Simple embedding based text classifier inspired by fastText, implemented in tensorflow
Switch branches/tags
Nothing to show
Clone or download
Latest commit 8120b5f Jan 12, 2018

FastText in Tensorflow

This project is based on the ideas in Facebook's FastText but implemented in Tensorflow. However, it is not an exact replica of fastText.

Classification is done by embedding each word, taking the mean embedding over the full text and classifying that using a linear classifier. The embedding is trained with the classifier. You can also specify to use 2+ character ngrams. These ngrams get hashed then embedded in a similar manner to the orginal words. Note, ngrams make training much slower but only make marginal improvements in performance, at least in English.

I may implement skipgram and cbow training later. Or preloading embedding tables.

<< Still WIP >>

You can use Horovod to distribute training across multiple GPUs, on one or multiple servers. See usage section below.

FastText Language Identification

I have added utilities to train a classifier to detect languages, as described in Fast and Accurate Language Identification using FastText

See usage below. It basically works in the same way as default usage.


  • classification of text using word embeddings
  • char ngrams, hashed to n bins
  • training and prediction program
  • serve models on tensorflow serving
  • preprocess facebook format, or text input into tensorflow records

Not Implemented:

  • separate word vector training (though can export embeddings)
  • heirarchical softmax.
  • quantize models (supported by tensorflow, but I haven't tried it yet)


The following are examples of how to use the applications. Get full help with --help option on any of the programs.

To transform input data into tensorflow Example format: --facebook_input=queries.txt --output_dir=. --ngrams=2,3,4

Or, using a text file with one example per line with an extra file for labels: --text_input=queries.txt --labels=labels.txt --output_dir=.

To train a text classifier: \
  --train_records=queries.tfrecords \
  --eval_records=queries.tfrecords \
  --label_file=labels.txt \
  --vocab_file=vocab.txt \
  --model_dir=model \

To predict classifications for text, use a saved_model from classifier. --export_dir stores a saved model in a numbered directory below export_dir. Pass this directory to the following to use that model for predictions:
  --text="some text to classify"

To export the embedding layer you can export from predictor. Note, this will only be the text embedding, not the ngram embeddings.
  --text="some text to classify"

Use the provided script to train easily: path-to-data-directory

Language Identification

To implement something similar to the method described in Fast and Accurate Language Identification using FastText you need to download the data: [datadir]

You can then process the training and validation data using and as described above.

There is a utility script to do this for you: datadir

It reaches about 96% accuracy using word embeddings and this increases to nearly 99% when adding --ngrams=2,3,4

Distributed Training

You can run training across multiple GPUs either on one or multiple servers. To do so you need to install MPI and Horovod then add the --horovod option. It runs very close to the GPU multiple in terms of performance. I.e. if you have 2 GPUs on your server, it should run close to 2x the speed.

mpirun -np $NUM_GPUS python \
  --horovod \
  --train_records=queries.tfrecords \
  --eval_records=queries.tfrecords \
  --label_file=labels.txt \
  --vocab_file=vocab.txt \
  --model_dir=model \

The training script has this option added:

Tensorflow Serving

As well as using to run a saved model to provide predictions, it is easy to serve a saved model using Tensorflow Serving with a client server setup. There is a supplied simple rpc client ( that provides predictions by using tensorflow server.

First make sure you install the tensorflow serving binaries. Instructions are here.

You then serve the latest saved model by supplying the base export directory where you exported saved models to. This directory will contain the numbered model directories:

tensorflow_model_server --port=9000 --model_base_path=model

Now you can make requests to the server using gRPC calls. An example simple client is provided in --text="Some text to classify"

Facebook Examples


You can compare with Facebook's fastText by running similar examples to what's provided in their repository.