This article will demonstrate how to build a Text Generator by building a Recurrent Long Short Term Memory 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 was downloaded from Kaggle. This dataset contains the articles published in the New York Times from April 2017 to April 2018. separated according to the month of publication. The dataset is in the form of .csv file which contains the url of the published article along with other details. Any one random url was chosen for the training process and then on visiting this url, the text was copied into a text file and this text file was used for the training process.
Step 1: Importing the required libraries
Python3
from __future__ import absolute_import, division, print_function, unicode_literalsimport numpy as npimport tensorflow as tffrom keras.models import Sequentialfrom keras.layers import Dense, Activationfrom keras.layers import LSTMfrom keras.optimizers import RMSpropfrom keras.callbacks import LambdaCallbackfrom keras.callbacks import ModelCheckpointfrom keras.callbacks import ReduceLROnPlateauimport randomimport sys |
Step 2: Loading the data into a string
Python3
# Changing the working location to the location of the text filecd C:\Users\Dev\Desktop\Kaggle\New York Times# Reading the text file into a stringwith open('article1.txt', 'r') as file: text = file.read()# A preview of the text file print(text) |
Step 3: Creating a mapping from each unique character in the text to a unique number
Python3
# Storing all the unique characters present in the textvocabulary = sorted(list(set(text)))# Creating dictionaries to map each character to an indexchar_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) |
Step 4: Pre-processing the data
Python3
# Dividing the text into subsequences of length max_length# So that at each time step the next max_length characters # are fed into the networkmax_length = 100steps = 5sentences = []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 vectorX = np.zeros((len(sentences), max_length, len(vocabulary)), dtype = np.bool)y = np.zeros((len(sentences), len(vocabulary)), dtype = np.bool)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 |
Step 5: Building the LSTM network
Python3
# Building the LSTM network for the taskmodel = Sequential()model.add(LSTM(128, input_shape =(max_length, len(vocabulary))))model.add(Dense(len(vocabulary)))model.add(Activation('softmax'))optimizer = RMSprop(lr = 0.01)model.compile(loss ='categorical_crossentropy', optimizer = optimizer) |
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:
Python3
# Helper function to sample an index from a probability arraydef sample_index(preds, temperature = 1.0): preds = np.asarray(preds).astype('float64') preds = np.log(preds) / temperature exp_preds = np.exp(preds) preds = exp_preds / np.sum(exp_preds) probas = np.random.multinomial(1, preds, 1) return np.argmax(probas) |
b) Helper function to generate text after each epoch
Python3
# Helper function to generate text after the end of each epochdef on_epoch_end(epoch, logs): print() print('----- Generating text after Epoch: % d' % epoch) start_index = random.randint(0, len(text) - max_length - 1) for diversity in [0.2, 0.5, 1.0, 1.2]: print('----- diversity:', diversity) generated = '' sentence = text[start_index: start_index + max_length] generated += sentence print('----- Generating with seed: "' + sentence + '"') sys.stdout.write(generated) for i in range(400): x_pred = np.zeros((1, max_length, len(vocabulary))) for t, char in enumerate(sentence): x_pred[0, t, char_to_indices[char]] = 1. preds = model.predict(x_pred, verbose = 0)[0] next_index = sample_index(preds, diversity) next_char = indices_to_char[next_index] generated += next_char sentence = sentence[1:] + next_char sys.stdout.write(next_char) sys.stdout.flush() print()print_callback = LambdaCallback(on_epoch_end = on_epoch_end) |
c) Helper function to save the model after each epoch in which loss decreases
Python3
# Defining a helper function to save the model after each epoch# in which the loss decreasesfilepath = "weights.hdf5"checkpoint = ModelCheckpoint(filepath, monitor ='loss', verbose = 1, save_best_only = True, mode ='min') |
d) Helper function to reduce the learning rate each time the learning plateaus
Python3
# Defining a helper function to reduce the learning rate each time# the learning plateausreduce_alpha = ReduceLROnPlateau(monitor ='loss', factor = 0.2, patience = 1, min_lr = 0.001)callbacks = [print_callback, checkpoint, reduce_alpha] |
Step 7: Training the LSTM model
Python3
# Training the LSTM modelmodel.fit(X, y, batch_size = 128, epochs = 500, callbacks = callbacks) |
Step 8: Generating new and random text
Python3
# Defining a utility function to generate new and random text based on the# network's learningsdef generate_text(length, diversity): # Get random starting text start_index = random.randint(0, len(text) - max_length - 1) generated = '' sentence = text[start_index: start_index + max_length] generated += sentence for i in range(length): x_pred = np.zeros((1, max_length, len(vocabulary))) for t, char in enumerate(sentence): x_pred[0, t, char_to_indices[char]] = 1. preds = model.predict(x_pred, verbose = 0)[0] next_index = sample_index(preds, diversity) next_char = indices_to_char[next_index] generated += next_char sentence = sentence[1:] + next_char return generatedprint(generate_text(500, 0.2)) |
