Google I/O is a wrap! Catch up on TensorFlow sessions View sessions

Text generation with an RNN

View on Run in Google Colab View source on GitHub Download notebook

This tutorial demonstrates how to generate text using a character-based RNN. You 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 the sample output when the model in this tutorial trained for 30 epochs, and started with the prompt "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', '')
Downloading data from
1115394/1115394 [==============================] - 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(f'Length of text: {len(text)} characters')
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(f'{len(vocab)} unique characters')
65 unique characters

Process the text

Vectorize the text

Before training, you need to convert the strings to a numerical representation.

The tf.keras.layers.StringLookup layer can convert each character into a numeric ID. It just needs the text to be split into tokens first.

example_texts = ['abcdefg', 'xyz']

chars = tf.strings.unicode_split(example_texts, input_encoding='UTF-8')
<tf.RaggedTensor [[b'a', b'b', b'c', b'd', b'e', b'f', b'g'], [b'x', b'y', b'z']]>

Now create the tf.keras.layers.StringLookup layer:

ids_from_chars = tf.keras.layers.StringLookup(
    vocabulary=list(vocab), mask_token=None)

It converts from tokens to character IDs:

ids = ids_from_chars(chars)
<tf.RaggedTensor [[40, 41, 42, 43, 44, 45, 46], [63, 64, 65]]>

Since the goal of this tutorial is to generate text, it will also be important to invert this representation and recover human-readable strings from it. For this you can use tf.keras.layers.StringLookup(..., invert=True).

chars_from_ids = tf.keras.layers.StringLookup(
    vocabulary=ids_from_chars.get_vocabulary(), invert=True, mask_token=None)

This layer recovers the characters from the vectors of IDs, and returns them as a tf.RaggedTensor of characters:

chars = chars_from_ids(ids)
<tf.RaggedTensor [[b'a', b'b', b'c', b'd', b'e', b'f', b'g'], [b'x', b'y', b'z']]>

You can tf.strings.reduce_join to join the characters back into strings.

tf.strings.reduce_join(chars, axis=-1).numpy()
array([b'abcdefg', b'xyz'], dtype=object)
def text_from_ids(ids):
  return tf.strings.reduce_join(chars_from_ids(ids), axis=-1)

The prediction task

Given a character, or a sequence of characters, what is the most probable next character? This is the task you're training the model to perform. The input to the model will be a sequence of characters, and you 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 function to convert the text vector into a stream of character indices.

all_ids = ids_from_chars(tf.strings.unicode_split(text, 'UTF-8'))
<tf.Tensor: shape=(1115394,), dtype=int64, numpy=array([19, 48, 57, ..., 46,  9,  1])>
ids_dataset =
for ids in ids_dataset.take(10):
seq_length = 100

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

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

for seq in sequences.take(1):
[b'F' b'i' b'r' b's' b't' b' ' b'C' b'i' b't' b'i' b'z' b'e' b'n' b':'
 b'\n' b'B' b'e' b'f' b'o' b'r' b'e' b' ' b'w' b'e' b' ' b'p' b'r' b'o'
 b'c' b'e' b'e' b'd' b' ' b'a' b'n' b'y' b' ' b'f' b'u' b'r' b't' b'h'
 b'e' b'r' b',' b' ' b'h' b'e' b'a' b'r' b' ' b'm' b'e' b' ' b's' b'p'
 b'e' b'a' b'k' b'.' b'\n' b'\n' b'A' b'l' b'l' b':' b'\n' b'S' b'p' b'e'
 b'a' b'k' b',' b' ' b's' b'p' b'e' b'a' b'k' b'.' b'\n' b'\n' b'F' b'i'
 b'r' b's' b't' b' ' b'C' b'i' b't' b'i' b'z' b'e' b'n' b':' b'\n' b'Y'
 b'o' b'u' b' '], shape=(101,), dtype=string)

It's easier to see what this is doing if you join the tokens back into strings:

for seq in sequences.take(5):
b'First Citizen:\nBefore we proceed any further, hear me speak.\n\nAll:\nSpeak, speak.\n\nFirst Citizen:\nYou '
b'are all resolved rather to die than to famish?\n\nAll:\nResolved. resolved.\n\nFirst Citizen:\nFirst, you k'
b"now Caius Marcius is chief enemy to the people.\n\nAll:\nWe know't, we know't.\n\nFirst Citizen:\nLet us ki"
b"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"
b'one: away, away!\n\nSecond Citizen:\nOne word, good citizens.\n\nFirst Citizen:\nWe are accounted poor citi'

For training you'll need a dataset of (input, label) pairs. Where input and label are sequences. At each time step the input is the current character and the label is the next character.

Here's a function that takes a sequence as input, duplicates, and shifts it to align the input and label for each timestep:

def split_input_target(sequence):
    input_text = sequence[:-1]
    target_text = sequence[1:]
    return input_text, target_text
(['T', 'e', 'n', 's', 'o', 'r', 'f', 'l', 'o'],
 ['e', 'n', 's', 'o', 'r', 'f', 'l', 'o', 'w'])
dataset =
for input_example, target_example in dataset.take(1):
    print("Input :", text_from_ids(input_example).numpy())
    print("Target:", text_from_ids(target_example).numpy())
Input : b'First Citizen:\nBefore we proceed any further, hear me speak.\n\nAll:\nSpeak, speak.\n\nFirst Citizen:\nYou'
Target: b'irst Citizen:\nBefore we proceed any further, hear me speak.\n\nAll:\nSpeak, speak.\n\nFirst Citizen:\nYou '

Create training batches

You used to split the text into manageable sequences. But before feeding this data into the model, you 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 = (
    .batch(BATCH_SIZE, drop_remainder=True)

<PrefetchDataset element_spec=(TensorSpec(shape=(64, 100), dtype=tf.int64, name=None), TensorSpec(shape=(64, 100), dtype=tf.int64, name=None))>

Build The Model

This section defines the model as a keras.Model subclass (For details see Making new Layers and Models via subclassing).

This model has three layers:

  • tf.keras.layers.Embedding: The input layer. A trainable lookup table that will map each character-ID to a vector with embedding_dim dimensions;
  • tf.keras.layers.GRU: A type of RNN with size units=rnn_units (You can also use an LSTM layer here.)
  • tf.keras.layers.Dense: The output layer, with vocab_size outputs. It outputs one logit for each character in the vocabulary. These are the log-likelihood of each character according to the model.
# Length of the vocabulary in StringLookup Layer
vocab_size = len(ids_from_chars.get_vocabulary())

# The embedding dimension
embedding_dim = 256

# Number of RNN units
rnn_units = 1024
class MyModel(tf.keras.Model):
  def __init__(self, vocab_size, embedding_dim, rnn_units):
    self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_dim)
    self.gru = tf.keras.layers.GRU(rnn_units,
    self.dense = tf.keras.layers.Dense(vocab_size)

  def call(self, inputs, states=None, return_state=False, training=False):
    x = inputs
    x = self.embedding(x, training=training)
    if states is None:
      states = self.gru.get_initial_state(x)
    x, states = self.gru(x, initial_state=states, training=training)
    x = self.dense(x, training=training)

    if return_state:
      return x, states
      return x
model = MyModel(

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

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, 66) # (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: "my_model"
 Layer (type)                Output Shape              Param #   
 embedding (Embedding)       multiple                  16896     
 gru (GRU)                   multiple                  3938304   
 dense (Dense)               multiple                  67650     
Total params: 4,022,850
Trainable params: 4,022,850
Non-trainable params: 0

To get actual predictions from the model you 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([49, 53,  2, 31, 35, 61, 51, 27, 22, 65, 40, 49,  2, 59,  4, 16, 43,
       35, 63, 65, 63, 61, 50,  5, 19, 21, 26, 30, 64,  1, 28, 46, 61,  5,
        2, 26, 28,  6, 20, 63, 13,  1, 46, 28, 10, 44, 57, 57, 19,  9, 48,
        1, 34, 23, 51, 62, 62, 38, 16,  0, 28, 11, 20, 43, 32, 55,  2, 32,
       59, 34, 26, 50, 41, 48, 40, 54, 36, 12, 33, 32,  5, 39,  7, 16,  7,
        3, 31, 23, 51, 23, 46, 40, 21, 28, 47, 37, 28,  2, 57, 18])

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

print("Input:\n", text_from_ids(input_example_batch[0]).numpy())
print("Next Char Predictions:\n", text_from_ids(sampled_indices).numpy())
 b'e did call me by my name:\nI urged our old acquaintance, and the drops\nThat we have bled together. Co'

Next Char Predictions:
 b"jn RVvlNIzaj t$CdVxzxvk&FHMQy\nOgv& MO'Gx?\ngO3errF.i\nUJlwwYC[UNK]O:GdSp StUMkbiaoW;TS&Z,C,!RJlJgaHOhXO rE"

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 your model returns logits, you need to set the from_logits flag.

loss = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
example_batch_mean_loss = loss(target_example_batch, example_batch_predictions)
print("Prediction shape: ", example_batch_predictions.shape, " # (batch_size, sequence_length, vocab_size)")
print("Mean loss:        ", example_batch_mean_loss)
Prediction shape:  (64, 100, 66)  # (batch_size, sequence_length, vocab_size)
Mean loss:         tf.Tensor(4.1881366, shape=(), dtype=float32)

A newly initialized model shouldn't be too sure of itself, the output logits should all have similar magnitudes. To confirm this you can check that the exponential of the mean loss is approximately equal to the vocabulary size. A much higher loss means the model is sure of its wrong answers, and is badly initialized:


Configure the training procedure using the tf.keras.Model.compile method. 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}")

checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(

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 =, epochs=EPOCHS, callbacks=[checkpoint_callback])
Epoch 1/20
172/172 [==============================] - 7s 25ms/step - loss: 2.7052
Epoch 2/20
172/172 [==============================] - 6s 24ms/step - loss: 1.9786
Epoch 3/20
172/172 [==============================] - 5s 24ms/step - loss: 1.7043
Epoch 4/20
172/172 [==============================] - 6s 24ms/step - loss: 1.5423
Epoch 5/20
172/172 [==============================] - 5s 24ms/step - loss: 1.4445
Epoch 6/20
172/172 [==============================] - 5s 24ms/step - loss: 1.3763
Epoch 7/20
172/172 [==============================] - 5s 24ms/step - loss: 1.3228
Epoch 8/20
172/172 [==============================] - 5s 25ms/step - loss: 1.2781
Epoch 9/20
172/172 [==============================] - 5s 25ms/step - loss: 1.2359
Epoch 10/20
172/172 [==============================] - 5s 24ms/step - loss: 1.1943
Epoch 11/20
172/172 [==============================] - 5s 24ms/step - loss: 1.1533
Epoch 12/20
172/172 [==============================] - 5s 24ms/step - loss: 1.1127
Epoch 13/20
172/172 [==============================] - 5s 25ms/step - loss: 1.0663
Epoch 14/20
172/172 [==============================] - 5s 24ms/step - loss: 1.0189
Epoch 15/20
172/172 [==============================] - 5s 24ms/step - loss: 0.9691
Epoch 16/20
172/172 [==============================] - 5s 25ms/step - loss: 0.9175
Epoch 17/20
172/172 [==============================] - 5s 25ms/step - loss: 0.8633
Epoch 18/20
172/172 [==============================] - 5s 24ms/step - loss: 0.8120
Epoch 19/20
172/172 [==============================] - 5s 25ms/step - loss: 0.7613
Epoch 20/20
172/172 [==============================] - 5s 25ms/step - loss: 0.7144

Generate text

The simplest way to generate text with this model is to run it in a loop, and keep track of the model's internal state as you execute it.

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

Each time you call the model you pass in some text and an internal state. The model returns a prediction for the next character and its new state. Pass the prediction and state back in to continue generating text.

The following makes a single step prediction:

class OneStep(tf.keras.Model):
  def __init__(self, model, chars_from_ids, ids_from_chars, temperature=1.0):
    self.temperature = temperature
    self.model = model
    self.chars_from_ids = chars_from_ids
    self.ids_from_chars = ids_from_chars

    # Create a mask to prevent "[UNK]" from being generated.
    skip_ids = self.ids_from_chars(['[UNK]'])[:, None]
    sparse_mask = tf.SparseTensor(
        # Put a -inf at each bad index.
        # Match the shape to the vocabulary
    self.prediction_mask = tf.sparse.to_dense(sparse_mask)

  def generate_one_step(self, inputs, states=None):
    # Convert strings to token IDs.
    input_chars = tf.strings.unicode_split(inputs, 'UTF-8')
    input_ids = self.ids_from_chars(input_chars).to_tensor()

    # Run the model.
    # predicted_logits.shape is [batch, char, next_char_logits]
    predicted_logits, states = self.model(inputs=input_ids, states=states,
    # Only use the last prediction.
    predicted_logits = predicted_logits[:, -1, :]
    predicted_logits = predicted_logits/self.temperature
    # Apply the prediction mask: prevent "[UNK]" from being generated.
    predicted_logits = predicted_logits + self.prediction_mask

    # Sample the output logits to generate token IDs.
    predicted_ids = tf.random.categorical(predicted_logits, num_samples=1)
    predicted_ids = tf.squeeze(predicted_ids, axis=-1)

    # Convert from token ids to characters
    predicted_chars = self.chars_from_ids(predicted_ids)

    # Return the characters and model state.
    return predicted_chars, states
one_step_model = OneStep(model, chars_from_ids, ids_from_chars)

Run it in a loop to generate some text. 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.

start = time.time()
states = None
next_char = tf.constant(['ROMEO:'])
result = [next_char]

for n in range(1000):
  next_char, states = one_step_model.generate_one_step(next_char, states=states)

result = tf.strings.join(result)
end = time.time()
print(result[0].numpy().decode('utf-8'), '\n\n' + '_'*80)
print('\nRun time:', end - start)
Why, sir, what think you, sir?

A dozen; shall I be deceased.
The enemy is parting with your general,
As bias should still combit them offend
That Montague is as devotions that did satisfied;
But not they are put your pleasure.

Peace, sing! do you must be all the law;
And overmuting Mercutio slain;
And stand betide that blows which wretched shame;
Which, I, that have been complaints me older hours.

What, marry, may shame, the forish priest-lay estimest you, sir,
Whom I will purchase with green limits o' the commons' ears!

Your enemy, she did contemn you.

And you, adieu.

My lord,
I hope, which, with thy hand,
And I will meet divine and mercy then.

Pardon me, madam: come as;
A cover in my power,
I have no nobler than came heaven with you.

And who, I'll never weep; I'll sing thy love.

Go, my country George!

Are you married, sweet; we will give you there.



Run time: 2.7961649894714355

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

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

If you want the model to generate text faster the easiest thing you can do is batch the text generation. In the example below the model generates 5 outputs in about the same time it took to generate 1 above.

start = time.time()
states = None
next_char = tf.constant(['ROMEO:', 'ROMEO:', 'ROMEO:', 'ROMEO:', 'ROMEO:'])
result = [next_char]

for n in range(1000):
  next_char, states = one_step_model.generate_one_step(next_char, states=states)

result = tf.strings.join(result)
end = time.time()
print(result, '\n\n' + '_'*80)
print('\nRun time:', end - start)
[b"ROMEO:\nGive me mine ear. Stan both your husband's sons,\nTYORY Lend Lord Northumberland,\nWhen he should be aught to bid him but a little,\nHimare the manife's head in our private.\n\nCARDIFILA:\nO Nerransa's death--children, man; a hundred shoe,\nA thousand men of nameting gentlewoman,\nGod bless the manner to the business.\n\nLUCENTIO:\nBut see how sweeting instructs here; Go, counsel thee;\nThat he is odes hare too rud away to-night:\nI have stoop'd my revenges were necessity\nThat it was wont to come to it.\n\nRICHARD:\nYour princely know your power\nThat I may enter in the wall!\nSo from their bellant in their helps up.\n\nSEBASTIAN:\nHer shakes of shame,\nWhen they shall starve with him.\n\nBRUTUS:\nCat leave us.\n\nMARIANA:\nI know not where;\nTherefore us, 'em and several pack; I first a goodly,\nMy heart with manages of my side,\nWith my repenate wretched Clarence' death?\nGo, single than that he is meril that e'er\ncannot earl'd: if it be sin, but severe was beseegh the best!\nA persecute more peace unwilling steeds"
 b"ROMEO:\nIt is; your son wishing horse: he can light with him,\nBut wherefore I find you all.\n\nDUKE VINCENTIO:\nWere it neat, not for my dismal thrancy:\nThere is another bear thy glory is but upon myself.\n\nNORTHUMBERLAND:\nWhate'er it be thinking up the garlies' enemy,\nWith outst tresple gentleman born.\n\nClown:\nConfusio's time; here's Romeo?\n\nSICINIUS:\nTrue, this letter would it pray?\n\nQUEEN MARGARET:\nMost loving, in my power?\n\nClown:\nIf it be speekens eyed no less honour.\n\nJULIET:\nThis fearful mean of this. Lords, for God's sooth,\nOur commonwealths in the common peace and the gardies!\nRives on, but being more moved, upon me! speak.\n\nMONTAGUE:\nBoth. Play, Plantagenet, he shall not and two thousand.\nWhat is, then muth he had but upon't, Isabel!\n\nANGELO:\nAnd wherefore stay her?\n\nARIEL:\nIs pite believe her to a good?\n\nHASTINGS:\nWhenever Mercutio and held, Geer, desired the sea,\nAnd let the survein of envoly of his good cannot lead.\n\nDUKE OF YORK:\nBut think'st thou with such consul?\n\nMessenger:\nStrik"
 b"ROMEO:\nFaith, he will not teer thee: love thee here.\nGo, hie thee further.\n\nNORTHUMBERLAND:\nYour spence could prain Bolingbroke, deliver 'S:\nNow, by my mother,--O, 'tis: men I am\nIn heart's dear, than the Lord Aumorous it looks.\nUS:\nNow the ground I live until a moving\nThat I have as free from time:\nYour court-hip in one wonder, if you please,\nLove agam Lucentio? What's your'e,\nTirtuse that hungerous all the father they\nUpon itself for Edward will that was your deed;\nOr mother window thee for thy trace belly;\nHis dearth, your father, is he believed.\nLet me hear me with some stulber,--there's this\nFlory-will be danged to make what you have been.\n\nBoatswain:\nI pray head, then your personace means.\n\nGLOUCESTER:\nGo, say 'em felt by: nay, then 'tis private--as penalty,\nyears, can streagful I do entreat her father:\nSome have I found a conqueror;\nAnd so still they tell her do I see thy dam,\nCan sympariabed to have saffied as smiles,\nWhere, my great gracious lord, with fiery soul\nThy vows work thus "
 b"ROMEO:\nI shall be satisfied, let me say 'I'll not be long;\nMy looks are full, sir, with four life, sir my\nsweet loys, by second matter to-day, who\nlately called against the Sight;\nAnd then she hath to save this queen.\n\nQUEEN ELIZABETH:\nOf thy foul ministers that are passling;\nMy heart to have itself is dead:' believe me, single!\nJoy to you, rumedous buried, she's cowardly,\nAnd not 'twere you hated to make foldiers. Speak, speak.\n\nSICINIUS:\nIt is your trave beloved? I must understand;\nAnd all my body's love themselves, to have\nMore thringing love she cricks to have them spurs;\nAnd now my best way hath appeared; somely is held\nCam not share therefore I will give him a pedperture.\n\nGRUMIO:\nStay, good: God forbid your captain hath not show\nThat Edward still-dread than feet or two, to play the\nshoe:\n'Tis most love since for a little permit;\nAnd the Lord Angelo, come hither.\n\nKING EDWARD IV:\nYou hear no hourly little car, is not hate till\nAttend thee here, but thou hast spaked\n'Tis more than every"
 b"ROMEO:\n\nJULIET:\nO, do not swear at, I am\nLucentio, or by Richard kindly,\nAnd paint inquirecy tre plain vexaltisman: I have stay'd\nTo seek the insinuate by thee;\nFor you Veronamn, yet I must wash long has!\n\nBISHOP OF ELY:\nNay, that is gain, and you, Sir Walter Richard murder'd,\nIf I lay down the subject swellows, in the deceit drowno\nWith all shed ever man shall request it.\nO pity, gentle hap, is dead to nothing?\nMy business is itlen of your cut proppery, rapier's purpose:\nMake where changed thousand is his son-in-law;\nAre you not prepared, pound to him and fly:\nWhat is it that the next heavy woman?\nAnd, for my brother: let him be calm.\n\nCORIOLANUS:\nWife, we have been out, or hope, if you must\nmy ears, O, meeting but one of your vows\nFor insured to full delight\nIn execution will half you arm, but\nset on her shed 'sir, now as he's anointed knots,--in--to us,\nShall quiet we be convey'd at especo me\nIn safegualmy subjects but that I may.\n\nGRUMIO:\nA beggars, pardon! madam, take my consent was wit"], shape=(5,), dtype=string) 


Run time: 2.880214214324951

Export the generator

This single-step model can easily be saved and restored, allowing you to use it anywhere a tf.saved_model is accepted., 'one_step')
one_step_reloaded = tf.saved_model.load('one_step')
WARNING:tensorflow:Skipping full serialization of Keras layer <__main__.OneStep object at 0x7fe57009a650>, because it is not built.
WARNING:absl:Found untraced functions such as gru_cell_layer_call_fn, gru_cell_layer_call_and_return_conditional_losses while saving (showing 2 of 2). These functions will not be directly callable after loading.
INFO:tensorflow:Assets written to: one_step/assets
INFO:tensorflow:Assets written to: one_step/assets
states = None
next_char = tf.constant(['ROMEO:'])
result = [next_char]

for n in range(100):
  next_char, states = one_step_reloaded.generate_one_step(next_char, states=states)

My liege, the hungry beck. What shall I do?
Say, thou liest; thou most loved currections and yourse

Advanced: Customized Training

The above training procedure is simple, but does not give you much control. It uses teacher-forcing which prevents bad predictions from being fed back to the model, so the model never learns to recover from mistakes.

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

The most important part of a custom training loop is the train step function.

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

The basic procedure is:

  1. Execute the model and calculate the loss under a tf.GradientTape.
  2. Calculate the updates and apply them to the model using the optimizer.
class CustomTraining(MyModel):
  def train_step(self, inputs):
      inputs, labels = inputs
      with tf.GradientTape() as tape:
          predictions = self(inputs, training=True)
          loss = self.loss(labels, predictions)
      grads = tape.gradient(loss, model.trainable_variables)
      self.optimizer.apply_gradients(zip(grads, model.trainable_variables))

      return {'loss': loss}

The above implementation of the train_step method follows Keras' train_step conventions. This is optional, but it allows you to change the behavior of the train step and still use keras' Model.compile and methods.

model = CustomTraining(
model.compile(optimizer = tf.keras.optimizers.Adam(),
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)), epochs=1)
172/172 [==============================] - 8s 24ms/step - loss: 2.7035
<keras.callbacks.History at 0x7fe0e3705850>

Or if you need more control, you can write your own complete custom training loop:


mean = tf.metrics.Mean()

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

    for (batch_n, (inp, target)) in enumerate(dataset):
        logs = model.train_step([inp, target])

        if batch_n % 50 == 0:
            template = f"Epoch {epoch+1} Batch {batch_n} Loss {logs['loss']:.4f}"

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

    print(f'Epoch {epoch+1} Loss: {mean.result().numpy():.4f}')
    print(f'Time taken for 1 epoch {time.time() - start:.2f} sec')

Epoch 1 Batch 0 Loss 2.1751
Epoch 1 Batch 50 Loss 2.0953
Epoch 1 Batch 100 Loss 1.9578
Epoch 1 Batch 150 Loss 1.8412

Epoch 1 Loss: 1.9787
Time taken for 1 epoch 6.26 sec
Epoch 2 Batch 0 Loss 1.8242
Epoch 2 Batch 50 Loss 1.7469
Epoch 2 Batch 100 Loss 1.6730
Epoch 2 Batch 150 Loss 1.6070

Epoch 2 Loss: 1.6991
Time taken for 1 epoch 5.40 sec
Epoch 3 Batch 0 Loss 1.5871
Epoch 3 Batch 50 Loss 1.5395
Epoch 3 Batch 100 Loss 1.5625
Epoch 3 Batch 150 Loss 1.5381

Epoch 3 Loss: 1.5412
Time taken for 1 epoch 5.29 sec
Epoch 4 Batch 0 Loss 1.5052
Epoch 4 Batch 50 Loss 1.4582
Epoch 4 Batch 100 Loss 1.4183
Epoch 4 Batch 150 Loss 1.4440

Epoch 4 Loss: 1.4446
Time taken for 1 epoch 5.43 sec
Epoch 5 Batch 0 Loss 1.3738
Epoch 5 Batch 50 Loss 1.3614
Epoch 5 Batch 100 Loss 1.3415
Epoch 5 Batch 150 Loss 1.3577

Epoch 5 Loss: 1.3781
Time taken for 1 epoch 5.57 sec
Epoch 6 Batch 0 Loss 1.2974
Epoch 6 Batch 50 Loss 1.3335
Epoch 6 Batch 100 Loss 1.2936
Epoch 6 Batch 150 Loss 1.3011

Epoch 6 Loss: 1.3254
Time taken for 1 epoch 5.32 sec
Epoch 7 Batch 0 Loss 1.2789
Epoch 7 Batch 50 Loss 1.3097
Epoch 7 Batch 100 Loss 1.3013
Epoch 7 Batch 150 Loss 1.3065

Epoch 7 Loss: 1.2815
Time taken for 1 epoch 5.32 sec
Epoch 8 Batch 0 Loss 1.2167
Epoch 8 Batch 50 Loss 1.2018
Epoch 8 Batch 100 Loss 1.2257
Epoch 8 Batch 150 Loss 1.2720

Epoch 8 Loss: 1.2402
Time taken for 1 epoch 5.33 sec
Epoch 9 Batch 0 Loss 1.2189
Epoch 9 Batch 50 Loss 1.2003
Epoch 9 Batch 100 Loss 1.2177
Epoch 9 Batch 150 Loss 1.1997

Epoch 9 Loss: 1.2004
Time taken for 1 epoch 5.27 sec
Epoch 10 Batch 0 Loss 1.1495
Epoch 10 Batch 50 Loss 1.1529
Epoch 10 Batch 100 Loss 1.1649
Epoch 10 Batch 150 Loss 1.1323

Epoch 10 Loss: 1.1599
Time taken for 1 epoch 5.60 sec