/**
* Serializes the source, target and feature data structures into interlinked binary files. Target
* is written first, into a skeletal (node don't carry any data) upward-pointing trie, updating
* the linking source trie nodes with the position once it is known. Source and feature data are
* written simultaneously. The source structure is written into a downward-pointing trie and
* stores the rule's lhs as well as links to the target and feature stream. The feature stream is
* prompted to write out a block
*
* @param source_trie
* @param target_trie
* @param feature_buffer
* @param id
* @throws IOException
*/
private PackingFileTuple flush(
PackingTrie<SourceValue> source_trie,
PackingTrie<TargetValue> target_trie,
FeatureBuffer feature_buffer,
AlignmentBuffer alignment_buffer,
int id)
throws IOException {
// Make a slice object for this piece of the grammar.
PackingFileTuple slice = new PackingFileTuple("slice_" + String.format("%05d", id));
// Pull out the streams for source, target and data output.
DataOutputStream source_stream = slice.getSourceOutput();
DataOutputStream target_stream = slice.getTargetOutput();
DataOutputStream target_lookup_stream = slice.getTargetLookupOutput();
DataOutputStream feature_stream = slice.getFeatureOutput();
DataOutputStream alignment_stream = slice.getAlignmentOutput();
Queue<PackingTrie<TargetValue>> target_queue;
Queue<PackingTrie<SourceValue>> source_queue;
// The number of bytes both written into the source stream and
// buffered in the source queue.
int source_position;
// The number of bytes written into the target stream.
int target_position;
// Add trie root into queue, set target position to 0 and set cumulated
// size to size of trie root.
target_queue = new LinkedList<PackingTrie<TargetValue>>();
target_queue.add(target_trie);
target_position = 0;
// Target lookup table for trie levels.
int current_level_size = 1;
int next_level_size = 0;
ArrayList<Integer> target_lookup = new ArrayList<Integer>();
// Packing loop for upwards-pointing target trie.
while (!target_queue.isEmpty()) {
// Pop top of queue.
PackingTrie<TargetValue> node = target_queue.poll();
// Register that this is where we're writing the node to.
node.address = target_position;
// Tell source nodes that we're writing to this position in the file.
for (TargetValue tv : node.values) tv.parent.target = node.address;
// Write link to parent.
if (node.parent != null) target_stream.writeInt(node.parent.address);
else target_stream.writeInt(-1);
target_stream.writeInt(node.symbol);
// Enqueue children.
for (int k : node.children.descendingKeySet()) {
PackingTrie<TargetValue> child = node.children.get(k);
target_queue.add(child);
}
target_position += node.size(false, true);
next_level_size += node.children.descendingKeySet().size();
current_level_size--;
if (current_level_size == 0) {
target_lookup.add(target_position);
current_level_size = next_level_size;
next_level_size = 0;
}
}
target_lookup_stream.writeInt(target_lookup.size());
for (int i : target_lookup) target_lookup_stream.writeInt(i);
target_lookup_stream.close();
// Setting up for source and data writing.
source_queue = new LinkedList<PackingTrie<SourceValue>>();
source_queue.add(source_trie);
source_position = source_trie.size(true, false);
source_trie.address = target_position;
// Ready data buffers for writing.
feature_buffer.initialize();
if (packAlignments) alignment_buffer.initialize();
// Packing loop for downwards-pointing source trie.
while (!source_queue.isEmpty()) {
// Pop top of queue.
PackingTrie<SourceValue> node = source_queue.poll();
// Write number of children.
source_stream.writeInt(node.children.size());
// Write links to children.
for (int k : node.children.descendingKeySet()) {
PackingTrie<SourceValue> child = node.children.get(k);
// Enqueue child.
source_queue.add(child);
// Child's address will be at the current end of the queue.
child.address = source_position;
// Advance cumulated size by child's size.
source_position += child.size(true, false);
// Write the link.
source_stream.writeInt(k);
source_stream.writeInt(child.address);
}
// Write number of data items.
source_stream.writeInt(node.values.size());
// Write lhs and links to target and data.
for (SourceValue sv : node.values) {
int feature_block_index = feature_buffer.write(sv.data);
if (packAlignments) {
int alignment_block_index = alignment_buffer.write(sv.data);
if (alignment_block_index != feature_block_index) {
logger.severe("Block index mismatch.");
throw new RuntimeException(
"Block index mismatch: alignment ("
+ alignment_block_index
+ ") and features ("
+ feature_block_index
+ ") don't match.");
}
}
source_stream.writeInt(sv.lhs);
source_stream.writeInt(sv.target);
source_stream.writeInt(feature_block_index);
}
}
// Flush the data stream.
feature_buffer.flush(feature_stream);
if (packAlignments) alignment_buffer.flush(alignment_stream);
target_stream.close();
source_stream.close();
feature_stream.close();
if (packAlignments) alignment_stream.close();
return slice;
}
private void binarize(
LineReader grammar_reader, LineReader alignment_reader, Queue<PackingFileTuple> slices)
throws IOException {
int counter = 0;
int slice_counter = 0;
int num_slices = 0;
boolean ready_to_flush = false;
String first_source_word = null;
PackingTrie<SourceValue> source_trie = new PackingTrie<SourceValue>();
PackingTrie<TargetValue> target_trie = new PackingTrie<TargetValue>();
FeatureBuffer feature_buffer = new FeatureBuffer();
AlignmentBuffer alignment_buffer = null;
if (packAlignments) alignment_buffer = new AlignmentBuffer();
TreeMap<Integer, Float> features = new TreeMap<Integer, Float>();
while (grammar_reader.hasNext()) {
String grammar_line = grammar_reader.next().trim();
counter++;
slice_counter++;
String[] fields = grammar_line.split("\\s\\|{3}\\s");
if (fields.length < 4) {
logger.warning("Incomplete grammar line at line " + counter);
continue;
}
String lhs_word = fields[0];
String[] source_words = fields[1].split("\\s");
String[] target_words = fields[2].split("\\s");
String[] feature_entries = fields[3].split("\\s");
// Reached slice limit size, indicate that we're closing up.
if (!ready_to_flush
&& (slice_counter > SLICE_SIZE
|| feature_buffer.overflowing()
|| (packAlignments && alignment_buffer.overflowing()))) {
ready_to_flush = true;
first_source_word = source_words[0];
}
// Finished closing up.
if (ready_to_flush && !first_source_word.equals(source_words[0])) {
slices.add(flush(source_trie, target_trie, feature_buffer, alignment_buffer, num_slices));
source_trie.clear();
target_trie.clear();
feature_buffer.clear();
if (packAlignments) alignment_buffer.clear();
num_slices++;
slice_counter = 0;
ready_to_flush = false;
}
int alignment_index = -1;
// If present, process alignments.
if (packAlignments) {
if (!alignment_reader.hasNext()) {
logger.severe("No more alignments starting in line " + counter);
throw new RuntimeException("No more alignments starting in line " + counter);
} else {
String alignment_line = alignment_reader.next().trim();
String[] alignment_entries = alignment_line.split("\\s");
byte[] alignments = new byte[alignment_entries.length * 2];
if (alignment_entries.length != 0) {
for (int i = 0; i < alignment_entries.length; i++) {
String[] parts = alignment_entries[i].split("-");
alignments[2 * i] = Byte.parseByte(parts[0]);
alignments[2 * i + 1] = Byte.parseByte(parts[1]);
}
}
alignment_index = alignment_buffer.add(alignments);
}
}
// Process features.
// Implicitly sort via TreeMap, write to data buffer, remember position
// to pass on to the source trie node.
features.clear();
for (int f = 0; f < feature_entries.length; ++f) {
String feature_entry = feature_entries[f];
if (this.labeled) {
String[] parts = feature_entry.split("=");
if (parts[0].equals("Alignment")) continue;
int feature_id = Vocabulary.id(parts[0]);
float feature_value = Float.parseFloat(parts[1]);
if (feature_value != 0) features.put(encoderConfig.innerId(feature_id), feature_value);
} else {
float feature_value = Float.parseFloat(feature_entry);
if (feature_value != 0) features.put(f, feature_value);
}
}
int features_index = feature_buffer.add(features);
// Sanity check on the data block index.
if (packAlignments && features_index != alignment_index) {
logger.severe(
"Block index mismatch between features ("
+ features_index
+ ") and alignments ("
+ alignment_index
+ ").");
throw new RuntimeException("Data block index mismatch.");
}
// Process source side.
SourceValue sv = new SourceValue(Vocabulary.id(lhs_word), features_index);
int[] source = new int[source_words.length];
for (int i = 0; i < source_words.length; i++) {
if (FormatUtils.isNonterminal(source_words[i]))
source[i] = Vocabulary.id(FormatUtils.stripNt(source_words[i]));
else source[i] = Vocabulary.id(source_words[i]);
}
source_trie.add(source, sv);
// Process target side.
TargetValue tv = new TargetValue(sv);
int[] target = new int[target_words.length];
for (int i = 0; i < target_words.length; i++) {
if (FormatUtils.isNonterminal(target_words[i])) {
target[target_words.length - (i + 1)] = -FormatUtils.getNonterminalIndex(target_words[i]);
} else {
target[target_words.length - (i + 1)] = Vocabulary.id(target_words[i]);
}
}
target_trie.add(target, tv);
}
slices.add(flush(source_trie, target_trie, feature_buffer, alignment_buffer, num_slices));
}
