In [1]:
import keras
keras.__version__
Out[1]:
This notebook contains the code samples found in Chapter 3, Section 5 of Deep Learning with Python. Note that the original text features far more content, in particular further explanations and figures: in this notebook, you will only find source code and related comments.
In the previous section we saw how to classify vector inputs into two mutually exclusive classes using a densely-connected neural network. But what happens when you have more than two classes?
In this section, we will build a network to classify Reuters newswires into 46 different mutually-exclusive topics. Since we have many classes, this problem is an instance of "multi-class classification", and since each data point should be classified into only one category, the problem is more specifically an instance of "single-label, multi-class classification". If each data point could have belonged to multiple categories (in our case, topics) then we would be facing a "multi-label, multi-class classification" problem.
We will be working with the Reuters dataset, a set of short newswires and their topics, published by Reuters in 1986. It's a very simple, widely used toy dataset for text classification. There are 46 different topics; some topics are more represented than others, but each topic has at least 10 examples in the training set.
Like IMDB and MNIST, the Reuters dataset comes packaged as part of Keras. Let's take a look right away:
In [2]:
from keras.datasets import reuters
(train_data, train_labels), (test_data, test_labels) = reuters.load_data(num_words=10000)
Like with the IMDB dataset, the argument num_words=10000 restricts the data to the 10,000 most frequently occurring words found in the
data.
We have 8,982 training examples and 2,246 test examples:
In [3]:
len(train_data)
Out[3]:
In [4]:
len(test_data)
Out[4]:
As with the IMDB reviews, each example is a list of integers (word indices):
In [5]:
train_data[10]
Out[5]:
Here's how you can decode it back to words, in case you are curious:
In [6]:
word_index = reuters.get_word_index()
reverse_word_index = dict([(value, key) for (key, value) in word_index.items()])
# Note that our indices were offset by 3
# because 0, 1 and 2 are reserved indices for "padding", "start of sequence", and "unknown".
decoded_newswire = ' '.join([reverse_word_index.get(i - 3, '?') for i in train_data[0]])
In [7]:
decoded_newswire
Out[7]:
The label associated with an example is an integer between 0 and 45: a topic index.
In [8]:
train_labels[10]
Out[8]:
In [9]:
import numpy as np
def vectorize_sequences(sequences, dimension=10000):
results = np.zeros((len(sequences), dimension))
for i, sequence in enumerate(sequences):
results[i, sequence] = 1.
return results
# Our vectorized training data
x_train = vectorize_sequences(train_data)
# Our vectorized test data
x_test = vectorize_sequences(test_data)
To vectorize the labels, there are two possibilities: we could just cast the label list as an integer tensor, or we could use a "one-hot" encoding. One-hot encoding is a widely used format for categorical data, also called "categorical encoding". For a more detailed explanation of one-hot encoding, you can refer to Chapter 6, Section 1. In our case, one-hot encoding of our labels consists in embedding each label as an all-zero vector with a 1 in the place of the label index, e.g.:
In [10]:
def to_one_hot(labels, dimension=46):
results = np.zeros((len(labels), dimension))
for i, label in enumerate(labels):
results[i, label] = 1.
return results
# Our vectorized training labels
one_hot_train_labels = to_one_hot(train_labels)
# Our vectorized test labels
one_hot_test_labels = to_one_hot(test_labels)
Note that there is a built-in way to do this in Keras, which you have already seen in action in our MNIST example:
In [11]:
from keras.utils.np_utils import to_categorical
one_hot_train_labels = to_categorical(train_labels)
one_hot_test_labels = to_categorical(test_labels)
This topic classification problem looks very similar to our previous movie review classification problem: in both cases, we are trying to classify short snippets of text. There is however a new constraint here: the number of output classes has gone from 2 to 46, i.e. the dimensionality of the output space is much larger.
In a stack of Dense layers like what we were using, each layer can only access information present in the output of the previous layer.
If one layer drops some information relevant to the classification problem, this information can never be recovered by later layers: each
layer can potentially become an "information bottleneck". In our previous example, we were using 16-dimensional intermediate layers, but a
16-dimensional space may be too limited to learn to separate 46 different classes: such small layers may act as information bottlenecks,
permanently dropping relevant information.
For this reason we will use larger layers. Let's go with 64 units:
In [12]:
from keras import models
from keras import layers
model = models.Sequential()
model.add(layers.Dense(64, activation='relu', input_shape=(10000,)))
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(46, activation='softmax'))
There are two other things you should note about this architecture:
Dense layer of size 46. This means that for each input sample, our network will output a
46-dimensional vector. Each entry in this vector (each dimension) will encode a different output class.softmax activation. You have already seen this pattern in the MNIST example. It means that the network will
output a probability distribution over the 46 different output classes, i.e. for every input sample, the network will produce a
46-dimensional output vector where output[i] is the probability that the sample belongs to class i. The 46 scores will sum to 1.The best loss function to use in this case is categorical_crossentropy. It measures the distance between two probability distributions:
in our case, between the probability distribution output by our network, and the true distribution of the labels. By minimizing the
distance between these two distributions, we train our network to output something as close as possible to the true labels.
In [13]:
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
In [14]:
x_val = x_train[:1000]
partial_x_train = x_train[1000:]
y_val = one_hot_train_labels[:1000]
partial_y_train = one_hot_train_labels[1000:]
Now let's train our network for 20 epochs:
In [15]:
history = model.fit(partial_x_train,
partial_y_train,
epochs=20,
batch_size=512,
validation_data=(x_val, y_val))
Let's display its loss and accuracy curves:
In [16]:
import matplotlib.pyplot as plt
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(1, len(loss) + 1)
plt.plot(epochs, loss, 'bo', label='Training loss')
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 [18]:
plt.clf() # clear figure
acc = history.history['acc']
val_acc = history.history['val_acc']
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('Loss')
plt.legend()
plt.show()
It seems that the network starts overfitting after 8 epochs. Let's train a new network from scratch for 8 epochs, then let's evaluate it on the test set:
In [27]:
model = models.Sequential()
model.add(layers.Dense(64, activation='relu', input_shape=(10000,)))
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(46, activation='softmax'))
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.fit(partial_x_train,
partial_y_train,
epochs=8,
batch_size=512,
validation_data=(x_val, y_val))
results = model.evaluate(x_test, one_hot_test_labels)
In [28]:
results
Out[28]:
Our approach reaches an accuracy of ~78%. With a balanced binary classification problem, the accuracy reached by a purely random classifier would be 50%, but in our case it is closer to 19%, so our results seem pretty good, at least when compared to a random baseline:
In [29]:
import copy
test_labels_copy = copy.copy(test_labels)
np.random.shuffle(test_labels_copy)
float(np.sum(np.array(test_labels) == np.array(test_labels_copy))) / len(test_labels)
Out[29]:
In [30]:
predictions = model.predict(x_test)
Each entry in predictions is a vector of length 46:
In [31]:
predictions[0].shape
Out[31]:
The coefficients in this vector sum to 1:
In [32]:
np.sum(predictions[0])
Out[32]:
The largest entry is the predicted class, i.e. the class with the highest probability:
In [33]:
np.argmax(predictions[0])
Out[33]:
In [35]:
y_train = np.array(train_labels)
y_test = np.array(test_labels)
The only thing it would change is the choice of the loss function. Our previous loss, categorical_crossentropy, expects the labels to
follow a categorical encoding. With integer labels, we should use sparse_categorical_crossentropy:
In [36]:
model.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy', metrics=['acc'])
This new loss function is still mathematically the same as categorical_crossentropy; it just has a different interface.
We mentioned earlier that since our final outputs were 46-dimensional, we should avoid intermediate layers with much less than 46 hidden units. Now let's try to see what happens when we introduce an information bottleneck by having intermediate layers significantly less than 46-dimensional, e.g. 4-dimensional.
In [42]:
model = models.Sequential()
model.add(layers.Dense(64, activation='relu', input_shape=(10000,)))
model.add(layers.Dense(4, activation='relu'))
model.add(layers.Dense(46, activation='softmax'))
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.fit(partial_x_train,
partial_y_train,
epochs=20,
batch_size=128,
validation_data=(x_val, y_val))
Out[42]:
Our network now seems to peak at ~71% test accuracy, a 8% absolute drop. This drop is mostly due to the fact that we are now trying to compress a lot of information (enough information to recover the separation hyperplanes of 46 classes) into an intermediate space that is too low-dimensional. The network is able to cram most of the necessary information into these 8-dimensional representations, but not all of it.
Here's what you should take away from this example:
Dense layer of size N.softmax activation, so that it will output a
probability distribution over the N output classes.categorical_crossentropy as your loss
function.
Encoding the labels as integers and using the sparse_categorical_crossentropy loss function.