This is an automated email from the ASF dual-hosted git repository.
joern pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/opennlp-sandbox.git
The following commit(s) were added to refs/heads/master by this push:
new faee815 Call close on Tensor objects to release memory
faee815 is described below
commit faee81526837f19d1e3780242144560cdf5c4c92
Author: Jörn Kottmann <[email protected]>
AuthorDate: Mon Jun 18 14:47:35 2018 +0200
Call close on Tensor objects to release memory
---
.../apache/opennlp/namefinder/FeedDictionary.java | 69 ++++++++++-----------
.../apache/opennlp/namefinder/SequenceTagging.java | 70 +++++++++++++---------
2 files changed, 72 insertions(+), 67 deletions(-)
diff --git
a/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/FeedDictionary.java
b/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/FeedDictionary.java
index 26293ef..c8fae3b 100644
--- a/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/FeedDictionary.java
+++ b/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/FeedDictionary.java
@@ -17,40 +17,25 @@
package org.apache.opennlp.namefinder;
-import org.tensorflow.Tensor;
-
import java.util.Arrays;
-public class FeedDictionary {
+import org.tensorflow.Tensor;
+
+public class FeedDictionary implements AutoCloseable {
static int PAD_VALUE = 0;
- private final float dropout;
- private final int[][][] charIds;
- private final int[][] wordLengths;
- private final int[][] wordIds;
+ private final Tensor<Float> dropoutTensor;
+ private final Tensor<Integer> charIdsTensor;
+ private final Tensor<Integer> wordLengthsTensor;
+ private final Tensor<Integer> wordIdsTensor;
private final int[] sentenceLengths;
+ private final Tensor<Integer> sentenceLengthsTensor;
private final int maxSentenceLength;
private final int maxCharLength;
private final int numberOfSentences;
- public float getDropout() {
- return dropout;
- }
-
- public int[][][] getCharIds() {
- return charIds;
- }
-
- public int[][] getWordLengths() {
- return wordLengths;
- }
-
- public int[][] getWordIds() {
- return wordIds;
- }
-
public int[] getSentenceLengths() {
return sentenceLengths;
}
@@ -59,32 +44,29 @@ public class FeedDictionary {
return maxSentenceLength;
}
- public int getMaxCharLength() {
- return maxCharLength;
- }
-
public int getNumberOfSentences() {
return numberOfSentences;
}
- public Tensor<Integer> getSentenceLengthsTensor() {
- return Tensor.create(sentenceLengths, Integer.class);
- }
-
public Tensor<Float> getDropoutTensor() {
- return Tensor.create(dropout, Float.class);
+ return dropoutTensor;
}
public Tensor<Integer> getCharIdsTensor() {
- return Tensor.create(charIds, Integer.class);
+ return charIdsTensor;
}
+ public Tensor<Integer> getSentenceLengthsTensor() {
+ return sentenceLengthsTensor;
+ }
+
+
public Tensor<Integer> getWordLengthsTensor() {
- return Tensor.create(wordLengths, Integer.class);
+ return wordLengthsTensor;
}
public Tensor<Integer> getWordIdsTensor() {
- return Tensor.create(wordIds, Integer.class);
+ return wordIdsTensor;
}
private FeedDictionary(final float dropout,
@@ -96,17 +78,26 @@ public class FeedDictionary {
final int maxCharLength,
final int numberOfSentences) {
- this.dropout = dropout;
- this.charIds = charIds;
- this.wordLengths = wordLengths;
- this.wordIds = wordIds;
+ dropoutTensor = Tensor.create(dropout, Float.class);
+ charIdsTensor = Tensor.create(charIds, Integer.class);
+ wordLengthsTensor = Tensor.create(wordLengths, Integer.class);
+ wordIdsTensor = Tensor.create(wordIds, Integer.class);
this.sentenceLengths = sentenceLengths;
+ sentenceLengthsTensor = Tensor.create(sentenceLengths, Integer.class);
this.maxSentenceLength = maxSentenceLength;
this.maxCharLength = maxCharLength;
this.numberOfSentences = numberOfSentences;
}
+ public void close() {
+ dropoutTensor.close();
+ charIdsTensor.close();
+ wordLengthsTensor.close();
+ wordIdsTensor.close();
+ sentenceLengthsTensor.close();
+ }
+
// multi sentences
public static FeedDictionary create(TokenIds sentences) {
diff --git
a/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/SequenceTagging.java
b/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/SequenceTagging.java
index 69a1748..5f3d513 100644
---
a/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/SequenceTagging.java
+++
b/tf-ner-poc/src/main/java/org/apache/opennlp/namefinder/SequenceTagging.java
@@ -32,9 +32,7 @@ import org.tensorflow.Session;
import org.tensorflow.Tensor;
import opennlp.tools.namefind.BioCodec;
-import opennlp.tools.namefind.NameSample;
import opennlp.tools.namefind.TokenNameFinder;
-import opennlp.tools.tokenize.SimpleTokenizer;
import opennlp.tools.util.Span;
public class SequenceTagging implements TokenNameFinder, AutoCloseable {
@@ -87,8 +85,13 @@ public class SequenceTagging implements TokenNameFinder,
AutoCloseable {
@Override
public Span[] find(String[] sentence) {
- TokenIds tokenIds = wordIndexer.toTokenIds(sentence);
- return new BioCodec().decode(Arrays.asList(predict(tokenIds)[0]));
+ if (sentence.length > 0) {
+ TokenIds tokenIds = wordIndexer.toTokenIds(sentence);
+ return new BioCodec().decode(Arrays.asList(predict(tokenIds)[0]));
+ }
+ else {
+ return new Span[0];
+ }
}
public String[][] predict(String[][] sentences) {
@@ -97,31 +100,42 @@ public class SequenceTagging implements TokenNameFinder,
AutoCloseable {
}
private String[][] predict(TokenIds tokenIds) {
- FeedDictionary fd = FeedDictionary.create(tokenIds);
-
- List<Tensor<?>> run = session.runner()
- .feed("chars/char_ids:0", fd.getCharIdsTensor())
- .feed("dropout_keep_prop:0", fd.getDropoutTensor())
- .feed("words/sequence_lengths:0", fd.getSentenceLengthsTensor())
- .feed("words/word_ids:0", fd.getWordIdsTensor())
- .feed("chars/word_lengths:0", fd.getWordLengthsTensor())
- .fetch("logits", 0)
- .fetch("trans_params", 0).run();
-
- float[][][] logits = new
float[fd.getNumberOfSentences()][fd.getMaxSentenceLength()][indexTagger.getNumberOfTags()];
- run.get(0).copyTo(logits);
-
- float[][] trans_params = new
float[indexTagger.getNumberOfTags()][indexTagger.getNumberOfTags()];
- run.get(1).copyTo(trans_params);
-
- String[][] returnValue = new String[fd.getNumberOfSentences()][];
- for (int i=0; i < logits.length; i++) {
- //logit = logit[:sequence_length] # keep only the valid steps
- float[][] logit = Arrays.copyOf(logits[i], fd.getSentenceLengths()[i]);
- returnValue[i] = Viterbi.decode(logit,
trans_params).stream().map(indexTagger::getTag).toArray(String[]::new);
- }
- return returnValue;
+ try (FeedDictionary fd = FeedDictionary.create(tokenIds)) {
+
+ List<Tensor<?>> run = session.runner()
+ .feed("chars/char_ids:0", fd.getCharIdsTensor())
+ .feed("dropout_keep_prop:0", fd.getDropoutTensor())
+ .feed("words/sequence_lengths:0", fd.getSentenceLengthsTensor())
+ .feed("words/word_ids:0", fd.getWordIdsTensor())
+ .feed("chars/word_lengths:0", fd.getWordLengthsTensor())
+ .fetch("logits", 0)
+ .fetch("trans_params", 0).run();
+
+ float[][][] logits = new
float[fd.getNumberOfSentences()][fd.getMaxSentenceLength()][indexTagger.getNumberOfTags()];
+ run.get(0).copyTo(logits);
+
+ float[][] trans_params = new
float[indexTagger.getNumberOfTags()][indexTagger.getNumberOfTags()];
+ run.get(1).copyTo(trans_params);
+
+ String[][] returnValue = new String[fd.getNumberOfSentences()][];
+ for (int i = 0; i < logits.length; i++) {
+ float[][] logit = Arrays.copyOf(logits[i], fd.getSentenceLengths()[i]);
+ returnValue[i] = Viterbi.decode(logit,
trans_params).stream().map(indexTagger::getTag).toArray(String[]::new);
+ }
+
+ for (int i = 0; i < returnValue[0].length; i++) {
+ if (returnValue[0][i] == null) {
+ returnValue[0][i] = "other";
+ }
+ }
+
+ for (Tensor t : run) {
+ t.close();
+ }
+
+ return returnValue;
+ }
}
@Override
