In [0]:
# Copyright 2019 The TensorFlow Hub Authors. All Rights Reserved.
#
# 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
#
# http://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, and TensorFlow Hub, a library and platform for transfer learning. For a more advanced text classification tutorial using tf.keras, see the MLCC Text Classification Guide.
In [0]:
import numpy as np
import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_datasets as tfds
import matplotlib.pyplot as plt
print("Version: ", tf.__version__)
print("Eager mode: ", tf.executing_eagerly())
print("Hub version: ", hub.__version__)
print("GPU is", "available" if tf.config.list_physical_devices('GPU') else "NOT AVAILABLE")
The IMDB dataset is available on TensorFlow datasets. The following code downloads the IMDB dataset to your machine (or the colab runtime):
In [0]:
train_data, test_data = tfds.load(name="imdb_reviews", split=["train", "test"],
batch_size=-1, as_supervised=True)
train_examples, train_labels = tfds.as_numpy(train_data)
test_examples, test_labels = tfds.as_numpy(test_data)
Let's take a moment to understand the format of the data. Each example is a sentence representing the movie review and a corresponding label. The sentence is not preprocessed in any way. The label is an integer value of either 0 or 1, where 0 is a negative review, and 1 is a positive review.
In [0]:
print("Training entries: {}, test entries: {}".format(len(train_examples), len(test_examples)))
Let's print first 10 examples.
In [0]:
train_examples[:10]
Let's also print the first 10 labels.
In [0]:
train_labels[:10]
The neural network is created by stacking layers—this requires three main architectural decisions:
In this example, the input data consists of sentences. The labels to predict are either 0 or 1.
One way to represent the text is to convert sentences into embeddings vectors. We can use a pre-trained text embedding as the first layer, which will have two advantages:
For this example we will use a model from TensorFlow Hub called google/tf2-preview/gnews-swivel-20dim/1.
There are three other models to test for the sake of this tutorial:
Let's first create a Keras layer that uses a TensorFlow Hub model to embed the sentences, and try it out on a couple of input examples. Note that the output shape of the produced embeddings is a expected: (num_examples, embedding_dimension).
In [0]:
model = "https://tfhub.dev/google/tf2-preview/gnews-swivel-20dim/1"
hub_layer = hub.KerasLayer(model, output_shape=[20], input_shape=[],
dtype=tf.string, trainable=True)
hub_layer(train_examples[:3])
Let's now build the full model:
In [0]:
model = tf.keras.Sequential()
model.add(hub_layer)
model.add(tf.keras.layers.Dense(16, activation='relu'))
model.add(tf.keras.layers.Dense(1))
model.summary()
The layers are stacked sequentially to build the classifier:
(num_examples, embedding_dimension).Dense) layer with 16 hidden units.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=tf.losses.BinaryCrossentropy(from_logits=True),
metrics=['accuracy'])
When training, we want to check the accuracy of the model on data it hasn't seen before. Create a validation set by setting apart 10,000 examples from the original training data. (Why not use the testing set now? Our goal is to develop and tune our model using only the training data, then use the test data just once to evaluate our accuracy).
In [0]:
x_val = train_examples[:10000]
partial_x_train = train_examples[10000:]
y_val = train_labels[:10000]
partial_y_train = train_labels[10000:]
In [0]:
history = model.fit(partial_x_train,
partial_y_train,
epochs=40,
batch_size=512,
validation_data=(x_val, y_val),
verbose=1)
In [0]:
results = model.evaluate(test_data, test_labels)
print(results)
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]:
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()
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.