serialize(fp)) return false;
00077   if (!verify_samples_.Serialize(fp)) return false;
00078   if (!master_shapes_.Serialize(fp)) return false;
00079   if (!flat_shapes_.Serialize(fp)) return false;
00080   if (!fontinfo_table_.write(fp, NewPermanentTessCallback(write_info)))
00081     return false;
00082   if (!fontinfo_table_.write(fp, NewPermanentTessCallback(write_spacing_info)))
00083     return false;
00084   if (!xheights_.Serialize(fp)) return false;
00085   return true;
00086 }
00087 
00088 // Reads from the given file. Returns false in case of error.
00089 // If swap is true, assumes a big/little-endian swap is needed.
00090 bool MasterTrainer::DeSerialize(bool swap, FILE* fp) {
00091   if (fread(&norm_mode_, sizeof(norm_mode_), 1, fp) != 1) return false;
00092   if (swap) {
00093     ReverseN(&norm_mode_, sizeof(norm_mode_));
00094   }
00095   if (!unicharset_.load_from_file(fp)) return false;
00096   charsetsize_ = unicharset_.size();
00097   if (!feature_space_.DeSerialize(swap, fp)) return false;
00098   feature_map_.Init(feature_space_);
00099   if (!samples_.DeSerialize(swap, fp)) return false;
00100   if (!junk_samples_.DeSerialize(swap, fp)) return false;
00101   if (!verify_samples_.DeSerialize(swap, fp)) return false;
00102   if (!master_shapes_.DeSerialize(swap, fp)) return false;
00103   if (!flat_shapes_.DeSerialize(swap, fp)) return false;
00104   if (!fontinfo_table_.read(fp, NewPermanentTessCallback(read_info), swap))
00105     return false;
00106   if (!fontinfo_table_.read(fp, NewPermanentTessCallback(read_spacing_info),
00107                             swap))
00108     return false;
00109   if (!xheights_.DeSerialize(swap, fp)) return false;
00110   return true;
00111 }
00112 
00113 // Load an initial unicharset, or set one up if the file cannot be read.
00114 void MasterTrainer::LoadUnicharset(const char* filename) {
00115   if (!unicharset_.load_from_file(filename)) {
00116     tprintf("Failed to load unicharset from file %s\n"
00117             "Building unicharset for training from scratch...\n",
00118             filename);
00119     unicharset_.clear();
00120     // Space character needed to represent NIL_LIST classification.
00121     unicharset_.unichar_insert(" ");
00122   }
00123   charsetsize_ = unicharset_.size();
00124   delete [] fragments_;
00125   fragments_ = new int[charsetsize_];
00126   memset(fragments_, 0, sizeof(*fragments_) * charsetsize_);
00127   samples_.LoadUnicharset(filename);
00128   junk_samples_.LoadUnicharset(filename);
00129   verify_samples_.LoadUnicharset(filename);
00130 }
00131 
00132 // Reads the samples and their features from the given .tr format file,
00133 // adding them to the trainer with the font_id from the content of the file.
00134 // See mftraining.cpp for a description of the file format.
00135 // If verification, then these are verification samples, not training.
00136 void MasterTrainer::ReadTrainingSamples(FILE  *fp,
00137                                         const FEATURE_DEFS_STRUCT& feature_defs,
00138                                         bool verification) {
00139   char buffer[2048];
00140   int int_feature_type = ShortNameToFeatureType(feature_defs, kIntFeatureType);
00141   int micro_feature_type = ShortNameToFeatureType(feature_defs,
00142                                                   kMicroFeatureType);
00143   int cn_feature_type = ShortNameToFeatureType(feature_defs, kCNFeatureType);
00144   int geo_feature_type = ShortNameToFeatureType(feature_defs, kGeoFeatureType);
00145 
00146   while (fgets(buffer, sizeof(buffer), fp) != NULL) {
00147     if (buffer[0] == '\n')
00148       continue;
00149 
00150     char* space = strchr(buffer, ' ');
00151     if (space == NULL) {
00152       tprintf("Bad format in tr file, reading fontname, unichar\n");
00153       continue;
00154     }
00155     *space++ = '\0';
00156     int font_id = GetFontInfoId(buffer);
00157     int page_number;
00158     STRING unichar;
00159     TBOX bounding_box;
00160     if (!ParseBoxFileStr(space, &page_number, &unichar, &bounding_box)) {
00161       tprintf("Bad format in tr file, reading box coords\n");
00162       continue;
00163     }
00164     CHAR_DESC char_desc = ReadCharDescription(feature_defs, fp);
00165     TrainingSample* sample = new TrainingSample;
00166     sample->set_font_id(font_id);
00167     sample->set_page_num(page_number + page_images_.size());
00168     sample->set_bounding_box(bounding_box);
00169     sample->ExtractCharDesc(int_feature_type, micro_feature_type,
00170                             cn_feature_type, geo_feature_type, char_desc);
00171     AddSample(verification, unichar.string(), sample);
00172     FreeCharDescription(char_desc);
00173   }
00174   charsetsize_ = unicharset_.size();
00175 }
00176 
00177 // Adds the given single sample to the trainer, setting the classid
00178 // appropriately from the given unichar_str.
00179 void MasterTrainer::AddSample(bool verification, const char* unichar,
00180                               TrainingSample* sample) {
00181   if (verification) {
00182     verify_samples_.AddSample(unichar, sample);
00183     prev_unichar_id_ = -1;
00184   } else if (unicharset_.contains_unichar(unichar)) {
00185     if (prev_unichar_id_ >= 0)
00186       fragments_[prev_unichar_id_] = -1;
00187     prev_unichar_id_ = samples_.AddSample(unichar, sample);
00188     if (flat_shapes_.FindShape(prev_unichar_id_, sample->font_id()) < 0)
00189       flat_shapes_.AddShape(prev_unichar_id_, sample->font_id());
00190   } else {
00191     int junk_id = junk_samples_.AddSample(unichar, sample);
00192     if (prev_unichar_id_ >= 0) {
00193       CHAR_FRAGMENT* frag = CHAR_FRAGMENT::parse_from_string(unichar);
00194       if (frag != NULL && frag->is_natural()) {
00195         if (fragments_[prev_unichar_id_] == 0)
00196           fragments_[prev_unichar_id_] = junk_id;
00197         else if (fragments_[prev_unichar_id_] != junk_id)
00198           fragments_[prev_unichar_id_] = -1;
00199       }
00200       delete frag;
00201     }
00202     prev_unichar_id_ = -1;
00203   }
00204 }
00205 
00206 // Loads all pages from the given tif filename and append to page_images_.
00207 // Must be called after ReadTrainingSamples, as the current number of images
00208 // is used as an offset for page numbers in the samples.
00209 void MasterTrainer::LoadPageImages(const char* filename) {
00210   int page;
00211   Pix* pix;
00212   for (page = 0; (pix = pixReadTiff(filename, page)) != NULL; ++page) {
00213     page_images_.push_back(pix);
00214   }
00215   tprintf("Loaded %d page images from %s\n", page, filename);
00216 }
00217 
00218 // Cleans up the samples after initial load from the tr files, and prior to
00219 // saving the MasterTrainer:
00220 // Remaps fragmented chars if running shape anaylsis.
00221 // Sets up the samples appropriately for class/fontwise access.
00222 // Deletes outlier samples.
00223 void MasterTrainer::PostLoadCleanup() {
00224   if (debug_level_ > 0)
00225     tprintf("PostLoadCleanup...\n");
00226   if (enable_shape_anaylsis_)
00227     ReplaceFragmentedSamples();
00228   SampleIterator sample_it;
00229   sample_it.Init(NULL, NULL, true, &verify_samples_);
00230   sample_it.NormalizeSamples();
00231   verify_samples_.OrganizeByFontAndClass();
00232 
00233   samples_.IndexFeatures(feature_space_);
00234   // TODO(rays) DeleteOutliers is currently turned off to prove NOP-ness
00235   // against current training.
00236   //  samples_.DeleteOutliers(feature_space_, debug_level_ > 0);
00237   samples_.OrganizeByFontAndClass();
00238   if (debug_level_ > 0)
00239     tprintf("ComputeCanonicalSamples...\n");
00240   samples_.ComputeCanonicalSamples(feature_map_, debug_level_ > 0);
00241 }
00242 
00243 // Gets the samples ready for training. Use after both
00244 // ReadTrainingSamples+PostLoadCleanup or DeSerialize.
00245 // Re-indexes the features and computes canonical and cloud features.
00246 void MasterTrainer::PreTrainingSetup() {
00247   if (debug_level_ > 0)
00248     tprintf("PreTrainingSetup...\n");
00249   samples_.IndexFeatures(feature_space_);
00250   samples_.ComputeCanonicalFeatures();
00251   if (debug_level_ > 0)
00252     tprintf("ComputeCloudFeatures...\n");
00253   samples_.ComputeCloudFeatures(feature_space_.Size());
00254 }
00255 
00256 // Sets up the master_shapes_ table, which tells which fonts should stay
00257 // together until they get to a leaf node classifier.
00258 void MasterTrainer::SetupMasterShapes() {
00259   tprintf("Building master shape table\n");
00260   int num_fonts = samples_.NumFonts();
00261 
00262   ShapeTable char_shapes_begin_fragment(samples_.unicharset());
00263   ShapeTable char_shapes_end_fragment(samples_.unicharset());
00264   ShapeTable char_shapes(samples_.unicharset());
00265   for (int c = 0; c < samples_.charsetsize(); ++c) {
00266     ShapeTable shapes(samples_.unicharset());
00267     for (int f = 0; f < num_fonts; ++f) {
00268       if (samples_.NumClassSamples(f, c, true) > 0)
00269         shapes.AddShape(c, f);
00270     }
00271     ClusterShapes(kMinClusteredShapes, 1, kFontMergeDistance, &shapes);
00272 
00273     const CHAR_FRAGMENT *fragment = samples_.unicharset().get_fragment(c);
00274 
00275     if (fragment == NULL)
00276       char_shapes.AppendMasterShapes(shapes);
00277     else if (fragment->is_beginning())
00278       char_shapes_begin_fragment.AppendMasterShapes(shapes);
00279     else if (fragment->is_ending())
00280       char_shapes_end_fragment.AppendMasterShapes(shapes);
00281     else
00282       char_shapes.AppendMasterShapes(shapes);
00283   }
00284   ClusterShapes(kMinClusteredShapes, kMaxUnicharsPerCluster,
00285                 kFontMergeDistance, &char_shapes_begin_fragment);
00286   char_shapes.AppendMasterShapes(char_shapes_begin_fragment);
00287   ClusterShapes(kMinClusteredShapes, kMaxUnicharsPerCluster,
00288                 kFontMergeDistance, &char_shapes_end_fragment);
00289   char_shapes.AppendMasterShapes(char_shapes_end_fragment);
00290   ClusterShapes(kMinClusteredShapes, kMaxUnicharsPerCluster,
00291                 kFontMergeDistance, &char_shapes);
00292   master_shapes_.AppendMasterShapes(char_shapes);
00293   tprintf("Master shape_table:%s\n", master_shapes_.SummaryStr().string());
00294 }
00295 
00296 // Adds the junk_samples_ to the main samples_ set. Junk samples are initially
00297 // fragments and n-grams (all incorrectly segmented characters).
00298 // Various training functions may result in incorrectly segmented characters
00299 // being added to the unicharset of the main samples, perhaps because they
00300 // form a "radical" decomposition of some (Indic) grapheme, or because they
00301 // just look the same as a real character (like rn/m)
00302 // This function moves all the junk samples, to the main samples_ set, but
00303 // desirable junk, being any sample for which the unichar already exists in
00304 // the samples_ unicharset gets the unichar-ids re-indexed to match, but
00305 // anything else gets re-marked as unichar_id 0 (space character) to identify
00306 // it as junk to the error counter.
00307 void MasterTrainer::IncludeJunk() {
00308   // Get ids of fragments in junk_samples_ that replace the dead chars.
00309   const UNICHARSET& junk_set = junk_samples_.unicharset();
00310   const UNICHARSET& sample_set = samples_.unicharset();
00311   int num_junks = junk_samples_.num_samples();
00312   tprintf("Moving %d junk samples to master sample set.\n", num_junks);
00313   for (int s = 0; s < num_junks; ++s) {
00314     TrainingSample* sample = junk_samples_.mutable_sample(s);
00315     int junk_id = sample->class_id();
00316     const char* junk_utf8 = junk_set.id_to_unichar(junk_id);
00317     int sample_id = sample_set.unichar_to_id(junk_utf8);
00318     if (sample_id == INVALID_UNICHAR_ID)
00319       sample_id = 0;
00320     sample->set_class_id(sample_id);
00321     junk_samples_.extract_sample(s);
00322     samples_.AddSample(sample_id, sample);
00323   }
00324   junk_samples_.DeleteDeadSamples();
00325   samples_.OrganizeByFontAndClass();
00326 }
00327 
00328 // Replicates the samples and perturbs them if the enable_replication_ flag
00329 // is set. MUST be used after the last call to OrganizeByFontAndClass on
00330 // the training samples, ie after IncludeJunk if it is going to be used, as
00331 // OrganizeByFontAndClass will eat the replicated samples into the regular
00332 // samples.
00333 void MasterTrainer::ReplicateAndRandomizeSamplesIfRequired() {
00334   if (enable_replication_) {
00335     if (debug_level_ > 0)
00336       tprintf("ReplicateAndRandomize...\n");
00337     verify_samples_.ReplicateAndRandomizeSamples();
00338     samples_.ReplicateAndRandomizeSamples();
00339     samples_.IndexFeatures(feature_space_);
00340   }
00341 }
00342 
00343 // Loads the basic font properties file into fontinfo_table_.
00344 // Returns false on failure.
00345 bool MasterTrainer::LoadFontInfo(const char* filename) {
00346   FILE* fp = fopen(filename, "rb");
00347   if (fp == NULL) {
00348     fprintf(stderr, "Failed to load font_properties from %s\n", filename);
00349     return false;
00350   }
00351   int italic, bold, fixed, serif, fraktur;
00352   while (!feof(fp)) {
00353     FontInfo fontinfo;
00354     char* font_name = new char[1024];
00355     fontinfo.name = font_name;
00356     fontinfo.properties = 0;
00357     fontinfo.universal_id = 0;
00358     if (fscanf(fp, "%1024s %i %i %i %i %i\n", font_name,
00359                &italic, &bold, &fixed, &serif, &fraktur) != 6)
00360       continue;
00361     fontinfo.properties =
00362         (italic << 0) +
00363         (bold << 1) +
00364         (fixed << 2) +
00365         (serif << 3) +
00366         (fraktur << 4);
00367     if (!fontinfo_table_.contains(fontinfo)) {
00368       fontinfo_table_.push_back(fontinfo);
00369     }
00370   }
00371   fclose(fp);
00372   return true;
00373 }
00374 
00375 // Loads the xheight font properties file into xheights_.
00376 // Returns false on failure.
00377 bool MasterTrainer::LoadXHeights(const char* filename) {
00378   tprintf("fontinfo table is of size %d\n", fontinfo_table_.size());
00379   xheights_.init_to_size(fontinfo_table_.size(), -1);
00380   if (filename == NULL) return true;
00381   FILE *f = fopen(filename, "rb");
00382   if (f == NULL) {
00383     fprintf(stderr, "Failed to load font xheights from %s\n", filename);
00384     return false;
00385   }
00386   tprintf("Reading x-heights from %s ...\n", filename);
00387   FontInfo fontinfo;
00388   fontinfo.properties = 0;  // Not used to lookup in the table.
00389   fontinfo.universal_id = 0;
00390   char buffer[1024];
00391   int xht;
00392   int total_xheight = 0;
00393   int xheight_count = 0;
00394   while (!feof(f)) {
00395     if (fscanf(f, "%1024s %d\n", buffer, &xht) != 2)
00396       continue;
00397     fontinfo.name = buffer;
00398     if (!fontinfo_table_.contains(fontinfo)) continue;
00399     int fontinfo_id = fontinfo_table_.get_id(fontinfo);
00400     xheights_[fontinfo_id] = xht;
00401     total_xheight += xht;
00402     ++xheight_count;
00403   }
00404   if (xheight_count == 0) {
00405     fprintf(stderr, "No valid xheights in %s!\n", filename);
00406     return false;
00407   }
00408   int mean_xheight = DivRounded(total_xheight, xheight_count);
00409   for (int i = 0; i < fontinfo_table_.size(); ++i) {
00410     if (xheights_[i] < 0)
00411       xheights_[i] = mean_xheight;
00412   }
00413   return true;
00414 }  // LoadXHeights
00415 
00416 // Reads spacing stats from filename and adds them to fontinfo_table.
00417 bool MasterTrainer::AddSpacingInfo(const char *filename) {
00418   FILE* fontinfo_file = fopen(filename, "rb");
00419   if (fontinfo_file == NULL)
00420     return true;  // We silently ignore missing files!
00421   // Find the fontinfo_id.
00422   int fontinfo_id = GetBestMatchingFontInfoId(filename);
00423   if (fontinfo_id < 0) {
00424     tprintf("No font found matching fontinfo filename %s\n", filename);
00425     fclose(fontinfo_file);
00426     return false;
00427   }
00428   tprintf("Reading spacing from %s for font %d...\n", filename, fontinfo_id);
00429   // TODO(rays) scale should probably be a double, but keep as an int for now
00430   // to duplicate current behavior.
00431   int scale = kBlnXHeight / xheights_[fontinfo_id];
00432   int num_unichars;
00433   char uch[UNICHAR_LEN];
00434   char kerned_uch[UNICHAR_LEN];
00435   int x_gap, x_gap_before, x_gap_after, num_kerned;
00436   ASSERT_HOST(fscanf(fontinfo_file, "%d\n", &num_unichars) == 1);
00437   FontInfo *fi = fontinfo_table_.get_mutable(fontinfo_id);
00438   fi->init_spacing(unicharset_.size());
00439   FontSpacingInfo *spacing = NULL;
00440   for (int l = 0; l < num_unichars; ++l) {
00441     if (fscanf(fontinfo_file, "%s %d %d %d",
00442                uch, &x_gap_before, &x_gap_after, &num_kerned) != 4) {
00443       tprintf("Bad format of font spacing file %s\n", filename);
00444       fclose(fontinfo_file);
00445       return false;
00446     }
00447     bool valid = unicharset_.contains_unichar(uch);
00448     if (valid) {
00449       spacing = new FontSpacingInfo();
00450       spacing->x_gap_before = static_cast<inT16>(x_gap_before * scale);
00451       spacing->x_gap_after = static_cast<inT16>(x_gap_after * scale);
00452     }
00453     for (int k = 0; k < num_kerned; ++k) {
00454       if (fscanf(fontinfo_file, "%s %d", kerned_uch, &x_gap) != 2) {
00455         tprintf("Bad format of font spacing file %s\n", filename);
00456         fclose(fontinfo_file);
00457         return false;
00458       }
00459       if (!valid || !unicharset_.contains_unichar(kerned_uch)) continue;
00460       spacing->kerned_unichar_ids.push_back(
00461           unicharset_.unichar_to_id(kerned_uch));
00462       spacing->kerned_x_gaps.push_back(static_cast<inT16>(x_gap * scale));
00463     }
00464     if (valid) fi->add_spacing(unicharset_.unichar_to_id(uch), spacing);
00465   }
00466   fclose(fontinfo_file);
00467   return true;
00468 }
00469 
00470 // Returns the font id corresponding to the given font name.
00471 // Returns -1 if the font cannot be found.
00472 int MasterTrainer::GetFontInfoId(const char* font_name) {
00473   FontInfo fontinfo;
00474   // We are only borrowing the string, so it is OK to const cast it.
00475   fontinfo.name = const_cast<char*>(font_name);
00476   fontinfo.properties = 0;  // Not used to lookup in the table
00477   fontinfo.universal_id = 0;
00478   if (!fontinfo_table_.contains(fontinfo)) {
00479     return -1;
00480   } else {
00481     return fontinfo_table_.get_id(fontinfo);
00482   }
00483 }
00484 // Returns the font_id of the closest matching font name to the given
00485 // filename. It is assumed that a substring of the filename will match
00486 // one of the fonts. If more than one is matched, the longest is returned.
00487 int MasterTrainer::GetBestMatchingFontInfoId(const char* filename) {
00488   int fontinfo_id = -1;
00489   int best_len = 0;
00490   for (int f = 0; f < fontinfo_table_.size(); ++f) {
00491     if (strstr(filename, fontinfo_table_.get(f).name) != NULL) {
00492       int len = strlen(fontinfo_table_.get(f).name);
00493       // Use the longest matching length in case a substring of a font matched.
00494       if (len > best_len) {
00495         best_len = len;
00496         fontinfo_id = f;
00497       }
00498     }
00499   }
00500   return fontinfo_id;
00501 }
00502 
00503 // Sets up a flat shapetable with one shape per class/font combination.
00504 void MasterTrainer::SetupFlatShapeTable(ShapeTable* shape_table) {
00505   // To exactly mimic the results of the previous implementation, the shapes
00506   // must be clustered in order the fonts arrived, and reverse order of the
00507   // characters within each font.
00508   // Get a list of the fonts in the order they appeared.
00509   GenericVector<int> active_fonts;
00510   int num_shapes = flat_shapes_.NumShapes();
00511   for (int s = 0; s < num_shapes; ++s) {
00512     int font = flat_shapes_.GetShape(s)[0].font_ids[0];
00513     int f = 0;
00514     for (f = 0; f < active_fonts.size(); ++f) {
00515       if (active_fonts[f] == font)
00516         break;
00517     }
00518     if (f == active_fonts.size())
00519       active_fonts.push_back(font);
00520   }
00521   // For each font in order, add all the shapes with that font in reverse order.
00522   int num_fonts = active_fonts.size();
00523   for (int f = 0; f < num_fonts; ++f) {
00524     for (int s = num_shapes - 1; s >= 0; --s) {
00525       int font = flat_shapes_.GetShape(s)[0].font_ids[0];
00526       if (font == active_fonts[f]) {
00527         shape_table->AddShape(flat_shapes_.GetShape(s));
00528       }
00529     }
00530   }
00531 }
00532 
00533 // Sets up a Clusterer for mftraining on a single shape_id.
00534 // Call FreeClusterer on the return value after use.
00535 CLUSTERER* MasterTrainer::SetupForClustering(
00536     const ShapeTable& shape_table,
00537     const FEATURE_DEFS_STRUCT& feature_defs,
00538     int shape_id,
00539     int* num_samples) {
00540 
00541   int desc_index = ShortNameToFeatureType(feature_defs, kMicroFeatureType);
00542   int num_params = feature_defs.FeatureDesc[desc_index]->NumParams;
00543   ASSERT_HOST(num_params == MFCount);
00544   CLUSTERER* clusterer = MakeClusterer(
00545       num_params, feature_defs.FeatureDesc[desc_index]->ParamDesc);
00546 
00547   // We want to iterate over the samples of just the one shape.
00548   IndexMapBiDi shape_map;
00549   shape_map.Init(shape_table.NumShapes(), false);
00550   shape_map.SetMap(shape_id, true);
00551   shape_map.Setup();
00552   // Reverse the order of the samples to match the previous behavior.
00553   GenericVector<const TrainingSample*> sample_ptrs;
00554   SampleIterator it;
00555   it.Init(&shape_map, &shape_table, false, &samples_);
00556   for (it.Begin(); !it.AtEnd(); it.Next()) {
00557     sample_ptrs.push_back(&it.GetSample());
00558   }
00559   int sample_id = 0;
00560   for (int i = sample_ptrs.size() - 1; i >= 0; --i) {
00561     const TrainingSample* sample = sample_ptrs[i];
00562     int num_features = sample->num_micro_features();
00563     for (int f = 0; f < num_features; ++f)
00564       MakeSample(clusterer, sample->micro_features()[f], sample_id);
00565     ++sample_id;
00566   }
00567   *num_samples = sample_id;
00568   return clusterer;
00569 }
00570 
00571 // Writes the given float_classes (produced by SetupForFloat2Int) as inttemp
00572 // to the given inttemp_file, and the corresponding pffmtable.
00573 // The unicharset is the original encoding of graphemes, and shape_set should
00574 // match the size of the shape_table, and may possibly be totally fake.
00575 void MasterTrainer::WriteInttempAndPFFMTable(const UNICHARSET& unicharset,
00576                                              const UNICHARSET& shape_set,
00577                                              const ShapeTable& shape_table,
00578                                              CLASS_STRUCT* float_classes,
00579                                              const char* inttemp_file,
00580                                              const char* pffmtable_file) {
00581   tesseract::Classify *classify = new tesseract::Classify();
00582   // Move the fontinfo table to classify.
00583   classify->get_fontinfo_table().move(&fontinfo_table_);
00584   INT_TEMPLATES int_templates = classify->CreateIntTemplates(float_classes,
00585                                                              shape_set);
00586   FILE* fp = fopen(inttemp_file, "wb");
00587   classify->WriteIntTemplates(fp, int_templates, shape_set);
00588   fclose(fp);
00589   // Now write pffmtable. This is complicated by the fact that the adaptive
00590   // classifier still wants one indexed by unichar-id, but the static
00591   // classifier needs one indexed by its shape class id.
00592   // We put the shapetable_cutoffs in a GenericVector, and compute the
00593   // unicharset cutoffs along the way.
00594   GenericVector<uinT16> shapetable_cutoffs;
00595   GenericVector<uinT16> unichar_cutoffs;
00596   for (int c = 0; c < unicharset.size(); ++c)
00597     unichar_cutoffs.push_back(0);
00598   /* then write out each class */
00599   for (int i = 0; i < int_templates->NumClasses; ++i) {
00600     INT_CLASS Class = ClassForClassId(int_templates, i);
00601     // Todo: Test with min instead of max
00602     // int MaxLength = LengthForConfigId(Class, 0);
00603     uinT16 max_length = 0;
00604     for (int config_id = 0; config_id < Class->NumConfigs; config_id++) {
00605       // Todo: Test with min instead of max
00606       // if (LengthForConfigId (Class, config_id) < MaxLength)
00607       uinT16 length = Class->ConfigLengths[config_id];
00608       if (length > max_length)
00609         max_length = Class->ConfigLengths[config_id];
00610       int shape_id = float_classes[i].font_set.get(config_id);
00611       const Shape& shape = shape_table.GetShape(shape_id);
00612       for (int c = 0; c < shape.size(); ++c) {
00613         int unichar_id = shape[c].unichar_id;
00614         if (length > unichar_cutoffs[unichar_id])
00615           unichar_cutoffs[unichar_id] = length;
00616       }
00617     }
00618     shapetable_cutoffs.push_back(max_length);
00619   }
00620   fp = fopen(pffmtable_file, "wb");
00621   shapetable_cutoffs.Serialize(fp);
00622   for (int c = 0; c < unicharset.size(); ++c) {
00623     const char *unichar = unicharset.id_to_unichar(c);
00624     if (strcmp(unichar, " ") == 0) {
00625       unichar = "NULL";
00626     }
00627     fprintf(fp, "%s %d\n", unichar, unichar_cutoffs[c]);
00628   }
00629   fclose(fp);
00630   free_int_templates(int_templates);
00631 }
00632 
00633 // Generate debug output relating to the canonical distance between the
00634 // two given UTF8 grapheme strings.
00635 void MasterTrainer::DebugCanonical(const char* unichar_str1,
00636                                    const char* unichar_str2) {
00637   int class_id1 = unicharset_.unichar_to_id(unichar_str1);
00638   int class_id2 = unicharset_.unichar_to_id(unichar_str2);
00639   if (class_id2 == INVALID_UNICHAR_ID)
00640     class_id2 = class_id1;
00641   if (class_id1 == INVALID_UNICHAR_ID) {
00642     tprintf("No unicharset entry found for %s\n", unichar_str1);
00643     return;
00644   } else {
00645     tprintf("Font ambiguities for unichar %d = %s and %d = %s\n",
00646             class_id1, unichar_str1, class_id2, unichar_str2);
00647   }
00648   int num_fonts = samples_.NumFonts();
00649   const IntFeatureMap& feature_map = feature_map_;
00650   // Iterate the fonts to get the similarity with other fonst of the same
00651   // class.
00652   tprintf("      ");
00653   for (int f = 0; f < num_fonts; ++f) {
00654     if (samples_.NumClassSamples(f, class_id2, false) == 0)
00655       continue;
00656     tprintf("%6d", f);
00657   }
00658   tprintf("\n");
00659   for (int f1 = 0; f1 < num_fonts; ++f1) {
00660     // Map the features of the canonical_sample.
00661     if (samples_.NumClassSamples(f1, class_id1, false) == 0)
00662       continue;
00663     tprintf("%4d  ", f1);
00664     for (int f2 = 0; f2 < num_fonts; ++f2) {
00665       if (samples_.NumClassSamples(f2, class_id2, false) == 0)
00666         continue;
00667       float dist = samples_.ClusterDistance(f1, class_id1, f2, class_id2,
00668                                             feature_map);
00669       tprintf(" %5.3f", dist);
00670     }
00671     tprintf("\n");
00672   }
00673   // Build a fake ShapeTable containing all the sample types.
00674   ShapeTable shapes(unicharset_);
00675   for (int f = 0; f < num_fonts; ++f) {
00676     if (samples_.NumClassSamples(f, class_id1, true) > 0)
00677       shapes.AddShape(class_id1, f);
00678     if (class_id1 != class_id2 &&
00679         samples_.NumClassSamples(f, class_id2, true) > 0)
00680       shapes.AddShape(class_id2, f);
00681   }
00682 }
00683 
00684 #ifndef GRAPHICS_DISABLED
00685 // Debugging for cloud/canonical features.
00686 // Displays a Features window containing:
00687 // If unichar_str2 is in the unicharset, and canonical_font is non-negative,
00688 // displays the canonical features of the char/font combination in red.
00689 // If unichar_str1 is in the unicharset, and cloud_font is non-negative,
00690 // displays the cloud feature of the char/font combination in green.
00691 // The canonical features are drawn first to show which ones have no
00692 // matches in the cloud features.
00693 // Until the features window is destroyed, each click in the features window
00694 // will display the samples that have that feature in a separate window.
00695 void MasterTrainer::DisplaySamples(const char* unichar_str1, int cloud_font,
00696                                    const char* unichar_str2,
00697                                    int canonical_font) {
00698   const IntFeatureMap& feature_map = feature_map_;
00699   const IntFeatureSpace& feature_space = feature_map.feature_space();
00700   ScrollView* f_window = CreateFeatureSpaceWindow("Features", 100, 500);
00701   ClearFeatureSpaceWindow(norm_mode_ == NM_BASELINE ? baseline : character,
00702                           f_window);
00703   int class_id2 = samples_.unicharset().unichar_to_id(unichar_str2);
00704   if (class_id2 != INVALID_UNICHAR_ID && canonical_font >= 0) {
00705     const TrainingSample* sample = samples_.GetCanonicalSample(canonical_font,
00706                                                                class_id2);
00707     for (int f = 0; f < sample->num_features(); ++f) {
00708       RenderIntFeature(f_window, &sample->features()[f], ScrollView::RED);
00709     }
00710   }
00711   int class_id1 = samples_.unicharset().unichar_to_id(unichar_str1);
00712   if (class_id1 != INVALID_UNICHAR_ID && cloud_font >= 0) {
00713     const BitVector& cloud = samples_.GetCloudFeatures(cloud_font, class_id1);
00714     for (int f = 0; f < cloud.size(); ++f) {
00715       if (cloud[f]) {
00716         INT_FEATURE_STRUCT feature =
00717             feature_map.InverseIndexFeature(f);
00718         RenderIntFeature(f_window, &feature, ScrollView::GREEN);
00719       }
00720     }
00721   }
00722   f_window->Update();
00723   ScrollView* s_window = CreateFeatureSpaceWindow("Samples", 100, 500);
00724   SVEventType ev_type;
00725   do {
00726     SVEvent* ev;
00727     // Wait until a click or popup event.
00728     ev = f_window->AwaitEvent(SVET_ANY);
00729     ev_type = ev->type;
00730     if (ev_type == SVET_CLICK) {
00731       int feature_index = feature_space.XYToFeatureIndex(ev->x, ev->y);
00732       if (feature_index >= 0) {
00733         // Iterate samples and display those with the feature.
00734         Shape shape;
00735         shape.AddToShape(class_id1, cloud_font);
00736         s_window->Clear();
00737         samples_.DisplaySamplesWithFeature(feature_index, shape,
00738                                            feature_space, ScrollView::GREEN,
00739                                            s_window);
00740         s_window->Update();
00741       }
00742     }
00743     delete ev;
00744   } while (ev_type != SVET_DESTROY);
00745 }
00746 #endif  // GRAPHICS_DISABLED
00747 
00748 // Tests the given test_classifier on the internal samples.
00749 // See TestClassifier for details.
00750 void MasterTrainer::TestClassifierOnSamples(int report_level,
00751                                             bool replicate_samples,
00752                                             ShapeClassifier* test_classifier,
00753                                             STRING* report_string) {
00754   TestClassifier(report_level, replicate_samples, &samples_,
00755                  test_classifier, report_string);
00756 }
00757 
00758 // Tests the given test_classifier on the given samples
00759 // report_levels:
00760 // 0 = no output.
00761 // 1 = bottom-line error rate.
00762 // 2 = bottom-line error rate + time.
00763 // 3 = font-level error rate + time.
00764 // 4 = list of all errors + short classifier debug output on 16 errors.
00765 // 5 = list of all errors + short classifier debug output on 25 errors.
00766 // If replicate_samples is true, then the test is run on an extended test
00767 // sample including replicated and systematically perturbed samples.
00768 // If report_string is non-NULL, a summary of the results for each font
00769 // is appended to the report_string.
00770 double MasterTrainer::TestClassifier(int report_level,
00771                                      bool replicate_samples,
00772                                      TrainingSampleSet* samples,
00773                                      ShapeClassifier* test_classifier,
00774                                      STRING* report_string) {
00775   SampleIterator sample_it;
00776   sample_it.Init(NULL, test_classifier->GetShapeTable(), replicate_samples,
00777                  samples);
00778   if (report_level > 0) {
00779     int num_samples = 0;
00780     for (sample_it.Begin(); !sample_it.AtEnd(); sample_it.Next())
00781       ++num_samples;
00782     tprintf("Iterator has charset size of %d/%d, %d shapes, %d samples\n",
00783             sample_it.SparseCharsetSize(), sample_it.CompactCharsetSize(),
00784             test_classifier->GetShapeTable()->NumShapes(), num_samples);
00785     tprintf("Testing %sREPLICATED:\n", replicate_samples ? "" : "NON-");
00786   }
00787   double unichar_error = 0.0;
00788   ErrorCounter::ComputeErrorRate(test_classifier, report_level,
00789                                  CT_SHAPE_TOP_ERR, fontinfo_table_,
00790                                  page_images_, &sample_it, &unichar_error,
00791                                  NULL, report_string);
00792   return unichar_error;
00793 }
00794 
00795 // Returns the average (in some sense) distance between the two given
00796 // shapes, which may contain multiple fonts and/or unichars.
00797 float MasterTrainer::ShapeDistance(const ShapeTable& shapes, int s1, int s2) {
00798   const IntFeatureMap& feature_map = feature_map_;
00799   const Shape& shape1 = shapes.GetShape(s1);
00800   const Shape& shape2 = shapes.GetShape(s2);
00801   int num_chars1 = shape1.size();
00802   int num_chars2 = shape2.size();
00803   float dist_sum = 0.0f;
00804   int dist_count = 0;
00805   if (num_chars1 > 1 || num_chars2 > 1) {
00806     // In the multi-char case try to optimize the calculation by computing
00807     // distances between characters of matching font where possible.
00808     for (int c1 = 0; c1 < num_chars1; ++c1) {
00809       for (int c2 = 0; c2 < num_chars2; ++c2) {
00810         dist_sum += samples_.UnicharDistance(shape1[c1], shape2[c2],
00811                                              true, feature_map);
00812         ++dist_count;
00813       }
00814     }
00815   } else {
00816     // In the single unichar case, there is little alternative, but to compute
00817     // the squared-order distance between pairs of fonts.
00818     dist_sum = samples_.UnicharDistance(shape1[0], shape2[0],
00819                                         false, feature_map);
00820     ++dist_count;
00821   }
00822   return dist_sum / dist_count;
00823 }
00824 
00825 // Replaces samples that are always fragmented with the corresponding
00826 // fragment samples.
00827 void MasterTrainer::ReplaceFragmentedSamples() {
00828   if (fragments_ == NULL) return;
00829   // Remove samples that are replaced by fragments. Each class that was
00830   // always naturally fragmented should be replaced by its fragments.
00831   int num_samples = samples_.num_samples();
00832   for (int s = 0; s < num_samples; ++s) {
00833     TrainingSample* sample = samples_.mutable_sample(s);
00834     if (fragments_[sample->class_id()] > 0)
00835       samples_.KillSample(sample);
00836   }
00837   samples_.DeleteDeadSamples();
00838 
00839   // Get ids of fragments in junk_samples_ that replace the dead chars.
00840   const UNICHARSET& frag_set = junk_samples_.unicharset();
00841 #if 0
00842   // TODO(rays) The original idea was to replace only graphemes that were
00843   // always naturally fragmented, but that left a lot of the Indic graphemes
00844   // out. Determine whether we can go back to that idea now that spacing
00845   // is fixed in the training images, or whether this code is obsolete.
00846   bool* good_junk = new bool[frag_set.size()];
00847   memset(good_junk, 0, sizeof(*good_junk) * frag_set.size());
00848   for (int dead_ch = 1; dead_ch < unicharset_.size(); ++dead_ch) {
00849     int frag_ch = fragments_[dead_ch];
00850     if (frag_ch <= 0) continue;
00851     const char* frag_utf8 = frag_set.id_to_unichar(frag_ch);
00852     CHAR_FRAGMENT* frag = CHAR_FRAGMENT::parse_from_string(frag_utf8);
00853     // Mark the chars for all parts of the fragment as good in good_junk.
00854     for (int part = 0; part < frag->get_total(); ++part) {
00855       frag->set_pos(part);
00856       int good_ch = frag_set.unichar_to_id(frag->to_string().string());
00857       if (good_ch != INVALID_UNICHAR_ID)
00858         good_junk[good_ch] = true;  // We want this one.
00859     }
00860   }
00861 #endif
00862   // For now just use all the junk that was from natural fragments.
00863   // Get samples of fragments in junk_samples_ that replace the dead chars.
00864   int num_junks = junk_samples_.num_samples();
00865   for (int s = 0; s < num_junks; ++s) {
00866     TrainingSample* sample = junk_samples_.mutable_sample(s);
00867     int junk_id = sample->class_id();
00868     const char* frag_utf8 = frag_set.id_to_unichar(junk_id);
00869     CHAR_FRAGMENT* frag = CHAR_FRAGMENT::parse_from_string(frag_utf8);
00870     if (frag != NULL && frag->is_natural()) {
00871       junk_samples_.extract_sample(s);
00872       samples_.AddSample(frag_set.id_to_unichar(junk_id), sample);
00873     }
00874   }
00875   junk_samples_.DeleteDeadSamples();
00876   junk_samples_.OrganizeByFontAndClass();
00877   samples_.OrganizeByFontAndClass();
00878   unicharset_.clear();
00879   unicharset_.AppendOtherUnicharset(samples_.unicharset());
00880   // delete [] good_junk;
00881   // Fragments_ no longer needed?
00882   delete [] fragments_;
00883   fragments_ = NULL;
00884 }
00885 
00886 // Runs a hierarchical agglomerative clustering to merge shapes in the given
00887 // shape_table, while satisfying the given constraints:
00888 // * End with at least min_shapes left in shape_table,
00889 // * No shape shall have more than max_shape_unichars in it,
00890 // * Don't merge shapes where the distance between them exceeds max_dist.
00891 const float kInfiniteDist = 999.0f;
00892 void MasterTrainer::ClusterShapes(int min_shapes,  int max_shape_unichars,
00893                                   float max_dist, ShapeTable* shapes) {
00894   int num_shapes = shapes->NumShapes();
00895   int max_merges = num_shapes - min_shapes;
00896   GenericVector<ShapeDist>* shape_dists =
00897       new GenericVector<ShapeDist>[num_shapes];
00898   float min_dist = kInfiniteDist;
00899   int min_s1 = 0;
00900   int min_s2 = 0;
00901   tprintf("Computing shape distances...");
00902   for (int s1 = 0; s1 < num_shapes; ++s1) {
00903     for (int s2 = s1 + 1; s2 < num_shapes; ++s2) {
00904       ShapeDist dist(s1, s2, ShapeDistance(*shapes, s1, s2));
00905       shape_dists[s1].push_back(dist);
00906       if (dist.distance < min_dist) {
00907         min_dist = dist.distance;
00908         min_s1 = s1;
00909         min_s2 = s2;
00910       }
00911     }
00912     tprintf(" %d", s1);
00913   }
00914   tprintf("\n");
00915   int num_merged = 0;
00916   while (num_merged < max_merges && min_dist < max_dist) {
00917     tprintf("Distance = %f: ", min_dist);
00918     int num_unichars = shapes->MergedUnicharCount(min_s1, min_s2);
00919     shape_dists[min_s1][min_s2 - min_s1 - 1].distance = kInfiniteDist;
00920     if (num_unichars > max_shape_unichars) {
00921       tprintf("Merge of %d and %d with %d would exceed max of %d unichars\n",
00922               min_s1, min_s2, num_unichars, max_shape_unichars);
00923     } else {
00924       shapes->MergeShapes(min_s1, min_s2);
00925       shape_dists[min_s2].clear();
00926       ++num_merged;
00927 
00928       for (int s = 0; s < min_s1; ++s) {
00929         if (!shape_dists[s].empty()) {
00930           shape_dists[s][min_s1 - s - 1].distance =
00931               ShapeDistance(*shapes, s, min_s1);
00932           shape_dists[s][min_s2 - s -1].distance = kInfiniteDist;
00933         }
00934       }
00935       for (int s2 = min_s1 + 1; s2 < num_shapes; ++s2) {
00936         if (shape_dists[min_s1][s2 - min_s1 - 1].distance < kInfiniteDist)
00937           shape_dists[min_s1][s2 - min_s1 - 1].distance =
00938               ShapeDistance(*shapes, min_s1, s2);
00939       }
00940       for (int s = min_s1 + 1; s < min_s2; ++s) {
00941         if (!shape_dists[s].empty()) {
00942           shape_dists[s][min_s2 - s - 1].distance = kInfiniteDist;
00943         }
00944       }
00945     }
00946     min_dist = kInfiniteDist;
00947     for (int s1 = 0; s1 < num_shapes; ++s1) {
00948       for (int i = 0; i < shape_dists[s1].size(); ++i) {
00949         if (shape_dists[s1][i].distance < min_dist) {
00950           min_dist = shape_dists[s1][i].distance;
00951           min_s1 = s1;
00952           min_s2 = s1 + 1 + i;
00953         }
00954       }
00955     }
00956   }
00957   tprintf("Stopped with %d merged, min dist %f\n", num_merged, min_dist);
00958   delete [] shape_dists;
00959   if (debug_level_ > 1) {
00960     for (int s1 = 0; s1 < num_shapes; ++s1) {
00961       if (shapes->MasterDestinationIndex(s1) == s1) {
00962         tprintf("Master shape:%s\n", shapes->DebugStr(s1).string());
00963       }
00964     }
00965   }
00966 }
00967 
00968 
00969 }  // namespace tesseract.