ML | Text Generation using Gated Recurrent Unit Networks

This article will demonstrate how to build a Text Generator by building a Gated Recurrent Unit Network. The conceptual procedure of training the network is to first feed the network a mapping of each character present in the text on which the network is training to a unique number. Each character is then hot-encoded into a vector which is the required format for the network.

The data for the described procedure is a collection of short and famous poems by famous poets and is in a .txt format. It can be downloaded from here.

Step 1: Importing the required libraries

filter_none

edit
close

play_arrow

link
brightness_4
code

from __future__ import absolute_import, division,
                       print_function, unicode_literals
  
import numpy as np
import tensorflow as tf
  
from keras.models import Sequential
from keras.layers import Dense, Activation
from keras.layers import LSTM
  
from keras.optimizers import RMSprop
  
from keras.callbacks import LambdaCallback
from keras.callbacks import ModelCheckpoint
from keras.callbacks import ReduceLROnPlateau
import random
import sys

chevron_right


Step 2: Loading the data into a string

filter_none

edit
close

play_arrow

link
brightness_4
code

# Changing the working location to the location of the text file
cd C:\Users\Dev\Desktop\Kaggle\Poems
  
# Reading the text file into a string
with open('poems.txt', 'r') as file:
    text = file.read()
  
# A preview of the text file    
print(text)

chevron_right


Step 3: Creating a mapping from each unique character in the text to a unique number

filter_none

edit
close

play_arrow

link
brightness_4
code

# Storing all the unique characters present in the text
vocabulary = sorted(list(set(text)))
  
# Creating dictionaries to map each character to an index
char_to_indices = dict((c, i) for i, c in enumerate(vocabulary))
indices_to_char = dict((i, c) for i, c in enumerate(vocabulary))
  
print(vocabulary)

chevron_right


Step 4: Pre-processing the data

filter_none

edit
close

play_arrow

link
brightness_4
code

# Dividing the text into subsequences of length max_length
# So that at each time step the next max_length characters 
# are fed into the network
max_length = 100
steps = 5
sentences = []
next_chars = []
for i in range(0, len(text) - max_length, steps):
    sentences.append(text[i: i + max_length])
    next_chars.append(text[i + max_length])
      
# Hot encoding each character into a boolean vector
  
# Initializing a matrix of boolean vectors with each column representing
# the hot encoded representation of the character
X = np.zeros((len(sentences), max_length, len(vocabulary)), dtype = np.bool)
y = np.zeros((len(sentences), len(vocabulary)), dtype = np.bool)
  
# Placing the value 1 at the appropriate position for each vector
# to complete the hot-encoding process
for i, sentence in enumerate(sentences):
    for t, char in enumerate(sentence):
        X[i, t, char_to_indices[char]] = 1
    y[i, char_to_indices[next_chars[i]]] = 1

chevron_right


Step 5: Building the GRU network

filter_none

edit
close

play_arrow

link
brightness_4
code

# Initializing the LSTM network
model = Sequential()
  
# Defining the cell type
model.add(GRU(128, input_shape =(max_length, len(vocabulary))))
  
# Defining the densely connected Neural Network layer
model.add(Dense(len(vocabulary)))
  
# Defining the activation function for the cell
model.add(Activation('softmax'))
  
# Defining the optimizing function
optimizer = RMSprop(lr = 0.01)
  
# Configuring the model for training
model.compile(loss ='categorical_crossentropy', optimizer = optimizer)

chevron_right


Step 6: Defining some helper functions which will be used during the training of the network

Note that the first two functions given below have been referred from the documentation of the official text generation example from the Keras team.

a) Helper function to sample the next character:

filter_none

edit
close

play_arrow

link
brightness_4
code

# Helper function to sample an index from a probability array
def sample_index(preds, temperature = 1.0):
# temperature determines the freedom the function has when generating text
  
    # Converting the predictions vector into a numpy array
    preds = np.asarray(preds).astype('float64')
  
    # Normalizing the predicitons array
    preds = np.log(preds) / temperature
    exp_preds = np.exp(preds)
    preds = exp_preds / np.sum(exp_preds)
  
    # The main sampling step. Creates an array of probablities signifying
    # the probability of each character to be the next character in the 
    # generated text
    probas = np.random.multinomial(1, preds, 1)
  
    # Returning the character with maximum probability to be the next character
    # in the generated text
    return np.argmax(probas)

chevron_right


b) Helper function to generate text after each epoch

filter_none

edit
close

play_arrow

link
brightness_4
code

# Helper function to generate text after the end of each epoch
def on_epoch_end(epoch, logs):
    print()
    print('----- Generating text after Epoch: % d' % epoch)
  
    # Choosing a random starting index for the text generation
    start_index = random.randint(0, len(text) - max_length - 1)
  
    # Sampling for different values of diversity
    for diversity in [0.2, 0.5, 1.0, 1.2]:
        print('----- diversity:', diversity)
  
        generated = ''
  
        # Seed sentence
        sentence = text[start_index: start_index + max_length]
  
        generated += sentence
        print('----- Generating with seed: "' + sentence + '"')
        sys.stdout.write(generated)
  
        for i in range(400):
            # Initializing the predicitons vector
            x_pred = np.zeros((1, max_length, len(vocabulary)))
  
            for t, char in enumerate(sentence):
                x_pred[0, t, char_to_indices[char]] = 1.
  
            # Making the predictions for the next character
            preds = model.predict(x_pred, verbose = 0)[0]
  
            # Getting the index of the most probable next character
            next_index = sample_index(preds, diversity)
  
            # Getting the most probable next character using the mapping built
            next_char = indices_to_char[next_index]
  
            # Building the generated text
            generated += next_char
            sentence = sentence[1:] + next_char
  
            sys.stdout.write(next_char)
            sys.stdout.flush()
        print()
  
# Defining a custom callback function to 
# describe the internal states of the network
print_callback = LambdaCallback(on_epoch_end = on_epoch_end)

chevron_right


c) Helper function to save the model after each epoch in which loss decreases

filter_none

edit
close

play_arrow

link
brightness_4
code

# Defining a helper function to save the model after each epoch
# in which the loss decreases
filepath = "weights.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor ='loss',
                             verbose = 1, save_best_only = True,
                             mode ='min')

chevron_right


d) Helper function to reduce the learning rate each time the learning plateaus

filter_none

edit
close

play_arrow

link
brightness_4
code

# Defining a helper function to reduce the learning rate each time
# the learning plateaus
reduce_alpha = ReduceLROnPlateau(monitor ='loss', factor = 0.2,
                              patience = 1, min_lr = 0.001)
callbacks = [print_callback, checkpoint, reduce_alpha]

chevron_right


Step 7: Training the GRU model

filter_none

edit
close

play_arrow

link
brightness_4
code

# Training the GRU model
model.fit(X, y, batch_size = 128, epochs = 30, callbacks = callbacks)

chevron_right


Step 8: Generating new and random text

filter_none

edit
close

play_arrow

link
brightness_4
code

def generate_text(length, diversity):
    # Get random starting text
    start_index = random.randint(0, len(text) - max_length - 1)
  
    # Defining the generated text
    generated = ''
    sentence = text[start_index: start_index + max_length]
    generated += sentence
  
    # Generating new text of given length
    for i in range(length):
  
            # Initializing the predicition vector
            x_pred = np.zeros((1, max_length, len(vocabulary)))
            for t, char in enumerate(sentence):
                x_pred[0, t, char_to_indices[char]] = 1.
  
            # Making the predicitons
            preds = model.predict(x_pred, verbose = 0)[0]
  
            # Getting the index of the next most probable index
            next_index = sample_index(preds, diversity)
  
            # Getting the most probable next character using the mapping built
            next_char = indices_to_char[next_index]
  
            # Generating new text
            generated += next_char
            sentence = sentence[1:] + next_char
    return generated
  
print(generate_text(500, 0.2))

chevron_right


Note: Although the output does not make much sense now, the output can be significantly improved by training the model for more epochs.



My Personal Notes arrow_drop_up

Check out this Author's contributed articles.

If you like GeeksforGeeks and would like to contribute, you can also write an article using contribute.geeksforgeeks.org or mail your article to contribute@geeksforgeeks.org. See your article appearing on the GeeksforGeeks main page and help other Geeks.

Please Improve this article if you find anything incorrect by clicking on the "Improve Article" button below.




Article Tags :
Practice Tags :


Be the First to upvote.


Please write to us at contribute@geeksforgeeks.org to report any issue with the above content.