From 5c1ebdcb9e631a9e7e59f343993b9916b15405b5 Mon Sep 17 00:00:00 2001
From: Vojtech Moravec <vojtech.moravec.st@vsb.cz>
Date: Tue, 8 Sep 2020 08:16:52 +0200
Subject: [PATCH] Add variant of quantizeIntoIndices with KDTree lookup.
In early experiments BBF algorithm with low maxE value is faster than
conventional brute force search for dimensions up to 27 (3x3x3).
VQImageCompressor still uses the old method but it can be switched
easily in the code.
---
.../compression/VQImageCompressor.java | 3 ++
.../quantization/vector/VectorQuantizer.java | 43 ++++++++++---------
2 files changed, 26 insertions(+), 20 deletions(-)
diff --git a/src/main/java/azgracompress/compression/VQImageCompressor.java b/src/main/java/azgracompress/compression/VQImageCompressor.java
index 8b522aa..b32a93d 100644
--- a/src/main/java/azgracompress/compression/VQImageCompressor.java
+++ b/src/main/java/azgracompress/compression/VQImageCompressor.java
@@ -159,6 +159,9 @@ public class VQImageCompressor extends CompressorDecompressorBase implements IIm
assert (quantizer != null);
+ // Use BestBinFirst KDTree for codebook lookup.
+// final int[] indices = quantizer.quantizeIntoIndicesUsingKDTree(planeVectors, options.getWorkerCount());
+ // Use BruteForce for codebook lookup.
final int[] indices = quantizer.quantizeIntoIndices(planeVectors, options.getWorkerCount());
planeDataSizes[planeCounter++] = writeHuffmanEncodedIndices(compressStream, huffman, indices);
diff --git a/src/main/java/azgracompress/quantization/vector/VectorQuantizer.java b/src/main/java/azgracompress/quantization/vector/VectorQuantizer.java
index a0e043a..4fcc8ca 100644
--- a/src/main/java/azgracompress/quantization/vector/VectorQuantizer.java
+++ b/src/main/java/azgracompress/quantization/vector/VectorQuantizer.java
@@ -6,6 +6,10 @@ import azgracompress.utilities.Utils;
public class VectorQuantizer {
+ private interface QuantizeVectorMethod {
+ int call(final int[] vector);
+ }
+
private final VectorDistanceMetric metric = VectorDistanceMetric.Euclidean;
private final CodebookEntry[] codebookVectors;
private final int vectorSize;
@@ -55,26 +59,16 @@ public class VectorQuantizer {
return quantizeIntoIndices(dataVectors, 1);
}
- public int[] quantizeIntoIndicesUsingKDTree(final int[][] dataVectors, final int maxWorkerCount) {
- assert (dataVectors.length > 0 && dataVectors[0].length % vectorSize == 0) : "Wrong vector size";
- int[] indices = new int[dataVectors.length];
-
- for (int vectorIndex = 0; vectorIndex < dataVectors.length; vectorIndex++) {
+ private int[] quantizeIntoIndicesImpl(final int[][] dataVectors,
+ final int maxWorkerCount,
+ final QuantizeVectorMethod method) {
- indices[vectorIndex] = kdTree.findNearestBBF(dataVectors[vectorIndex], 32);
- }
- return indices;
- }
-
- public int[] quantizeIntoIndices(final int[][] dataVectors, final int maxWorkerCount) {
- assert (dataVectors.length > 0 && dataVectors[0].length % vectorSize == 0) : "Wrong vector size";
+ assert (dataVectors.length > 0 && dataVectors[0].length == vectorSize) : "Wrong vector size";
int[] indices = new int[dataVectors.length];
if (maxWorkerCount == 1) {
- int closestIndex;
for (int vectorIndex = 0; vectorIndex < dataVectors.length; vectorIndex++) {
- closestIndex = findClosestCodebookEntryIndex(dataVectors[vectorIndex], metric);
- indices[vectorIndex] = closestIndex;
+ indices[vectorIndex] = method.call(dataVectors[vectorIndex]);
}
} else {
// Cap the worker count on 8
@@ -86,12 +80,9 @@ public class VectorQuantizer {
final int fromIndex = wId * workSize;
final int toIndex = (wId == workerCount - 1) ? dataVectors.length : (workSize + (wId * workSize));
-
workers[wId] = new Thread(() -> {
- int closestIndex;
for (int vectorIndex = fromIndex; vectorIndex < toIndex; vectorIndex++) {
- closestIndex = findClosestCodebookEntryIndex(dataVectors[vectorIndex], metric);
- indices[vectorIndex] = closestIndex;
+ indices[vectorIndex] = method.call(dataVectors[vectorIndex]);
}
});
@@ -104,11 +95,23 @@ public class VectorQuantizer {
} catch (InterruptedException e) {
e.printStackTrace();
}
-
}
+
return indices;
}
+ public int[] quantizeIntoIndicesUsingKDTree(final int[][] dataVectors, final int maxWorkerCount) {
+
+ return quantizeIntoIndicesImpl(dataVectors, maxWorkerCount, (final int[] queryVector) ->
+ kdTree.findNearestBBF(queryVector, 8));
+ }
+
+ public int[] quantizeIntoIndices(final int[][] dataVectors, final int maxWorkerCount) {
+
+ return quantizeIntoIndicesImpl(dataVectors, maxWorkerCount, (final int[] queryVector) ->
+ findClosestCodebookEntryIndex(queryVector, metric));
+ }
+
public static double distanceBetweenVectors(final int[] originalDataVector,
final int[] codebookEntry,
final VectorDistanceMetric metric) {
--
GitLab