Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge

Text classification with an RNN

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

This text classification tutorial trains a recurrent neural network on the IMDB large movie review dataset for sentiment analysis.

Setup

import numpy as np

import tensorflow_datasets as tfds
import tensorflow as tf

tfds.disable_progress_bar()

Import matplotlib and create a helper function to plot graphs:

import matplotlib.pyplot as plt


def plot_graphs(history, metric):
  plt.plot(history.history[metric])
  plt.plot(history.history['val_'+metric], '')
  plt.xlabel("Epochs")
  plt.ylabel(metric)
  plt.legend([metric, 'val_'+metric])

Setup input pipeline

The IMDB large movie review dataset is a binary classification dataset—all the reviews have either a positive or negative sentiment.

Download the dataset using TFDS. See the loading text tutorial for details on how to load this sort of data manually.

dataset, info = tfds.load('imdb_reviews', with_info=True,
                          as_supervised=True)
train_dataset, test_dataset = dataset['train'], dataset['test']

train_dataset.element_spec
(TensorSpec(shape=(), dtype=tf.string, name=None),
 TensorSpec(shape=(), dtype=tf.int64, name=None))

Initially this returns a dataset of (text, label pairs):

for example, label in train_dataset.take(1):
  print('text: ', example.numpy())
  print('label: ', label.numpy())
text:  b"This was an absolutely terrible movie. Don't be lured in by Christopher Walken or Michael Ironside. Both are great actors, but this must simply be their worst role in history. Even their great acting could not redeem this movie's ridiculous storyline. This movie is an early nineties US propaganda piece. The most pathetic scenes were those when the Columbian rebels were making their cases for revolutions. Maria Conchita Alonso appeared phony, and her pseudo-love affair with Walken was nothing but a pathetic emotional plug in a movie that was devoid of any real meaning. I am disappointed that there are movies like this, ruining actor's like Christopher Walken's good name. I could barely sit through it."
label:  0

Next shuffle the data for training and create batches of these (text, label) pairs:

BUFFER_SIZE = 10000
BATCH_SIZE = 64
train_dataset = train_dataset.shuffle(BUFFER_SIZE).batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
test_dataset = test_dataset.batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
for example, label in train_dataset.take(1):
  print('texts: ', example.numpy()[:3])
  print()
  print('labels: ', label.numpy()[:3])
texts:  [b'This is arguably the worst film I have ever seen, and I have quite an appetite for awful (and good) movies. It could (just) have managed a kind of adolescent humour if it had been consistently tongue-in-cheek --\xc3\xa0 la ROCKY HORROR PICTURE SHOW, which was really very funny. Other movies, like PLAN NINE FROM OUTER SPACE, manage to be funny while (apparently) trying to be serious. As to the acting, it looks like they rounded up brain-dead teenagers and asked them to ad-lib the whole production. Compared to them, Tom Cruise looks like Alec Guinness. There was one decent interpretation -- that of the older ghoul-busting broad on the motorcycle.'
 b"I saw this film in the worst possible circumstance. I'd already missed 15 minutes when I woke up to it on an international flight between Sydney and Seoul. I didn't know what I was watching, I thought maybe it was a movie of the week, but quickly became riveted by the performance of the lead actress playing a young woman who's child had been kidnapped. The premise started taking twist and turns I didn't see coming and by the end credits I was scrambling through the the in-flight guide to figure out what I had just watched. Turns out I was belatedly discovering Do-yeon Jeon who'd won Best Actress at Cannes for the role. I don't know if Secret Sunshine is typical of Korean cinema but I'm off to the DVD store to discover more."
 b"Hello. I am Paul Raddick, a.k.a. Panic Attack of WTAF, Channel 29 in Philadelphia. Let me tell you about this god awful movie that powered on Adam Sandler's film career but was digitized after a short time.<br /><br />Going Overboard is about an aspiring comedian played by Sandler who gets a job on a cruise ship and fails...or so I thought. Sandler encounters babes that like History of the World Part 1 and Rebound. The babes were supposed to be engaged, but, actually, they get executed by Sawtooth, the meanest cannibal the world has ever known. Adam Sandler fared bad in Going Overboard, but fared better in Big Daddy, Billy Madison, and Jen Leone's favorite, 50 First Dates. Man, Drew Barrymore was one hot chick. Spanglish is red hot, Going Overboard ain't Dooley squat! End of file."]

labels:  [0 1 0]

Create the text encoder

The raw text loaded by tfds needs to be processed before it can be used in a model. The simplest way to process text for training is using the TextVectorization layer. This layer has many capabilities, but this tutorial sticks to the default behavior.

Create the layer, and pass the dataset's text to the layer's .adapt method:

VOCAB_SIZE = 1000
encoder = tf.keras.layers.TextVectorization(
    max_tokens=VOCAB_SIZE)
encoder.adapt(train_dataset.map(lambda text, label: text))

The .adapt method sets the layer's vocabulary. Here are the first 20 tokens. After the padding and unknown tokens they're sorted by frequency:

vocab = np.array(encoder.get_vocabulary())
vocab[:20]
array(['', '[UNK]', 'the', 'and', 'a', 'of', 'to', 'is', 'in', 'it', 'i',
       'this', 'that', 'br', 'was', 'as', 'for', 'with', 'movie', 'but'],
      dtype='<U14')

Once the vocabulary is set, the layer can encode text into indices. The tensors of indices are 0-padded to the longest sequence in the batch (unless you set a fixed output_sequence_length):

encoded_example = encoder(example)[:3].numpy()
encoded_example
array([[ 11,   7,   1, ...,   0,   0,   0],
       [ 10, 208,  11, ...,   0,   0,   0],
       [  1,  10, 237, ...,   0,   0,   0]])

With the default settings, the process is not completely reversible. There are three main reasons for that:

  1. The default value for preprocessing.TextVectorization's standardize argument is "lower_and_strip_punctuation".
  2. The limited vocabulary size and lack of character-based fallback results in some unknown tokens.
for n in range(3):
  print("Original: ", example[n].numpy())
  print("Round-trip: ", " ".join(vocab[encoded_example[n]]))
  print()
Original:  b'This is arguably the worst film I have ever seen, and I have quite an appetite for awful (and good) movies. It could (just) have managed a kind of adolescent humour if it had been consistently tongue-in-cheek --\xc3\xa0 la ROCKY HORROR PICTURE SHOW, which was really very funny. Other movies, like PLAN NINE FROM OUTER SPACE, manage to be funny while (apparently) trying to be serious. As to the acting, it looks like they rounded up brain-dead teenagers and asked them to ad-lib the whole production. Compared to them, Tom Cruise looks like Alec Guinness. There was one decent interpretation -- that of the older ghoul-busting broad on the motorcycle.'
Round-trip:  this is [UNK] the worst film i have ever seen and i have quite an [UNK] for awful and good movies it could just have [UNK] a kind of [UNK] [UNK] if it had been [UNK] [UNK] [UNK] la [UNK] horror picture show which was really very funny other movies like [UNK] [UNK] from [UNK] space [UNK] to be funny while apparently trying to be serious as to the acting it looks like they [UNK] up [UNK] [UNK] and [UNK] them to [UNK] the whole production [UNK] to them tom [UNK] looks like [UNK] [UNK] there was one decent [UNK] that of the older [UNK] [UNK] on the [UNK]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               

Original:  b"I saw this film in the worst possible circumstance. I'd already missed 15 minutes when I woke up to it on an international flight between Sydney and Seoul. I didn't know what I was watching, I thought maybe it was a movie of the week, but quickly became riveted by the performance of the lead actress playing a young woman who's child had been kidnapped. The premise started taking twist and turns I didn't see coming and by the end credits I was scrambling through the the in-flight guide to figure out what I had just watched. Turns out I was belatedly discovering Do-yeon Jeon who'd won Best Actress at Cannes for the role. I don't know if Secret Sunshine is typical of Korean cinema but I'm off to the DVD store to discover more."
Round-trip:  i saw this film in the worst possible [UNK] id already [UNK] [UNK] minutes when i [UNK] up to it on an [UNK] [UNK] between [UNK] and [UNK] i didnt know what i was watching i thought maybe it was a movie of the [UNK] but quickly became [UNK] by the performance of the lead actress playing a young woman whos child had been [UNK] the premise started taking twist and turns i didnt see coming and by the end credits i was [UNK] through the the [UNK] [UNK] to figure out what i had just watched turns out i was [UNK] [UNK] [UNK] [UNK] [UNK] [UNK] best actress at [UNK] for the role i dont know if secret [UNK] is typical of [UNK] cinema but im off to the dvd [UNK] to [UNK] more                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     

Original:  b"Hello. I am Paul Raddick, a.k.a. Panic Attack of WTAF, Channel 29 in Philadelphia. Let me tell you about this god awful movie that powered on Adam Sandler's film career but was digitized after a short time.<br /><br />Going Overboard is about an aspiring comedian played by Sandler who gets a job on a cruise ship and fails...or so I thought. Sandler encounters babes that like History of the World Part 1 and Rebound. The babes were supposed to be engaged, but, actually, they get executed by Sawtooth, the meanest cannibal the world has ever known. Adam Sandler fared bad in Going Overboard, but fared better in Big Daddy, Billy Madison, and Jen Leone's favorite, 50 First Dates. Man, Drew Barrymore was one hot chick. Spanglish is red hot, Going Overboard ain't Dooley squat! End of file."
Round-trip:  [UNK] i am paul [UNK] [UNK] [UNK] [UNK] of [UNK] [UNK] [UNK] in [UNK] let me tell you about this god awful movie that [UNK] on [UNK] [UNK] film career but was [UNK] after a short [UNK] br going [UNK] is about an [UNK] [UNK] played by [UNK] who gets a job on a [UNK] [UNK] and [UNK] so i thought [UNK] [UNK] [UNK] that like history of the world part 1 and [UNK] the [UNK] were supposed to be [UNK] but actually they get [UNK] by [UNK] the [UNK] [UNK] the world has ever known [UNK] [UNK] [UNK] bad in going [UNK] but [UNK] better in big [UNK] [UNK] [UNK] and [UNK] [UNK] favorite [UNK] first [UNK] man [UNK] [UNK] was one hot [UNK] [UNK] is red hot going [UNK] [UNK] [UNK] [UNK] end of [UNK]

Create the model

A drawing of the information flow in the model

Above is a diagram of the model.

  1. This model can be build as a tf.keras.Sequential.

  2. The first layer is the encoder, which converts the text to a sequence of token indices.

  3. After the encoder is an embedding layer. An embedding layer stores one vector per word. When called, it converts the sequences of word indices to sequences of vectors. These vectors are trainable. After training (on enough data), words with similar meanings often have similar vectors.

    This index-lookup is much more efficient than the equivalent operation of passing a one-hot encoded vector through a tf.keras.layers.Dense layer.

  4. A recurrent neural network (RNN) processes sequence input by iterating through the elements. RNNs pass the outputs from one timestep to their input on the next timestep.

    The tf.keras.layers.Bidirectional wrapper can also be used with an RNN layer. This propagates the input forward and backwards through the RNN layer and then concatenates the final output.

    • The main advantage of a bidirectional RNN is that the signal from the beginning of the input doesn't need to be processed all the way through every timestep to affect the output.

    • The main disadvantage of a bidirectional RNN is that you can't efficiently stream predictions as words are being added to the end.

  5. After the RNN has converted the sequence to a single vector the two layers.Dense do some final processing, and convert from this vector representation to a single logit as the classification output.

The code to implement this is below:

model = tf.keras.Sequential([
    encoder,
    tf.keras.layers.Embedding(
        input_dim=len(encoder.get_vocabulary()),
        output_dim=64,
        # Use masking to handle the variable sequence lengths
        mask_zero=True),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(64)),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dense(1)
])

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

The embedding layer uses masking to handle the varying sequence-lengths. All the layers after the Embedding support masking:

print([layer.supports_masking for layer in model.layers])
[False, True, True, True, True]

To confirm that this works as expected, evaluate a sentence twice. First, alone so there's no padding to mask:

# predict on a sample text without padding.

sample_text = ('The movie was cool. The animation and the graphics '
               'were out of this world. I would recommend this movie.')
predictions = model.predict(np.array([sample_text]))
print(predictions[0])
[-0.00012211]

Now, evaluate it again in a batch with a longer sentence. The result should be identical:

# predict on a sample text with padding

padding = "the " * 2000
predictions = model.predict(np.array([sample_text, padding]))
print(predictions[0])
[-0.00012211]

Compile the Keras model to configure the training process:

model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
              optimizer=tf.keras.optimizers.Adam(1e-4),
              metrics=['accuracy'])

Train the model

history = model.fit(train_dataset, epochs=10,
                    validation_data=test_dataset,
                    validation_steps=30)
Epoch 1/10
391/391 [==============================] - 39s 84ms/step - loss: 0.6454 - accuracy: 0.5630 - val_loss: 0.4888 - val_accuracy: 0.7568
Epoch 2/10
391/391 [==============================] - 30s 75ms/step - loss: 0.3925 - accuracy: 0.8200 - val_loss: 0.3663 - val_accuracy: 0.8464
Epoch 3/10
391/391 [==============================] - 30s 75ms/step - loss: 0.3319 - accuracy: 0.8525 - val_loss: 0.3402 - val_accuracy: 0.8385
Epoch 4/10
391/391 [==============================] - 30s 75ms/step - loss: 0.3183 - accuracy: 0.8616 - val_loss: 0.3289 - val_accuracy: 0.8438
Epoch 5/10
391/391 [==============================] - 30s 75ms/step - loss: 0.3088 - accuracy: 0.8656 - val_loss: 0.3254 - val_accuracy: 0.8646
Epoch 6/10
391/391 [==============================] - 32s 81ms/step - loss: 0.3043 - accuracy: 0.8686 - val_loss: 0.3242 - val_accuracy: 0.8521
Epoch 7/10
391/391 [==============================] - 30s 76ms/step - loss: 0.3019 - accuracy: 0.8696 - val_loss: 0.3315 - val_accuracy: 0.8609
Epoch 8/10
391/391 [==============================] - 32s 76ms/step - loss: 0.3007 - accuracy: 0.8688 - val_loss: 0.3245 - val_accuracy: 0.8609
Epoch 9/10
391/391 [==============================] - 31s 77ms/step - loss: 0.2981 - accuracy: 0.8707 - val_loss: 0.3294 - val_accuracy: 0.8599
Epoch 10/10
391/391 [==============================] - 31s 78ms/step - loss: 0.2969 - accuracy: 0.8742 - val_loss: 0.3218 - val_accuracy: 0.8547
test_loss, test_acc = model.evaluate(test_dataset)

print('Test Loss:', test_loss)
print('Test Accuracy:', test_acc)
391/391 [==============================] - 15s 38ms/step - loss: 0.3185 - accuracy: 0.8582
Test Loss: 0.3184521794319153
Test Accuracy: 0.8581600189208984
plt.figure(figsize=(16, 8))
plt.subplot(1, 2, 1)
plot_graphs(history, 'accuracy')
plt.ylim(None, 1)
plt.subplot(1, 2, 2)
plot_graphs(history, 'loss')
plt.ylim(0, None)
(0.0, 0.6627909764647484)

png

Run a prediction on a new sentence:

If the prediction is >= 0.0, it is positive else it is negative.

sample_text = ('The movie was cool. The animation and the graphics '
               'were out of this world. I would recommend this movie.')
predictions = model.predict(np.array([sample_text]))

Stack two or more LSTM layers

Keras recurrent layers have two available modes that are controlled by the return_sequences constructor argument:

  • If False it returns only the last output for each input sequence (a 2D tensor of shape (batch_size, output_features)). This is the default, used in the previous model.

  • If True the full sequences of successive outputs for each timestep is returned (a 3D tensor of shape (batch_size, timesteps, output_features)).

Here is what the flow of information looks like with return_sequences=True:

layered_bidirectional

The interesting thing about using an RNN with return_sequences=True is that the output still has 3-axes, like the input, so it can be passed to another RNN layer, like this:

model = tf.keras.Sequential([
    encoder,
    tf.keras.layers.Embedding(len(encoder.get_vocabulary()), 64, mask_zero=True),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(64,  return_sequences=True)),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32)),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(1)
])
model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
              optimizer=tf.keras.optimizers.Adam(1e-4),
              metrics=['accuracy'])
history = model.fit(train_dataset, epochs=10,
                    validation_data=test_dataset,
                    validation_steps=30)
Epoch 1/10
391/391 [==============================] - 71s 149ms/step - loss: 0.6502 - accuracy: 0.5625 - val_loss: 0.4923 - val_accuracy: 0.7573
Epoch 2/10
391/391 [==============================] - 55s 138ms/step - loss: 0.4067 - accuracy: 0.8198 - val_loss: 0.3727 - val_accuracy: 0.8271
Epoch 3/10
391/391 [==============================] - 54s 136ms/step - loss: 0.3417 - accuracy: 0.8543 - val_loss: 0.3343 - val_accuracy: 0.8510
Epoch 4/10
391/391 [==============================] - 53s 134ms/step - loss: 0.3242 - accuracy: 0.8607 - val_loss: 0.3268 - val_accuracy: 0.8568
Epoch 5/10
391/391 [==============================] - 53s 135ms/step - loss: 0.3174 - accuracy: 0.8652 - val_loss: 0.3213 - val_accuracy: 0.8516
Epoch 6/10
391/391 [==============================] - 52s 132ms/step - loss: 0.3098 - accuracy: 0.8671 - val_loss: 0.3294 - val_accuracy: 0.8547
Epoch 7/10
391/391 [==============================] - 53s 134ms/step - loss: 0.3063 - accuracy: 0.8697 - val_loss: 0.3158 - val_accuracy: 0.8594
Epoch 8/10
391/391 [==============================] - 52s 132ms/step - loss: 0.3043 - accuracy: 0.8692 - val_loss: 0.3184 - val_accuracy: 0.8521
Epoch 9/10
391/391 [==============================] - 53s 133ms/step - loss: 0.3016 - accuracy: 0.8704 - val_loss: 0.3208 - val_accuracy: 0.8609
Epoch 10/10
391/391 [==============================] - 54s 136ms/step - loss: 0.2975 - accuracy: 0.8740 - val_loss: 0.3301 - val_accuracy: 0.8651
test_loss, test_acc = model.evaluate(test_dataset)

print('Test Loss:', test_loss)
print('Test Accuracy:', test_acc)
391/391 [==============================] - 26s 65ms/step - loss: 0.3293 - accuracy: 0.8646
Test Loss: 0.329334557056427
Test Accuracy: 0.8646399974822998
# predict on a sample text without padding.

sample_text = ('The movie was not good. The animation and the graphics '
               'were terrible. I would not recommend this movie.')
predictions = model.predict(np.array([sample_text]))
print(predictions)
[[-1.6796288]]
plt.figure(figsize=(16, 6))
plt.subplot(1, 2, 1)
plot_graphs(history, 'accuracy')
plt.subplot(1, 2, 2)
plot_graphs(history, 'loss')

png

Check out other existing recurrent layers such as GRU layers.

If you're interestied in building custom RNNs, see the Keras RNN Guide.