Convert a FastText model to a form suitable for viewing in Tensorboard

fasttext_to_tensorboard.py

from tensorflow.contrib.tensorboard.plugins import projector
import tensorflow as tf
import numpy as np
import os

meta_file = "g2x_metadata.tsv"
output_path = "./projections"

# read embedding file into list and get the size
with open('./ft_model.vec', 'r') as embedding_file:
    embedding_content = embedding_file.readlines()
    embedding_content = [x.strip() for x in embedding_content] 

    num_lines = len(embedding_content) - 1 # skip the header
    num_dims = len(embedding_content[1].split()) - 1 # -1 because of the label column
    print("Detected dimensions:", num_lines, " X ", num_dims)

    placeholder = np.zeros((num_lines, num_dims))

    print(placeholder.shape)


    z = 0
    with open(os.path.join(output_path, meta_file), 'w') as file_metadata:

        i = 0
        for line in embedding_content[1:]:  # skip the header line
            values = line.split()
            raw_label = values[0]
            #print(label)
            col = 0
            for val in values[1:]: # skip the label
                placeholder[i][col] = val
                z = i + col
                col = col + 1
            i = i + 1

            if raw_label == '':
                file_metadata.write("<Empty Line>\n")
            else:
                label = raw_label
                file_metadata.write(label + "\n")

        print("z = ", z)

    # define the model without training
    sess = tf.InteractiveSession()

    embedding = tf.Variable(placeholder, trainable=False, name='g2x_metadata')
    tf.global_variables_initializer().run()

    saver = tf.train.Saver()
    writer = tf.summary.FileWriter(output_path, sess.graph)

    # adding into projector
    config = projector.ProjectorConfig()
    embed = config.embeddings.add()
    embed.tensor_name = 'g2x_metadata'
    embed.metadata_path = meta_file

    # Specify the width and height of a single thumbnail.
    projector.visualize_embeddings(writer, config)
    saver.save(sess, os.path.join(output_path, 'g2x_metadata.ckpt'))
    print('Num nodes: {}'.format(num_lines))
    print('Run `tensorboard --logdir={0}` to run visualize result on tensorboard'.format(output_path))

Hello,

My question may seem silly but I am desperate to know how to generate the tsv file. I already trained fasttext on a dataset of articles and I would like to plot my vectors in tf projector to show the proximity between patterns.

Could you explain to me how to do that please?

Many thanks

This script generates the tsv file for you right here: https://gist.github.com/nlothian/0cd4540389f7091717ece6f4b89b6604#file-fasttext_to_tensorboard-py-L59

It requires the trained vector file (called ft_model.vec) in the same directory as the script.

(This is really a pretty rough script I created for my own use. But I hope you find it helpful)