In [0]:
#@title Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
In [0]:
#@title MIT License
#
# Copyright (c) 2017 François Chollet
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.
This notebook classifies movie reviews as positive or negative using the text of the review. This is an example of binary—or two-class—classification, an important and widely applicable kind of machine learning problem.
We'll use the IMDB dataset that contains the text of 50,000 movie reviews from the Internet Movie Database. These are split into 25,000 reviews for training and 25,000 reviews for testing. The training and testing sets are balanced, meaning they contain an equal number of positive and negative reviews.
This notebook uses tf.keras, a high-level API to build and train models in TensorFlow. For a more advanced text classification tutorial using tf.keras, see the MLCC Text Classification Guide.
In [0]:
from __future__ import absolute_import, division, print_function, unicode_literals
In [0]:
try:
# %tensorflow_version only exists in Colab.
%tensorflow_version 2.x
except Exception:
pass
import tensorflow as tf
In [0]:
from tensorflow import keras
import tensorflow_datasets as tfds
tfds.disable_progress_bar()
import numpy as np
print(tf.__version__)
The IMDB movie reviews dataset comes packaged in tfds. It has already been preprocessed so that the reviews (sequences of words) have been converted to sequences of integers, where each integer represents a specific word in a dictionary.
The following code downloads the IMDB dataset to your machine (or uses a cached copy if you've already downloaded it):
To encode your own text see the Loading text tutorial
In [0]:
(train_data, test_data), info = tfds.load(
# Use the version pre-encoded with an ~8k vocabulary.
'imdb_reviews/subwords8k',
# Return the train/test datasets as a tuple.
split = (tfds.Split.TRAIN, tfds.Split.TEST),
# Return (example, label) pairs from the dataset (instead of a dictionary).
as_supervised=True,
# Also return the `info` structure.
with_info=True)
In [0]:
encoder = info.features['text'].encoder
In [0]:
print ('Vocabulary size: {}'.format(encoder.vocab_size))
This text encoder will reversibly encode any string:
In [0]:
sample_string = 'Hello TensorFlow.'
encoded_string = encoder.encode(sample_string)
print ('Encoded string is {}'.format(encoded_string))
original_string = encoder.decode(encoded_string)
print ('The original string: "{}"'.format(original_string))
assert original_string == sample_string
The encoder encodes the string by breaking it into subwords or characters if the word is not in its dictionary. So the more a string resembles the dataset, the shorter the encoded representation will be.
In [0]:
for ts in encoded_string:
print ('{} ----> {}'.format(ts, encoder.decode([ts])))
Let's take a moment to understand the format of the data. The dataset comes preprocessed: each example is an array of integers representing the words of the movie review.
The text of reviews have been converted to integers, where each integer represents a specific word-piece in the dictionary.
Each label is an integer value of either 0 or 1, where 0 is a negative review, and 1 is a positive review.
Here's what the first review looks like:
In [0]:
for train_example, train_label in train_data.take(1):
print('Encoded text:', train_example[:10].numpy())
print('Label:', train_label.numpy())
The info structure contains the encoder/decoder. The encoder can be used to recover the original text:
In [0]:
encoder.decode(train_example)
In [0]:
BUFFER_SIZE = 1000
train_batches = (
train_data
.shuffle(BUFFER_SIZE)
.padded_batch(32, train_data.output_shapes))
test_batches = (
test_data
.padded_batch(32, train_data.output_shapes))
Each batch will have a shape of (batch_size, sequence_length) because the padding is dynamic each batch will have a different length:
In [0]:
for example_batch, label_batch in train_batches.take(2):
print("Batch shape:", example_batch.shape)
print("label shape:", label_batch.shape)
The neural network is created by stacking layers—this requires two main architectural decisions:
In this example, the input data consists of an array of word-indices. The labels to predict are either 0 or 1. Let's build a "Continuous bag of words" style model for this problem:
Caution: This model doesn't use masking, so the zero-padding is used as part of the input, so the padding length may affect the output. To fix this, see the masking and padding guide.
In [0]:
model = keras.Sequential([
keras.layers.Embedding(encoder.vocab_size, 16),
keras.layers.GlobalAveragePooling1D(),
keras.layers.Dense(1, activation='sigmoid')])
model.summary()
The layers are stacked sequentially to build the classifier:
Embedding layer. This layer takes the integer-encoded vocabulary and looks up the embedding vector for each word-index. These vectors are learned as the model trains. The vectors add a dimension to the output array. The resulting dimensions are: (batch, sequence, embedding).GlobalAveragePooling1D layer returns a fixed-length output vector for each example by averaging over the sequence dimension. This allows the model to handle input of variable length, in the simplest way possible.Dense) layer with 16 hidden units.sigmoid activation function, this value is a float between 0 and 1, representing a probability, or confidence level.The above model has two intermediate or "hidden" layers, between the input and output. The number of outputs (units, nodes, or neurons) is the dimension of the representational space for the layer. In other words, the amount of freedom the network is allowed when learning an internal representation.
If a model has more hidden units (a higher-dimensional representation space), and/or more layers, then the network can learn more complex representations. However, it makes the network more computationally expensive and may lead to learning unwanted patterns—patterns that improve performance on training data but not on the test data. This is called overfitting, and we'll explore it later.
A model needs a loss function and an optimizer for training. Since this is a binary classification problem and the model outputs a probability (a single-unit layer with a sigmoid activation), we'll use the binary_crossentropy loss function.
This isn't the only choice for a loss function, you could, for instance, choose mean_squared_error. But, generally, binary_crossentropy is better for dealing with probabilities—it measures the "distance" between probability distributions, or in our case, between the ground-truth distribution and the predictions.
Later, when we are exploring regression problems (say, to predict the price of a house), we will see how to use another loss function called mean squared error.
Now, configure the model to use an optimizer and a loss function:
In [0]:
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
In [0]:
history = model.fit(train_batches,
epochs=10,
validation_data=test_batches,
validation_steps=30)
In [0]:
loss, accuracy = model.evaluate(test_batches)
print("Loss: ", loss)
print("Accuracy: ", accuracy)
This fairly naive approach achieves an accuracy of about 87%. With more advanced approaches, the model should get closer to 95%.
In [0]:
history_dict = history.history
history_dict.keys()
There are four entries: one for each monitored metric during training and validation. We can use these to plot the training and validation loss for comparison, as well as the training and validation accuracy:
In [0]:
import matplotlib.pyplot as plt
acc = history_dict['accuracy']
val_acc = history_dict['val_accuracy']
loss = history_dict['loss']
val_loss = history_dict['val_loss']
epochs = range(1, len(acc) + 1)
# "bo" is for "blue dot"
plt.plot(epochs, loss, 'bo', label='Training loss')
# b is for "solid blue line"
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()
plt.show()
In [0]:
plt.clf() # clear figure
plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(loc='lower right')
plt.show()
In this plot, the dots represent the training loss and accuracy, and the solid lines are the validation loss and accuracy.
Notice the training loss decreases with each epoch and the training accuracy increases with each epoch. This is expected when using a gradient descent optimization—it should minimize the desired quantity on every iteration.
This isn't the case for the validation loss and accuracy—they seem to peak after about twenty epochs. This is an example of overfitting: the model performs better on the training data than it does on data it has never seen before. After this point, the model over-optimizes and learns representations specific to the training data that do not generalize to test data.
For this particular case, we could prevent overfitting by simply stopping the training after twenty or so epochs. Later, you'll see how to do this automatically with a callback.
In [0]: