[
https://issues.apache.org/jira/browse/OPENNLP-1366?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17624214#comment-17624214
]
ASF GitHub Bot commented on OPENNLP-1366:
-----------------------------------------
mawiesne commented on code in PR #427:
URL: https://github.com/apache/opennlp/pull/427#discussion_r1005268338
##########
opennlp-tools/src/main/java/opennlp/tools/ml/model/ModelParameterChunker.java:
##########
@@ -0,0 +1,142 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package opennlp.tools.ml.model;
+
+import java.io.DataInputStream;
+import java.io.DataOutputStream;
+import java.io.IOException;
+import java.io.UTFDataFormatException;
+import java.nio.ByteBuffer;
+import java.nio.CharBuffer;
+import java.nio.charset.CharsetEncoder;
+import java.nio.charset.CoderResult;
+import java.nio.charset.StandardCharsets;
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * A helper class that handles Strings with more than 64k (65535 bytes) in
length.
+ * This is achieved via the signature {@link #SIGNATURE_CHUNKED_PARAMS} at the
beginning of
+ * the String instance to be written to a {@link DataOutputStream}.
+ * <p>
+ * Background: In OpenNLP, for large(r) corpora, we train models whose (UTF
String) parameters will exceed
+ * the {@link #MAX_CHUNK_SIZE_BYTES} bytes limit set in {@link
DataOutputStream}.
+ * For writing and reading those models, we have to chunk up those string
instances in 64kB blocks and
+ * recombine them correctly upon reading a (binary) model file.
+ * <p>
+ * The problem was raised in <a
href="https://issues.apache.org/jira/browse/OPENNLP-1366";>ticket
OPENNLP-1366</a>.
+ * <p>
+ * Solution strategy:
+ * <ul>
+ * <li>If writing parameters to a {@link DataOutputStream} blows up with a
{@link UTFDataFormatException} a
+ * large String instance is chunked up and written as appropriate blocks.</li>
+ * <li>To indicate that chunking was conducted, we start with the {@link
#SIGNATURE_CHUNKED_PARAMS} indicator,
+ * directly followed by the number of chunks used. This way, when reading in
chunked model parameters,
+ * recombination is achieved transparently.</li>
+ * </ul>
+ * <p>
+ * Note: Both, existing (binary) model files and newly trained models which
don't require the chunking
+ * technique, will be supported like in previous OpenNLP versions.
+ *
+ * @author <a href="mailto:martin.wies...@hs-heilbronn.de";>Martin Wiesner</a>
+ * @author <a href="mailto:strub...@apache.org";>Mark Struberg</a>
+ */
+public final class ModelParameterChunker {
+
+ /*
+ * A signature that denotes the start of a String that required chunking.
+ *
+ * Semantics:
+ * If a model parameter (String) carries the below signature at the very
beginning, this indicates
+ * that 'n > 1' chunks must be processed to obtain the whole model
parameters. Otherwise, those would not be
+ * written to the binary model files (as reported in OPENNLP-1366) if the
training occurs on large corpora
+ * as used, for instance, in the context of (very large) German NLP models.
+ */
+ public static final String SIGNATURE_CHUNKED_PARAMS =
"CHUNKED-MODEL-PARAMS:"; // followed by no of chunks!
+
+ private static final int MAX_CHUNK_SIZE_BYTES = 65535; // the maximum
'utflen' DataOutputStream can handle
+
+ private ModelParameterChunker(){
+ // private utility class ct s
+ }
+
+ /**
+ * Reads model parameters from {@code dis}. In case the stream start with
{@link #SIGNATURE_CHUNKED_PARAMS},
+ * the number of chunks is detected and the original large parameter string
is reconstructed from several
+ * chunks.
+ *
+ * @param dis The stream which will be used to read the model parameter
from.
+ */
+ public static String readUTF(DataInputStream dis) throws IOException {
+ String data = dis.readUTF();
+ if (data.startsWith(SIGNATURE_CHUNKED_PARAMS)) {
+ String chunkElements = data.replace(SIGNATURE_CHUNKED_PARAMS, "");
+ int chunkSize = Integer.parseInt(chunkElements);
+ StringBuilder sb = new StringBuilder();
+ for (int i = 0; i < chunkSize; i++) {
+ sb.append(dis.readUTF());
+ }
+ return sb.toString(); // the reconstructed model parameter string
+ } else { // default case: no chunked data -> just return the read data /
parameter information
+ return data;
+ }
+ }
+
+ /**
+ * Writes the model parameter {@code s} to {@code dos}. In case {@code s}
does exceed
+ * {@link #MAX_CHUNK_SIZE_BYTES} in length, the chunking mechanism is used;
otherwise the parameter is
+ * written 'as is'.
+ *
+ * @param dos The {@link DataOutputStream} stream which will be used to
persist the model.
+ * @param s The input string that is checked for length and chunked if
{@link #MAX_CHUNK_SIZE_BYTES} is
+ * exceeded.
+ */
+ public static void writeUTF(DataOutputStream dos, String s) throws
IOException {
+ try {
+ dos.writeUTF(s);
+ } catch (UTFDataFormatException dfe) {
Review Comment:
@jzonthemtn In general, this looks entirely plausible. Yet, there are
specific details for encoding characters in UTF(8) format. Therefore, byte
length of `s` is not sufficient or valid for all cases.
Thus, @struberg and I decided to leave it to the actual implementation of
`DataOutputStream.writeUTF(...)` to tell us if the string contains too many
characters. Have a look and check the code there, if interested, for the
"difficult" part. We decided not to cnp the code over to OpenNLP, but to rely
on the checks made in `DataOutputStream`.
I hope, you can follow our idea/decision to go for "avoid cnp" in favor of
catching `UTFDataFormatException` here once, in case the limit exceeds 64k.
This seems acceptable for those "rare" occasions when training with large
corpora.
As a background: The 'Tueba Wikipedia' treebank only runs into this
exception twice; other (well- known) German corpora, like the one from Hamburg
I also used for validation, hit it 3 or 4 times during training.
> Training of MaxEnt Model with large corpora fails with
> java.io.UTFDataFormatException
> --------------------------------------------------------------------------------------
>
> Key: OPENNLP-1366
> URL: https://issues.apache.org/jira/browse/OPENNLP-1366
> Project: OpenNLP
> Issue Type: Improvement
> Components: Lemmatizer
> Affects Versions: 1.9.4, 2.0.0
> Reporter: Richard Zowalla
> Assignee: Jeff Zemerick
> Priority: Major
> Labels: pull-request-available
>
> As written on
> [dev@opennlp.a.o|[https://lists.apache.org/thread/vc5lfzj81tco703noqxpvy8sfj8fw8b1]],
> we are working on training a large opennlp maxent model for lemmatizing
> German texts. We use a wikipedia tree bank from Tübingen.
> This consumes > 2 TB of RAM during training but will finish after some time.
> However, writing this model will result in a java.io.UTFDataFormatException
> However, training such a big model isn't feasable for debugging. Gladly, a
> similar with a smaller dataset is found on Stackoverflow:
> [https://stackoverflow.com/questions/70064477/opennlp-lemmatizertrainer-utfdataformatexception-encoded-string-too-long]
>
> It contains the OpenNLP CLI command to train a lemmatizer on a much smaller
> dataset.
> The stacktrace is raced while writing a String as UTF in DataOutputStream,
> which has a hard-coded size limitation in the JDK (for reasons behind my
> knowledge ;))
> Stacktrace:
> {code:java}
> java.io.UTFDataFormatException: encoded string too long: 383769 bytes
> at
> java.base/java.io.DataOutputStream.writeUTF(DataOutputStream.java:364)
> at
> java.base/java.io.DataOutputStream.writeUTF(DataOutputStream.java:323)
> at
> opennlp.tools.ml.maxent.io.BinaryGISModelWriter.writeUTF(BinaryGISModelWriter.java:71)
> at
> opennlp.tools.ml.maxent.io.GISModelWriter.persist(GISModelWriter.java:97)
> at
> opennlp.tools.ml.model.GenericModelWriter.persist(GenericModelWriter.java:75)
> at opennlp.tools.util.model.ModelUtil.writeModel(ModelUtil.java:71)
> at
> opennlp.tools.util.model.GenericModelSerializer.serialize(GenericModelSerializer.java:36)
> at
> opennlp.tools.util.model.GenericModelSerializer.serialize(GenericModelSerializer.java:29)
> at opennlp.tools.util.model.BaseModel.serialize(BaseModel.java:597)
> at opennlp.tools.cmdline.CmdLineUtil.writeModel(CmdLineUtil.java:182)
> at
> opennlp.tools.cmdline.lemmatizer.LemmatizerTrainerTool.run(LemmatizerTrainerTool.java:77)
> at opennlp.tools.cmdline.CLI.main(CLI.java:256) {code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
