Mallet Api - Get consistent results - lda

Question

I am new to LDA and mallet. I have the following query
I tried running Mallet-LDA with the command line and by setting the --random-seed to a fixed value, I was able to get consistent results for multiple runs of the algorithm
However, I did try with the Mallet-Java-API and everytime I run the program I get different output.
I did google around and found out that random-seed needs to be fixed and I have it fixed in my java code. I still am getting different results.
Could anyone let me know what other parameters do I need to consider for consistent results (when run multiple times)
I might want to add that train-topics when ran multiple times(command line) yields same result. However, when I rerun import-dir and then run train-topics, the results do not match with previous one. (Probably as expected).
I am ok with running import-dir just once and then experiment with different number of topics and iterations by running train-topics.
Similarly, what needs to be changed/ kept constant if I want to replicate the same when I use Java-Api.

Answer 1

I was able to solve this.
I will respond in detail here:
There are two ways in which Mallet could be run.
a. Command mode
b. Using Java API
To get consistent results for different runs, we need to fix the 'random seed' and in the command line we have an option of setting it. We have no surprises there.
However, while using APIs, though we have an option of setting 'random seed', we need to know that it needs to be done at proper point, else it does not work. (see code)
I have pasted the code here which would create a model(read InstanceList) file from the data
and then we could use the same model file and set the random seed and see to it that we get consistent(read same) results every time we run.
Creating and saving model for later use.
Note: Follow this link to know the format of input file.
http://mallet.cs.umass.edu/ap.txt
public void getModelReady(String inputFile) throws IOException {
if(inputFile != null && (! inputFile.isEmpty())) {
List<Pipe> pipeList = new ArrayList<Pipe>();
pipeList.add(new Target2Label());
pipeList.add(new Input2CharSequence("UTF-8"));
pipeList.add(new CharSequence2TokenSequence());
pipeList.add(new TokenSequenceLowercase());
pipeList.add(new TokenSequenceRemoveStopwords());
pipeList.add(new TokenSequence2FeatureSequence());
Reader fileReader = new InputStreamReader(new FileInputStream(new File(inputFile)), "UTF-8");
CsvIterator ci = new CsvIterator (fileReader, Pattern.compile("^(\\S*)[\\s,]*(\\S*)[\\s,]*(.*)$"),
3, 2, 1); // data, label, name fields
InstanceList instances = new InstanceList(new SerialPipes(pipeList));
instances.addThruPipe(ci);
ObjectOutputStream oos;
oos = new ObjectOutputStream(new FileOutputStream("Resources\\Input\\Model\\Model.vectors"));
oos.writeObject(instances);
oos.close();
}
}
Once model file is saved, this uses the above saved file to generate topics
public void applyLDA(ParallelTopicModel model) throws IOException {
InstanceList training = InstanceList.load (new File("Resources\\Input\\Model\\Model.vectors"));
logger.debug("InstanceList Data loaded.");
if (training.size() > 0 &&
training.get(0) != null) {
Object data = training.get(0).getData();
if (! (data instanceof FeatureSequence)) {
logger.error("Topic modeling currently only supports feature sequences.");
System.exit(1);
}
}
// IT HAS TO BE SET HERE, BEFORE CALLING ADDINSTANCE METHOD.
model.setRandomSeed(5);
model.addInstances(training);
model.estimate();
model.printTopWords(new File("Resources\\Output\\OutputFile\\topic_keys_java.txt"), 25,
false);
model.printDocumentTopics(new File ("Resources\\Output\\OutputFile\\document_topicssplit_java.txt"));
}