// Helper function to get the index of the first result with the required
// unichar_id. If the results are sorted by rating, this will also be the
// best result with the required unichar_id.
// Returns -1 if the unichar_id is not found
int ShapeRating::FirstResultWithUnichar(
const GenericVector<ShapeRating>& results,
const ShapeTable& shape_table,
UNICHAR_ID unichar_id) {
for (int r = 0; r < results.size(); ++r) {
int shape_id = results[r].shape_id;
const Shape& shape = shape_table.GetShape(shape_id);
if (shape.ContainsUnichar(unichar_id)) {
return r;
}
}
return -1;
}
// Accumulates counts for junk. Counts only whether the junk was correctly
// rejected or not.
void ErrorCounter::AccumulateJunk(const ShapeTable& shape_table,
const GenericVector<ShapeRating>& results,
TrainingSample* sample) {
// For junk we accept no answer, or an explicit shape answer matching the
// class id of the sample.
int num_results = results.size();
int font_id = sample->font_id();
int unichar_id = sample->class_id();
if (num_results > 0 &&
!shape_table.GetShape(results[0].shape_id).ContainsUnichar(unichar_id)) {
// This is a junk error.
++font_counts_[font_id].n[CT_ACCEPTED_JUNK];
sample->set_is_error(true);
// It counts as an error for boosting too so sum the weight.
scaled_error_ += sample->weight();
} else {
// Correctly rejected.
++font_counts_[font_id].n[CT_REJECTED_JUNK];
sample->set_is_error(false);
}
}
// Expands all the classes/fonts in the shape individually to build
// a ShapeTable.
int ShapeTable::BuildFromShape(const Shape& shape,
const ShapeTable& master_shapes) {
int num_masters = 0;
for (int u_ind = 0; u_ind < shape.size(); ++u_ind) {
for (int f_ind = 0; f_ind < shape[u_ind].font_ids.size(); ++f_ind) {
int c = shape[u_ind].unichar_id;
int f = shape[u_ind].font_ids[f_ind];
if (FindShape(c, f) < 0) {
int shape_id = AddShape(c, f);
int master_id = master_shapes.FindShape(c, f);
if (master_id >= 0 && shape.size() > 1) {
const Shape& master = master_shapes.GetShape(master_id);
if (master.IsSubsetOf(shape) && !shape.IsSubsetOf(master)) {
// Add everything else from the master shape.
shape_table_[shape_id]->AddShape(master);
++num_masters;
}
}
}
}
}
return num_masters;
}
// Expands all the classes/fonts in the shape individually to build
// a ShapeTable.
int ShapeTable::BuildFromShape(const Shape& shape,
const ShapeTable& master_shapes) {
BitVector shape_map(master_shapes.NumShapes());
for (int u_ind = 0; u_ind < shape.size(); ++u_ind) {
for (int f_ind = 0; f_ind < shape[u_ind].font_ids.size(); ++f_ind) {
int c = shape[u_ind].unichar_id;
int f = shape[u_ind].font_ids[f_ind];
int master_id = master_shapes.FindShape(c, f);
if (master_id >= 0) {
shape_map.SetBit(master_id);
} else if (FindShape(c, f) < 0) {
AddShape(c, f);
}
}
}
int num_masters = 0;
for (int s = 0; s < master_shapes.NumShapes(); ++s) {
if (shape_map[s]) {
AddShape(master_shapes.GetShape(s));
++num_masters;
}
}
return num_masters;
}
// Accumulates the errors from the classifier results on a single sample.
// Returns true if debug is true and a CT_UNICHAR_TOPN_ERR error occurred.
// boosting_mode selects the type of error to be used for boosting and the
// is_error_ member of sample is set according to whether the required type
// of error occurred. The font_table provides access to font properties
// for error counting and shape_table is used to understand the relationship
// between unichar_ids and shape_ids in the results
bool ErrorCounter::AccumulateErrors(bool debug, CountTypes boosting_mode,
const UnicityTable<FontInfo>& font_table,
const ShapeTable& shape_table,
const GenericVector<ShapeRating>& results,
TrainingSample* sample) {
int num_results = results.size();
int res_index = 0;
bool debug_it = false;
int font_id = sample->font_id();
int unichar_id = sample->class_id();
sample->set_is_error(false);
if (num_results == 0) {
// Reject. We count rejects as a separate category, but still mark the
// sample as an error in case any training module wants to use that to
// improve the classifier.
sample->set_is_error(true);
++font_counts_[font_id].n[CT_REJECT];
} else if (shape_table.GetShape(results[0].shape_id).
ContainsUnicharAndFont(unichar_id, font_id)) {
++font_counts_[font_id].n[CT_SHAPE_TOP_CORRECT];
// Unichar and font OK, but count if multiple unichars.
if (shape_table.GetShape(results[0].shape_id).size() > 1)
++font_counts_[font_id].n[CT_OK_MULTI_UNICHAR];
} else {
// This is a top shape error.
++font_counts_[font_id].n[CT_SHAPE_TOP_ERR];
// Check to see if any font in the top choice has attributes that match.
bool attributes_match = false;
uinT32 font_props = font_table.get(font_id).properties;
const Shape& shape = shape_table.GetShape(results[0].shape_id);
for (int c = 0; c < shape.size() && !attributes_match; ++c) {
for (int f = 0; f < shape[c].font_ids.size(); ++f) {
if (font_table.get(shape[c].font_ids[f]).properties == font_props) {
attributes_match = true;
break;
}
}
}
// TODO(rays) It is easy to add counters for individual font attributes
// here if we want them.
if (!attributes_match)
++font_counts_[font_id].n[CT_FONT_ATTR_ERR];
if (boosting_mode == CT_SHAPE_TOP_ERR) sample->set_is_error(true);
// Find rank of correct unichar answer. (Ignoring the font.)
while (res_index < num_results &&
!shape_table.GetShape(results[res_index].shape_id).
ContainsUnichar(unichar_id)) {
++res_index;
}
if (res_index == 0) {
// Unichar OK, but count if multiple unichars.
if (shape_table.GetShape(results[res_index].shape_id).size() > 1) {
++font_counts_[font_id].n[CT_OK_MULTI_UNICHAR];
}
} else {
// Count maps from unichar id to shape id.
if (num_results > 0)
++unichar_counts_(unichar_id, results[0].shape_id);
// This is a unichar error.
++font_counts_[font_id].n[CT_UNICHAR_TOP1_ERR];
if (boosting_mode == CT_UNICHAR_TOP1_ERR) sample->set_is_error(true);
if (res_index >= MIN(2, num_results)) {
// It is also a 2nd choice unichar error.
++font_counts_[font_id].n[CT_UNICHAR_TOP2_ERR];
if (boosting_mode == CT_UNICHAR_TOP2_ERR) sample->set_is_error(true);
}
if (res_index >= num_results) {
// It is also a top-n choice unichar error.
++font_counts_[font_id].n[CT_UNICHAR_TOPN_ERR];
if (boosting_mode == CT_UNICHAR_TOPN_ERR) sample->set_is_error(true);
debug_it = debug;
}
}
}
// Compute mean number of return values and mean rank of correct answer.
font_counts_[font_id].n[CT_NUM_RESULTS] += num_results;
font_counts_[font_id].n[CT_RANK] += res_index;
// If it was an error for boosting then sum the weight.
if (sample->is_error()) {
scaled_error_ += sample->weight();
}
if (debug_it) {
tprintf("%d results for char %s font %d :",
num_results, shape_table.unicharset().id_to_unichar(unichar_id),
font_id);
for (int i = 0; i < num_results; ++i) {
tprintf(" %.3f/%.3f:%s",
results[i].rating, results[i].font,
shape_table.DebugStr(results[i].shape_id).string());
}
tprintf("\n");
return true;
}
return false;
}