Text generation with an RNN

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

This tutorial demonstrates how to generate text using a character-based RNN. We will work with a dataset of Shakespeare's writing from Andrej Karpathy's The Unreasonable Effectiveness of Recurrent Neural Networks. Given a sequence of characters from this data ("Shakespear"), train a model to predict the next character in the sequence ("e"). Longer sequences of text can be generated by calling the model repeatedly.

This tutorial includes runnable code implemented using tf.keras and eager execution. The following is sample output when the model in this tutorial trained for 30 epochs, and started with the string "Q":

I had thought thou hadst a Roman; for the oracle,
Thus by All bids the man against the word,
Which are so weak of care, by old care done;
Your children were in your holy love,
And the precipitation through the bleeding throne.

Marry, and will, my lord, to weep in such a one were prettiest;
Yet now I was adopted heir
Of the world's lamentable day,
To watch the next way with his father with his face?

The cause why then we are all resolved more sons.

O, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, it is no sin it should be dead,
And love and pale as any will to that word.

But how long have I heard the soul for this world,
And show his hands of life be proved to stand.

I say he look'd on, if I must be content
To stay him from the fatal of our country's bliss.
His lordship pluck'd from this sentence then for prey,
And then let us twain, being the moon,
were she such a case as fills m

While some of the sentences are grammatical, most do not make sense. The model has not learned the meaning of words, but consider:

  • The model is character-based. When training started, the model did not know how to spell an English word, or that words were even a unit of text.

  • The structure of the output resembles a play—blocks of text generally begin with a speaker name, in all capital letters similar to the dataset.

  • As demonstrated below, the model is trained on small batches of text (100 characters each), and is still able to generate a longer sequence of text with coherent structure.


Import TensorFlow and other libraries

import tensorflow as tf

import numpy as np
import os
import time

Download the Shakespeare dataset

Change the following line to run this code on your own data.

path_to_file = tf.keras.utils.get_file('shakespeare.txt', 'https://storage.googleapis.com/download.tensorflow.org/data/shakespeare.txt')
Downloading data from https://storage.googleapis.com/download.tensorflow.org/data/shakespeare.txt
1122304/1115394 [==============================] - 0s 0us/step

Read the data

First, look in the text:

# Read, then decode for py2 compat.
text = open(path_to_file, 'rb').read().decode(encoding='utf-8')
# length of text is the number of characters in it
print ('Length of text: {} characters'.format(len(text)))
Length of text: 1115394 characters

# Take a look at the first 250 characters in text
First Citizen:
Before we proceed any further, hear me speak.

Speak, speak.

First Citizen:
You are all resolved rather to die than to famish?

Resolved. resolved.

First Citizen:
First, you know Caius Marcius is chief enemy to the people.

# The unique characters in the file
vocab = sorted(set(text))
print ('{} unique characters'.format(len(vocab)))
65 unique characters

Process the text

Vectorize the text

Before training, we need to map strings to a numerical representation. Create two lookup tables: one mapping characters to numbers, and another for numbers to characters.

# Creating a mapping from unique characters to indices
char2idx = {u:i for i, u in enumerate(vocab)}
idx2char = np.array(vocab)

text_as_int = np.array([char2idx[c] for c in text])

Now we have an integer representation for each character. Notice that we mapped the character as indexes from 0 to len(unique).

for char,_ in zip(char2idx, range(20)):
    print('  {:4s}: {:3d},'.format(repr(char), char2idx[char]))
print('  ...\n}')
  '\n':   0,
  ' ' :   1,
  '!' :   2,
  '$' :   3,
  '&' :   4,
  "'" :   5,
  ',' :   6,
  '-' :   7,
  '.' :   8,
  '3' :   9,
  ':' :  10,
  ';' :  11,
  '?' :  12,
  'A' :  13,
  'B' :  14,
  'C' :  15,
  'D' :  16,
  'E' :  17,
  'F' :  18,
  'G' :  19,

# Show how the first 13 characters from the text are mapped to integers
print ('{} ---- characters mapped to int ---- > {}'.format(repr(text[:13]), text_as_int[:13]))
'First Citizen' ---- characters mapped to int ---- > [18 47 56 57 58  1 15 47 58 47 64 43 52]

The prediction task

Given a character, or a sequence of characters, what is the most probable next character? This is the task we're training the model to perform. The input to the model will be a sequence of characters, and we train the model to predict the output—the following character at each time step.

Since RNNs maintain an internal state that depends on the previously seen elements, given all the characters computed until this moment, what is the next character?

Create training examples and targets

Next divide the text into example sequences. Each input sequence will contain seq_length characters from the text.

For each input sequence, the corresponding targets contain the same length of text, except shifted one character to the right.

So break the text into chunks of seq_length+1. For example, say seq_length is 4 and our text is "Hello". The input sequence would be "Hell", and the target sequence "ello".

To do this first use the tf.data.Dataset.from_tensor_slices function to convert the text vector into a stream of character indices.

# The maximum length sentence we want for a single input in characters
seq_length = 100
examples_per_epoch = len(text)//(seq_length+1)

# Create training examples / targets
char_dataset = tf.data.Dataset.from_tensor_slices(text_as_int)

for i in char_dataset.take(5):

The batch method lets us easily convert these individual characters to sequences of the desired size.

sequences = char_dataset.batch(seq_length+1, drop_remainder=True)

for item in sequences.take(5):
'First Citizen:\nBefore we proceed any further, hear me speak.\n\nAll:\nSpeak, speak.\n\nFirst Citizen:\nYou '
'are all resolved rather to die than to famish?\n\nAll:\nResolved. resolved.\n\nFirst Citizen:\nFirst, you k'
"now Caius Marcius is chief enemy to the people.\n\nAll:\nWe know't, we know't.\n\nFirst Citizen:\nLet us ki"
"ll him, and we'll have corn at our own price.\nIs't a verdict?\n\nAll:\nNo more talking on't; let it be d"
'one: away, away!\n\nSecond Citizen:\nOne word, good citizens.\n\nFirst Citizen:\nWe are accounted poor citi'

For each sequence, duplicate and shift it to form the input and target text by using the map method to apply a simple function to each batch:

def split_input_target(chunk):
    input_text = chunk[:-1]
    target_text = chunk[1:]
    return input_text, target_text

dataset = sequences.map(split_input_target)

Print the first examples input and target values:

for input_example, target_example in  dataset.take(1):
  print ('Input data: ', repr(''.join(idx2char[input_example.numpy()])))
  print ('Target data:', repr(''.join(idx2char[target_example.numpy()])))
Input data:  'First Citizen:\nBefore we proceed any further, hear me speak.\n\nAll:\nSpeak, speak.\n\nFirst Citizen:\nYou'
Target data: 'irst Citizen:\nBefore we proceed any further, hear me speak.\n\nAll:\nSpeak, speak.\n\nFirst Citizen:\nYou '

Each index of these vectors are processed as one time step. For the input at time step 0, the model receives the index for "F" and trys to predict the index for "i" as the next character. At the next timestep, it does the same thing but the RNN considers the previous step context in addition to the current input character.

for i, (input_idx, target_idx) in enumerate(zip(input_example[:5], target_example[:5])):
    print("Step {:4d}".format(i))
    print("  input: {} ({:s})".format(input_idx, repr(idx2char[input_idx])))
    print("  expected output: {} ({:s})".format(target_idx, repr(idx2char[target_idx])))
Step    0
  input: 18 ('F')
  expected output: 47 ('i')
Step    1
  input: 47 ('i')
  expected output: 56 ('r')
Step    2
  input: 56 ('r')
  expected output: 57 ('s')
Step    3
  input: 57 ('s')
  expected output: 58 ('t')
Step    4
  input: 58 ('t')
  expected output: 1 (' ')

Create training batches

We used tf.data to split the text into manageable sequences. But before feeding this data into the model, we need to shuffle the data and pack it into batches.

# Batch size

# Buffer size to shuffle the dataset
# (TF data is designed to work with possibly infinite sequences,
# so it doesn't attempt to shuffle the entire sequence in memory. Instead,
# it maintains a buffer in which it shuffles elements).

dataset = dataset.shuffle(BUFFER_SIZE).batch(BATCH_SIZE, drop_remainder=True)

<BatchDataset shapes: ((64, 100), (64, 100)), types: (tf.int64, tf.int64)>

Build The Model

Use tf.keras.Sequential to define the model. For this simple example three layers are used to define our model:

# Length of the vocabulary in chars
vocab_size = len(vocab)

# The embedding dimension
embedding_dim = 256

# Number of RNN units
rnn_units = 1024
def build_model(vocab_size, embedding_dim, rnn_units, batch_size):
  model = tf.keras.Sequential([
    tf.keras.layers.Embedding(vocab_size, embedding_dim,
                              batch_input_shape=[batch_size, None]),
  return model
model = build_model(
  vocab_size = len(vocab),

For each character the model looks up the embedding, runs the GRU one timestep with the embedding as input, and applies the dense layer to generate logits predicting the log-likelihood of the next character:

A drawing of the data passing through the model

Please note that we choose to Keras sequential model here since all the layers in the model only have single input and produce single output. In case you want to retrieve and reuse the states from stateful RNN layer, you might want to build your model with Keras functional API or model subclassing. Please check Keras RNN guide for more details.

Try the model

Now run the model to see that it behaves as expected.

First check the shape of the output:

for input_example_batch, target_example_batch in dataset.take(1):
  example_batch_predictions = model(input_example_batch)
  print(example_batch_predictions.shape, "# (batch_size, sequence_length, vocab_size)")
(64, 100, 65) # (batch_size, sequence_length, vocab_size)

In the above example the sequence length of the input is 100 but the model can be run on inputs of any length:

Model: "sequential"
Layer (type)                 Output Shape              Param #   
embedding (Embedding)        (64, None, 256)           16640     
gru (GRU)                    (64, None, 1024)          3938304   
dense (Dense)                (64, None, 65)            66625     
Total params: 4,021,569
Trainable params: 4,021,569
Non-trainable params: 0

To get actual predictions from the model we need to sample from the output distribution, to get actual character indices. This distribution is defined by the logits over the character vocabulary.

Try it for the first example in the batch:

sampled_indices = tf.random.categorical(example_batch_predictions[0], num_samples=1)
sampled_indices = tf.squeeze(sampled_indices,axis=-1).numpy()

This gives us, at each timestep, a prediction of the next character index:

array([ 3, 21, 53,  7, 45,  5, 64, 26, 37, 39, 17, 34, 25, 30, 29, 56, 25,
       22, 27, 62, 60, 59, 29, 51,  6, 27, 55, 39,  9, 46, 37, 19, 31, 38,
       63, 56,  6, 41, 40, 47,  5, 15, 22, 55, 24, 24, 36, 24,  3, 50, 46,
       47, 40,  7,  2, 62, 44, 27, 59, 12, 15, 61, 35,  2, 36, 60, 51, 64,
       40, 26, 31, 21, 46, 48, 33, 26, 31, 19, 20, 18, 64, 36, 61, 35, 21,
       51, 56, 49,  0, 41, 51,  8, 32, 10, 45, 29, 10, 23, 54, 49])

Decode these to see the text predicted by this untrained model:

print("Input: \n", repr("".join(idx2char[input_example_batch[0]])))
print("Next Char Predictions: \n", repr("".join(idx2char[sampled_indices ])))
 'judgment,\nTo fail in the disposing of those chances\nWhich he was lord of; or whether nature,\nNot to '

Next Char Predictions: 

Train the model

At this point the problem can be treated as a standard classification problem. Given the previous RNN state, and the input this time step, predict the class of the next character.

Attach an optimizer, and a loss function

The standard tf.keras.losses.sparse_categorical_crossentropy loss function works in this case because it is applied across the last dimension of the predictions.

Because our model returns logits, we need to set the from_logits flag.

def loss(labels, logits):
  return tf.keras.losses.sparse_categorical_crossentropy(labels, logits, from_logits=True)

example_batch_loss  = loss(target_example_batch, example_batch_predictions)
print("Prediction shape: ", example_batch_predictions.shape, " # (batch_size, sequence_length, vocab_size)")
print("scalar_loss:      ", example_batch_loss.numpy().mean())
Prediction shape:  (64, 100, 65)  # (batch_size, sequence_length, vocab_size)
scalar_loss:       4.173578

Configure the training procedure using the tf.keras.Model.compile method. We'll use tf.keras.optimizers.Adam with default arguments and the loss function.

model.compile(optimizer='adam', loss=loss)

Configure checkpoints

Use a tf.keras.callbacks.ModelCheckpoint to ensure that checkpoints are saved during training:

# Directory where the checkpoints will be saved
checkpoint_dir = './training_checkpoints'
# Name of the checkpoint files
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt_{epoch}")


Execute the training

To keep training time reasonable, use 10 epochs to train the model. In Colab, set the runtime to GPU for faster training.

history = model.fit(dataset, epochs=EPOCHS, callbacks=[checkpoint_callback])
Epoch 1/10
172/172 [==============================] - 4s 26ms/step - loss: 2.6754
Epoch 2/10
172/172 [==============================] - 4s 26ms/step - loss: 1.9630
Epoch 3/10
172/172 [==============================] - 4s 26ms/step - loss: 1.6974
Epoch 4/10
172/172 [==============================] - 4s 26ms/step - loss: 1.5468
Epoch 5/10
172/172 [==============================] - 4s 26ms/step - loss: 1.4582
Epoch 6/10
172/172 [==============================] - 4s 26ms/step - loss: 1.3984
Epoch 7/10
172/172 [==============================] - 4s 26ms/step - loss: 1.3528
Epoch 8/10
172/172 [==============================] - 4s 26ms/step - loss: 1.3146
Epoch 9/10
172/172 [==============================] - 4s 26ms/step - loss: 1.2793
Epoch 10/10
172/172 [==============================] - 4s 26ms/step - loss: 1.2477

Generate text

Restore the latest checkpoint

To keep this prediction step simple, use a batch size of 1.

Because of the way the RNN state is passed from timestep to timestep, the model only accepts a fixed batch size once built.

To run the model with a different batch_size, we need to rebuild the model and restore the weights from the checkpoint.

model = build_model(vocab_size, embedding_dim, rnn_units, batch_size=1)


model.build(tf.TensorShape([1, None]))
Model: "sequential_1"
Layer (type)                 Output Shape              Param #   
embedding_1 (Embedding)      (1, None, 256)            16640     
gru_1 (GRU)                  (1, None, 1024)           3938304   
dense_1 (Dense)              (1, None, 65)             66625     
Total params: 4,021,569
Trainable params: 4,021,569
Non-trainable params: 0

The prediction loop

The following code block generates the text:

  • It Starts by choosing a start string, initializing the RNN state and setting the number of characters to generate.

  • Get the prediction distribution of the next character using the start string and the RNN state.

  • Then, use a categorical distribution to calculate the index of the predicted character. Use this predicted character as our next input to the model.

  • The RNN state returned by the model is fed back into the model so that it now has more context, instead than only one character. After predicting the next character, the modified RNN states are again fed back into the model, which is how it learns as it gets more context from the previously predicted characters.

To generate text the model's output is fed back to the input

Looking at the generated text, you'll see the model knows when to capitalize, make paragraphs and imitates a Shakespeare-like writing vocabulary. With the small number of training epochs, it has not yet learned to form coherent sentences.

def generate_text(model, start_string):
  # Evaluation step (generating text using the learned model)

  # Number of characters to generate
  num_generate = 1000

  # Converting our start string to numbers (vectorizing)
  input_eval = [char2idx[s] for s in start_string]
  input_eval = tf.expand_dims(input_eval, 0)

  # Empty string to store our results
  text_generated = []

  # Low temperatures results in more predictable text.
  # Higher temperatures results in more surprising text.
  # Experiment to find the best setting.
  temperature = 1.0

  # Here batch size == 1
  for i in range(num_generate):
      predictions = model(input_eval)
      # remove the batch dimension
      predictions = tf.squeeze(predictions, 0)

      # using a categorical distribution to predict the character returned by the model
      predictions = predictions / temperature
      predicted_id = tf.random.categorical(predictions, num_samples=1)[-1,0].numpy()

      # We pass the predicted character as the next input to the model
      # along with the previous hidden state
      input_eval = tf.expand_dims([predicted_id], 0)


  return (start_string + ''.join(text_generated))
print(generate_text(model, start_string=u"ROMEO: "))
ROMEO: I remember then;
What says he terts his, by God, thou wilt be
the larg: but s your honour than follow to her?

LADY Angelo, to Savultior home;
For 'tis I heard a moon of place;
Good entertainment there not
A puasts of trangs? Wilt thou tear the last,
Like it o'er me a part of little;
It will the Edrook the court?

Gay have I should require to the best,
And shaked and harpour the deathfore resesses the occks?
Nor to lay.

Good and my too deceivedem and faintly:
I, no, as I have a sweet for, cry 'say we wall:
For this afternoone how here pardon me:
If the Rome all miseries, who ever Isabel!
Was he patient, hast thou were as if the fire of truth, the world to lost it,
From court had he is, ten despisition at,
Provoking, a kiss, a prize of man,
For defended droped as the like?

Now, sir!

Then stay where I fire, sirrah, Or one Wantimore,
Erechment so. Everalt there
My death like us and sake to a pine. Come,
Is so disprove a thorefood, and he'll p

The easiest thing you can do to improve the results it to train it for longer (try EPOCHS=30).

You can also experiment with a different start string, or try adding another RNN layer to improve the model's accuracy, or adjusting the temperature parameter to generate more or less random predictions.

Advanced: Customized Training

The above training procedure is simple, but does not give you much control.

So now that you've seen how to run the model manually let's unpack the training loop, and implement it ourselves. This gives a starting point if, for example, to implement curriculum learning to help stabilize the model's open-loop output.

We will use tf.GradientTape to track the gradients. You can learn more about this approach by reading the eager execution guide.

The procedure works as follows:

  • First, initialize the RNN state. We do this by calling the tf.keras.Model.reset_states method.

  • Next, iterate over the dataset (batch by batch) and calculate the predictions associated with each.

  • Open a tf.GradientTape, and calculate the predictions and loss in that context.

  • Calculate the gradients of the loss with respect to the model variables using the tf.GradientTape.grads method.

  • Finally, take a step downwards by using the optimizer's tf.train.Optimizer.apply_gradients method.

model = build_model(
  vocab_size = len(vocab),
optimizer = tf.keras.optimizers.Adam()
def train_step(inp, target):
  with tf.GradientTape() as tape:
    predictions = model(inp)
    loss = tf.reduce_mean(
            target, predictions, from_logits=True))
  grads = tape.gradient(loss, model.trainable_variables)
  optimizer.apply_gradients(zip(grads, model.trainable_variables))

  return loss
# Training step

for epoch in range(EPOCHS):
  start = time.time()

  # initializing the hidden state at the start of every epoch
  # initally hidden is None
  hidden = model.reset_states()

  for (batch_n, (inp, target)) in enumerate(dataset):
    loss = train_step(inp, target)

    if batch_n % 100 == 0:
      template = 'Epoch {} Batch {} Loss {}'
      print(template.format(epoch+1, batch_n, loss))

  # saving (checkpoint) the model every 5 epochs
  if (epoch + 1) % 5 == 0:

  print ('Epoch {} Loss {:.4f}'.format(epoch+1, loss))
  print ('Time taken for 1 epoch {} sec\n'.format(time.time() - start))

Epoch 1 Batch 0 Loss 4.1742167472839355
Epoch 1 Batch 100 Loss 2.35125470161438
Epoch 1 Loss 2.1512
Time taken for 1 epoch 6.28100323677063 sec

Epoch 2 Batch 0 Loss 2.1255431175231934
Epoch 2 Batch 100 Loss 1.9002211093902588
Epoch 2 Loss 1.8136
Time taken for 1 epoch 5.1717822551727295 sec

Epoch 3 Batch 0 Loss 1.7405412197113037
Epoch 3 Batch 100 Loss 1.7021349668502808
Epoch 3 Loss 1.6206
Time taken for 1 epoch 5.088864088058472 sec

Epoch 4 Batch 0 Loss 1.569535732269287
Epoch 4 Batch 100 Loss 1.5282319784164429
Epoch 4 Loss 1.4413
Time taken for 1 epoch 5.10555362701416 sec

Epoch 5 Batch 0 Loss 1.4405953884124756
Epoch 5 Batch 100 Loss 1.401581883430481
Epoch 5 Loss 1.4063
Time taken for 1 epoch 5.196904182434082 sec

Epoch 6 Batch 0 Loss 1.3883678913116455
Epoch 6 Batch 100 Loss 1.3651986122131348
Epoch 6 Loss 1.4065
Time taken for 1 epoch 5.089401483535767 sec

Epoch 7 Batch 0 Loss 1.313085675239563
Epoch 7 Batch 100 Loss 1.3296411037445068
Epoch 7 Loss 1.2981
Time taken for 1 epoch 5.165014982223511 sec

Epoch 8 Batch 0 Loss 1.2713452577590942
Epoch 8 Batch 100 Loss 1.316956639289856
Epoch 8 Loss 1.3161
Time taken for 1 epoch 5.1581645011901855 sec

Epoch 9 Batch 0 Loss 1.2387388944625854
Epoch 9 Batch 100 Loss 1.2668814659118652
Epoch 9 Loss 1.3104
Time taken for 1 epoch 5.157152414321899 sec

Epoch 10 Batch 0 Loss 1.1915398836135864
Epoch 10 Batch 100 Loss 1.2359037399291992
Epoch 10 Loss 1.2784
Time taken for 1 epoch 5.1973793506622314 sec