sszuev / fasttext_java

Java port of c++ version of facebook fasttext

fasttext_java

Java port of c++ facebook fastText and fork of ivanhk fastText_java port.

The main differences with the original port:

• synchronization with c++ version
• support any kind of I/O streams (including hadoop and web in extra)
• tests
• java8
• changes in java-coding & OOP styles

Version

release c++ version: 0.1.0

last checked c++ revision number: 09/12/2017

bin model version: 12

Resources

You can find more information and resources at https://github.com/facebookresearch/fastText

License

fastText is BSD-licensed. Facebook also provide an additional patent grant.

Maven

<dependency>
    <groupId>com.github.sszuev</groupId>
    <artifactId>fasttext</artifactId>
    <version>1.0.0</version>
</dependency>

Tools

to build command-line tools use mvn package -Pmain or mvn package -Pextra

fasttext_java's Issues

Unsupervised cbow or skipgram of a medium sized corpus slow and memory hungry

Hi, trying it to generate embeddings of a medium sized corpus like http://mattmahoney.net/dc/enwik8.zip does not train with less 12g of memory, aside of that the speed is 9 words/sec resulting in thousands of hours, while C++ implementation does it in 10min.

Have you ever tried to train such a big corpus?

org.apache.hadoop.hdfs not exist

On line 76 of src/main/java/cc/fasttext/extra/ExraMain.java, org.apache.hadoop.hdfs does not exist

cannot find symbol: method forEach((f,i)->res[...]), f))

Hi there,

I run into a compile error:

FastText.java:860: error: cannot find symbol
map.forEach((f, i) -> res.put(dict.getLabel(i), f));
^
symbol: method forEach((f,i)->res[...]), f))
location: variable map of type TreeMultimap<Float,Integer>

Any hint? Thanks!

different predict results

i have model trained by c++ version fastText and the model is used for predicting in java version. But the two results is much different and the result of java version is obviously wrong.

Any plans for official release

Hi there,

Is there a roadmap for an non snapshot release into jitpack or other maven repo?

Cheers
George

OOV

Hello,

I want to use your version of fasttext in my poject, but i want to check before if this version handle the OOV words.

Thank in advance for your confirmation on how your code deals with the OOV during test phase
Regards

Optimize Dictionary#computeSubwords

which was changed in pull request #8

at least, there is duplicated #toString() call; that is not optimal

Wrong result when using -pretrainedVectors

Hi. Thanks for your work on keeping java version of fastText up-to-date.
I have an issue.
When I train model in supervised mode using pretrained vectors (wiki-news-300d-1M.vec) and then use it to classify sentences, it's result is quite different from original facebook implementation. This happens because you don't add words to dictionary in this loop in FastText.java (lines 1371-1386):

                for (int i = 0; i < n; i++) {
                    String line = in.readLine();
                    String word;
                    String[] array;
                    if (StringUtils.isEmpty(line) || (array = line.split(" ")).length == 0 || StringUtils.isEmpty(word = array[0])) {
                        throw new IllegalArgumentException("Wrong line: " + line);
                    }
                    List<Float> numbers = Arrays.stream(array).skip(1).limit(dim + 1).map(Float::parseFloat).collect(Collectors.toList());
                    if (numbers.size() < dim)
                        throw new IllegalArgumentException("Wrong numbers in the line: " + numbers.size() + ". Expected " + dim);
                    words.add(word);
                    for (int j = 0; j < numbers.size(); j++) {
                        mat.set(i, j, numbers.get(j));
                    }
                }

Could you please fix it?