void Wordrec::ProcessSegSearchPainPoint(
float pain_point_priority,
const MATRIX_COORD &pain_point, const char* pain_point_type,
GenericVector<SegSearchPending>* pending, WERD_RES *word_res,
LMPainPoints *pain_points, BlamerBundle *blamer_bundle) {
if (segsearch_debug_level > 0) {
tprintf("Classifying pain point %s priority=%.4f, col=%d, row=%d\n",
pain_point_type, pain_point_priority,
pain_point.col, pain_point.row);
}
ASSERT_HOST(pain_points != NULL);
MATRIX *ratings = word_res->ratings;
// Classify blob [pain_point.col pain_point.row]
if (!pain_point.Valid(*ratings)) {
ratings->IncreaseBandSize(pain_point.row + 1 - pain_point.col);
}
ASSERT_HOST(pain_point.Valid(*ratings));
BLOB_CHOICE_LIST *classified = classify_piece(word_res->seam_array,
pain_point.col, pain_point.row,
pain_point_type,
word_res->chopped_word,
blamer_bundle);
BLOB_CHOICE_LIST *lst = ratings->get(pain_point.col, pain_point.row);
if (lst == NULL) {
ratings->put(pain_point.col, pain_point.row, classified);
} else {
// We can not delete old BLOB_CHOICEs, since they might contain
// ViterbiStateEntries that are parents of other "active" entries.
// Thus if the matrix cell already contains classifications we add
// the new ones to the beginning of the list.
BLOB_CHOICE_IT it(lst);
it.add_list_before(classified);
delete classified; // safe to delete, since empty after add_list_before()
classified = NULL;
}
if (segsearch_debug_level > 0) {
print_ratings_list("Updated ratings matrix with a new entry:",
ratings->get(pain_point.col, pain_point.row),
getDict().getUnicharset());
ratings->print(getDict().getUnicharset());
}
// Insert initial "pain points" to join the newly classified blob
// with its left and right neighbors.
if (classified != NULL && !classified->empty()) {
if (pain_point.col > 0) {
pain_points->GeneratePainPoint(
pain_point.col - 1, pain_point.row, LM_PPTYPE_SHAPE, 0.0,
true, segsearch_max_char_wh_ratio, word_res);
}
if (pain_point.row + 1 < ratings->dimension()) {
pain_points->GeneratePainPoint(
pain_point.col, pain_point.row + 1, LM_PPTYPE_SHAPE, 0.0,
true, segsearch_max_char_wh_ratio, word_res);
}
}
(*pending)[pain_point.col].SetBlobClassified(pain_point.row);
}
ChoiceIterator::ChoiceIterator(const LTRResultIterator& result_it) {
ASSERT_HOST(result_it.it_->word() != NULL);
word_res_ = result_it.it_->word();
BLOB_CHOICE_LIST* choices = NULL;
if (word_res_->ratings != NULL)
choices = word_res_->GetBlobChoices(result_it.blob_index_);
if (choices != NULL && !choices->empty()) {
choice_it_ = new BLOB_CHOICE_IT(choices);
choice_it_->mark_cycle_pt();
} else {
choice_it_ = NULL;
}
}
/**
* @name improve_by_chopping
*
* Repeatedly chops the worst blob, classifying the new blobs fixing up all
* the data, and incrementally runs the segmentation search until a good word
* is found, or no more chops can be found.
*/
void Wordrec::improve_by_chopping(float rating_cert_scale,
WERD_RES* word,
BestChoiceBundle* best_choice_bundle,
BlamerBundle* blamer_bundle,
LMPainPoints* pain_points,
GenericVector<SegSearchPending>* pending) {
int blob_number;
do { // improvement loop.
// Make a simple vector of BLOB_CHOICEs to make it easy to pick which
// one to chop.
GenericVector<BLOB_CHOICE*> blob_choices;
int num_blobs = word->ratings->dimension();
for (int i = 0; i < num_blobs; ++i) {
BLOB_CHOICE_LIST* choices = word->ratings->get(i, i);
if (choices == NULL || choices->empty()) {
blob_choices.push_back(NULL);
} else {
BLOB_CHOICE_IT bc_it(choices);
blob_choices.push_back(bc_it.data());
}
}
SEAM* seam = improve_one_blob(blob_choices, &best_choice_bundle->fixpt,
false, false, word, &blob_number);
if (seam == NULL) break;
// A chop has been made. We have to correct all the data structures to
// take into account the extra bottom-level blob.
// Put the seam into the seam_array and correct everything else on the
// word: ratings matrix (including matrix location in the BLOB_CHOICES),
// states in WERD_CHOICEs, and blob widths.
word->InsertSeam(blob_number, seam);
// Insert a new entry in the beam array.
best_choice_bundle->beam.insert(new LanguageModelState, blob_number);
// Fixpts are outdated, but will get recalculated.
best_choice_bundle->fixpt.clear();
// Remap existing pain points.
pain_points->RemapForSplit(blob_number);
// Insert a new pending at the chop point.
pending->insert(SegSearchPending(), blob_number);
// Classify the two newly created blobs using ProcessSegSearchPainPoint,
// as that updates the pending correctly and adds new pain points.
MATRIX_COORD pain_point(blob_number, blob_number);
ProcessSegSearchPainPoint(0.0f, pain_point, "Chop1", pending, word,
pain_points, blamer_bundle);
pain_point.col = blob_number + 1;
pain_point.row = blob_number + 1;
ProcessSegSearchPainPoint(0.0f, pain_point, "Chop2", pending, word,
pain_points, blamer_bundle);
if (language_model_->language_model_ngram_on) {
// N-gram evaluation depends on the number of blobs in a chunk, so we
// have to re-evaluate everything in the word.
ResetNGramSearch(word, best_choice_bundle, pending);
blob_number = 0;
}
// Run language model incrementally. (Except with the n-gram model on.)
UpdateSegSearchNodes(rating_cert_scale, blob_number, pending,
word, pain_points, best_choice_bundle, blamer_bundle);
} while (!language_model_->AcceptableChoiceFound() &&
word->ratings->dimension() < kMaxNumChunks);
// If after running only the chopper best_choice is incorrect and no blame
// has been yet set, blame the classifier if best_choice is classifier's
// top choice and is a dictionary word (i.e. language model could not have
// helped). Otherwise blame the tradeoff between the classifier and
// the old language model (permuters).
if (word->blamer_bundle != NULL &&
word->blamer_bundle->incorrect_result_reason() == IRR_CORRECT &&
!word->blamer_bundle->ChoiceIsCorrect(word->best_choice)) {
bool valid_permuter = word->best_choice != NULL &&
Dict::valid_word_permuter(word->best_choice->permuter(), false);
word->blamer_bundle->BlameClassifierOrLangModel(word,
getDict().getUnicharset(),
valid_permuter,
wordrec_debug_blamer);
}
}
void Wordrec::SegSearch(CHUNKS_RECORD *chunks_record,
WERD_CHOICE *best_choice,
BLOB_CHOICE_LIST_VECTOR *best_char_choices,
WERD_CHOICE *raw_choice,
STATE *output_best_state) {
int row, col = 0;
if (segsearch_debug_level > 0) {
tprintf("Starting SegSearch on ratings matrix:\n");
chunks_record->ratings->print(getDict().getUnicharset());
}
// Start with a fresh best_choice since rating adjustments
// used by the chopper and the new segmentation search are not compatible.
best_choice->set_rating(WERD_CHOICE::kBadRating);
// Clear best choice accumulator (that is used for adaption), so that
// choices adjusted by chopper do not interfere with the results from the
// segmentation search.
getDict().ClearBestChoiceAccum();
MATRIX *ratings = chunks_record->ratings;
// Priority queue containing pain points generated by the language model
// The priority is set by the language model components, adjustments like
// seam cost and width priority are factored into the priority.
HEAP *pain_points = MakeHeap(segsearch_max_pain_points);
// best_path_by_column records the lowest cost path found so far for each
// column of the chunks_record->ratings matrix over all the rows.
BestPathByColumn *best_path_by_column =
new BestPathByColumn[ratings->dimension()];
for (col = 0; col < ratings->dimension(); ++col) {
best_path_by_column[col].avg_cost = WERD_CHOICE::kBadRating;
best_path_by_column[col].best_vse = NULL;
}
language_model_->InitForWord(prev_word_best_choice_, &denorm_,
assume_fixed_pitch_char_segment,
best_choice->certainty(),
segsearch_max_char_wh_ratio,
pain_points, chunks_record);
MATRIX_COORD *pain_point;
float pain_point_priority;
BestChoiceBundle best_choice_bundle(
output_best_state, best_choice, raw_choice, best_char_choices);
// pending[i] stores a list of the parent/child pair of BLOB_CHOICE_LISTs,
// where i is the column of the child. Initially all the classified entries
// in the ratings matrix from column 0 (with parent NULL) are inserted into
// pending[0]. As the language model state is updated, new child/parent
// pairs are inserted into the lists. Next, the entries in pending[1] are
// considered, and so on. It is important that during the update the
// children are considered in the non-decreasing order of their column, since
// this guarantess that all the parents would be up to date before an update
// of a child is done.
SEG_SEARCH_PENDING_LIST *pending =
new SEG_SEARCH_PENDING_LIST[ratings->dimension()];
// Search for the ratings matrix for the initial best path.
for (row = 0; row < ratings->dimension(); ++row) {
if (ratings->get(0, row) != NOT_CLASSIFIED) {
pending[0].add_sorted(
SEG_SEARCH_PENDING::compare, true,
new SEG_SEARCH_PENDING(row, NULL, LanguageModel::kAllChangedFlag));
}
}
UpdateSegSearchNodes(0, &pending, &best_path_by_column, chunks_record,
pain_points, &best_choice_bundle);
// Keep trying to find a better path by fixing the "pain points".
int num_futile_classifications = 0;
while (!(language_model_->AcceptableChoiceFound() ||
num_futile_classifications >=
segsearch_max_futile_classifications)) {
// Get the next valid "pain point".
int pop;
while (true) {
pop = HeapPop(pain_points, &pain_point_priority, &pain_point);
if (pop == EMPTY) break;
if (pain_point->Valid(*ratings) &&
ratings->get(pain_point->col, pain_point->row) == NOT_CLASSIFIED) {
break;
} else {
delete pain_point;
}
}
if (pop == EMPTY) {
if (segsearch_debug_level > 0) tprintf("Pain points queue is empty\n");
break;
}
if (segsearch_debug_level > 0) {
tprintf("Classifying pain point priority=%.4f, col=%d, row=%d\n",
pain_point_priority, pain_point->col, pain_point->row);
}
BLOB_CHOICE_LIST *classified = classify_piece(
chunks_record->chunks, chunks_record->splits,
pain_point->col, pain_point->row);
ratings->put(pain_point->col, pain_point->row, classified);
if (segsearch_debug_level > 0) {
print_ratings_list("Updated ratings matrix with a new entry:",
ratings->get(pain_point->col, pain_point->row),
getDict().getUnicharset());
chunks_record->ratings->print(getDict().getUnicharset());
}
// Insert initial "pain points" to join the newly classified blob
// with its left and right neighbors.
if (!classified->empty()) {
float worst_piece_cert;
bool fragmented;
if (pain_point->col > 0) {
language_model_->GetWorstPieceCertainty(
pain_point->col-1, pain_point->row, chunks_record->ratings,
&worst_piece_cert, &fragmented);
language_model_->GeneratePainPoint(
pain_point->col-1, pain_point->row, false,
LanguageModel::kInitialPainPointPriorityAdjustment,
worst_piece_cert, fragmented, best_choice->certainty(),
segsearch_max_char_wh_ratio, NULL, NULL,
chunks_record, pain_points);
}
if (pain_point->row+1 < ratings->dimension()) {
language_model_->GetWorstPieceCertainty(
pain_point->col, pain_point->row+1, chunks_record->ratings,
&worst_piece_cert, &fragmented);
language_model_->GeneratePainPoint(
pain_point->col, pain_point->row+1, true,
LanguageModel::kInitialPainPointPriorityAdjustment,
worst_piece_cert, fragmented, best_choice->certainty(),
segsearch_max_char_wh_ratio, NULL, NULL,
chunks_record, pain_points);
}
}
// Record a pending entry with the pain_point and each of its parents.
int parent_row = pain_point->col - 1;
if (parent_row < 0) { // this node has no parents
pending[pain_point->col].add_sorted(
SEG_SEARCH_PENDING::compare, true,
new SEG_SEARCH_PENDING(pain_point->row, NULL,
LanguageModel::kAllChangedFlag));
} else {
for (int parent_col = 0; parent_col < pain_point->col; ++parent_col) {
if (ratings->get(parent_col, parent_row) != NOT_CLASSIFIED) {
pending[pain_point->col].add_sorted(
SEG_SEARCH_PENDING::compare, true,
new SEG_SEARCH_PENDING(pain_point->row,
ratings->get(parent_col, parent_row),
LanguageModel::kAllChangedFlag));
}
}
}
UpdateSegSearchNodes(pain_point->col, &pending, &best_path_by_column,
chunks_record, pain_points, &best_choice_bundle);
if (!best_choice_bundle.updated) ++num_futile_classifications;
if (segsearch_debug_level > 0) {
tprintf("num_futile_classifications %d\n", num_futile_classifications);
}
// Clean up
best_choice_bundle.updated = false;
delete pain_point; // done using this pain point
}
if (segsearch_debug_level > 0) {
tprintf("Done with SegSearch (AcceptableChoiceFound: %d\n",
language_model_->AcceptableChoiceFound());
}
// Clean up.
FreeHeapData(pain_points, MATRIX_COORD::Delete);
delete[] best_path_by_column;
delete[] pending;
for (row = 0; row < ratings->dimension(); ++row) {
for (col = 0; col <= row; ++col) {
BLOB_CHOICE_LIST *rating = ratings->get(col, row);
if (rating != NOT_CLASSIFIED) language_model_->DeleteState(rating);
}
}
}
