In [ ]:
#@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 [ ]:
#@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 tutorial demonstrates text classification starting from plain text files stored on disk. You'll train a binary classifier to perform sentiment analysis on an IMDB dataset. At the end of the notebook, there is an exercise for you to try, in which you'll train a multiclass classifier to predict the tag for a programming question on Stack Overflow.
In [ ]:
!pip install tf-nightly
In [ ]:
import matplotlib.pyplot as plt
import os
import re
import shutil
import string
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras import losses
from tensorflow.keras import preprocessing
from tensorflow.keras.layers.experimental.preprocessing import TextVectorization
In [ ]:
print(tf.__version__)
This notebook trains a sentiment analysis model to classify movie reviews as positive or negative, based on 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.
You'll use the Large Movie Review 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.
In [ ]:
url = "https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz"
dataset = tf.keras.utils.get_file("aclImdb_v1.tar.gz", url,
untar=True, cache_dir='.',
cache_subdir='')
dataset_dir = os.path.join(os.path.dirname(dataset), 'aclImdb')
In [ ]:
os.listdir(dataset_dir)
In [ ]:
train_dir = os.path.join(dataset_dir, 'train')
os.listdir(train_dir)
The aclImdb/train/pos and aclImdb/train/neg directories contain many text files, each of which is a single movie review. Let's take a look at one of them.
In [ ]:
sample_file = os.path.join(train_dir, 'pos/1181_9.txt')
with open(sample_file) as f:
print(f.read())
Next, you will load the data off disk and prepare it into a format suitable for training. To do so, you will use the helpful text_dataset_from_directory utility, which expects a directory structure as follows.
main_directory/
...class_a/
......a_text_1.txt
......a_text_2.txt
...class_b/
......b_text_1.txt
......b_text_2.txtTo prepare a dataset for binary classification, you will need two folders on disk, corresponding to class_a and class_b. These will be the positive and negative movie reviews, which can be found in aclImdb/train/pos and aclImdb/train/neg. As the IMDB dataset contains additional folders, you will remove them before using this utility.
In [ ]:
remove_dir = os.path.join(train_dir, 'unsup')
shutil.rmtree(remove_dir)
Next, you will use the text_dataset_from_directory utility to create a labeled tf.data.Dataset. tf.data is a powerful collection of tools for working with data.
When running a machine learning experiment, it is a best practice to divide your dataset into three splits: train, validation, and test.
The IMDB dataset has already been divided into train and test, but it lacks a validation set. Let's create a validation set using an 80:20 split of the training data by using the validation_split argument below.
In [ ]:
batch_size = 32
seed = 42
raw_train_ds = tf.keras.preprocessing.text_dataset_from_directory(
'aclImdb/train',
batch_size=batch_size,
validation_split=0.2,
subset='training',
seed=seed)
As you can see above, there are 25,000 examples in the training folder, of which you will use 80% (or 20,000) for training. As you will see in a moment, you can train a model by passing a dataset directly to model.fit. If you're new to tf.data, you can also iterate over the dataset and print out a few examples as follows.
In [ ]:
for text_batch, label_batch in raw_train_ds.take(1):
for i in range(3):
print("Review", text_batch.numpy()[i])
print("Label", label_batch.numpy()[i])
Notice the reviews contain raw text (with punctuation and occasional HTML tags like <br/>). You will show how to handle these in the following section.
The labels are 0 or 1. To see which of these correspond to positive and negative movie reviews, you can check the class_names property on the dataset.
In [ ]:
print("Label 0 corresponds to", raw_train_ds.class_names[0])
print("Label 1 corresponds to", raw_train_ds.class_names[1])
Next, you will create a validation and test dataset. You will use the remaining 5,000 reviews from the training set for validation.
Note: When using the validation_split and subset arguments, make sure to either specify a random seed, or to pass shuffle=False, so that the validation and training splits have no overlap.
In [ ]:
raw_val_ds = tf.keras.preprocessing.text_dataset_from_directory(
'aclImdb/train',
batch_size=batch_size,
validation_split=0.2,
subset='validation',
seed=seed)
In [ ]:
raw_test_ds = tf.keras.preprocessing.text_dataset_from_directory(
'aclImdb/test',
batch_size=batch_size)
Note: The Preprocessing APIs used in the following section are experimental in TensorFlow 2.3 and subject to change.
Next, you will standardize, tokenize, and vectorize the data using the helpful preprocessing.TextVectorization layer.
Standardization refers to preprocessing the text, typically to remove punctuation or HTML elements to simplify the dataset. Tokenization refers to splitting strings into tokens (for example, splitting a sentence into individual words, by splitting on whitespace). Vectorization refers to converting tokens into numbers so they can be fed into a neural network. All of these tasks can be accomplished with this layer.
As you saw above, the reviews contain various HTML tags like <br />. These tags will not be removed by the default standardizer in the TextVectorization layer (which converts text to lowecase and strips punctuation by default, but doesn't strip HTML). You will write a custom standardization function to remove the HTML.
Note: to prevent train/test skew (also know as train/serving skew), it is important to preprocess the data identically at train and test time. To facilitate this, the TextVectorization layer can be included directly inside your model, as shown later in this tutorial.
In [ ]:
def custom_standardization(input_data):
lowercase = tf.strings.lower(input_data)
stripped_html = tf.strings.regex_replace(lowercase, '<br />', ' ')
return tf.strings.regex_replace(stripped_html,
'[%s]' % re.escape(string.punctuation),
'')
Next, you will create a TextVectorization layer. you will use this layer to standardize, tokenize, and vectorize our data. You set the output_mode to int to create unique integer indices for each token.
Note that you're using the default split function, and the custom standardization function you defined above. You'll also define some constants for the model, like an explicit maximum sequence_length, which will cause the layer to pad or truncate sequences to exactly sequence_length values.
In [ ]:
max_features = 10000
sequence_length = 250
vectorize_layer = TextVectorization(
standardize=custom_standardization,
max_tokens=max_features,
output_mode='int',
output_sequence_length=sequence_length)
Next, you will call adapt to fit the state of the preprocessing layer to the dataset. This will cause the model to build an index of strings to integers.
Note: it's important to only use your training data when calling adapt (using the test set would leak information).
In [ ]:
# Make a text-only dataset (without labels), then call adapt
train_text = raw_train_ds.map(lambda x, y: x)
vectorize_layer.adapt(train_text)
Let's create a function to see the result of using this layer to preprocess some data.
In [ ]:
def vectorize_text(text, label):
text = tf.expand_dims(text, -1)
return vectorize_layer(text), label
In [ ]:
# retrieve a batch (of 32 reviews and labels) from the dataset
text_batch, label_batch = next(iter(raw_train_ds))
first_review, first_label = text_batch[0], label_batch[0]
print("Review", first_review)
print("Label", raw_train_ds.class_names[first_label])
print("Vectorized review", vectorize_text(first_review, first_label))
As you can see above, each token has been replaced by an integer. You can lookup the token (string) that each integer corresponds to by calling .get_vocabulary() on the layer.
In [ ]:
print("1287 ---> ",vectorize_layer.get_vocabulary()[1287])
print(" 313 ---> ",vectorize_layer.get_vocabulary()[313])
print('Vocabulary size: {}'.format(len(vectorize_layer.get_vocabulary())))
You are nearly ready to train your model. As a final preprocessing step, you will apply the TextVectorization layer you created earlier to the train, validation, and test dataset.
In [ ]:
train_ds = raw_train_ds.map(vectorize_text)
val_ds = raw_val_ds.map(vectorize_text)
test_ds = raw_test_ds.map(vectorize_text)
These are two important methods you should use when loading data to make sure that I/O does not become blocking.
.cache() keeps data in memory after it's loaded off disk. This will ensure the dataset does not become a bottleneck while training your model. If your dataset is too large to fit into memory, you can also use this method to create a performant on-disk cache, which is more efficient to read than many small files.
.prefetch() overlaps data preprocessing and model execution while training.
You can learn more about both methods, as well as how to cache data to disk in the data performance guide.
In [ ]:
AUTOTUNE = tf.data.experimental.AUTOTUNE
train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)
test_ds = test_ds.cache().prefetch(buffer_size=AUTOTUNE)
In [ ]:
embedding_dim = 16
In [ ]:
model = tf.keras.Sequential([
layers.Embedding(max_features + 1, embedding_dim),
layers.Dropout(0.2),
layers.GlobalAveragePooling1D(),
layers.Dropout(0.2),
layers.Dense(1)])
model.summary()
The layers are stacked sequentially to build the classifier:
Embedding layer. This layer takes the integer-encoded reviews and looks up an 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). To learn more about embeddings, see the word embedding tutorial.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. 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), you'll use losses.BinaryCrossentropy loss function.
Now, configure the model to use an optimizer and a loss function:
In [ ]:
model.compile(loss=losses.BinaryCrossentropy(from_logits=True), optimizer='adam', metrics=tf.metrics.BinaryAccuracy(threshold=0.0))
In [ ]:
epochs = 10
history = model.fit(
train_ds,
validation_data=val_ds,
epochs=epochs)
In [ ]:
loss, accuracy = model.evaluate(test_ds)
print("Loss: ", loss)
print("Accuracy: ", accuracy)
This fairly naive approach achieves an accuracy of about 86%.
In [ ]:
history_dict = history.history
history_dict.keys()
There are four entries: one for each monitored metric during training and validation. You can use these to plot the training and validation loss for comparison, as well as the training and validation accuracy:
In [ ]:
acc = history_dict['binary_accuracy']
val_acc = history_dict['val_binary_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 [ ]:
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 before the training accuracy. 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, you could prevent overfitting by simply stopping the training when the validation accuracy is no longer increasing. One way to do so is to use the EarlyStopping callback.
In the code above, you applied the TextVectorization layer to the dataset before feeding text to the model. If you want to make your model capable of processing raw strings (for example, to simplify deploying it), you can include the TextVectorization layer inside your model. To do so, you can create a new model using the weights you just trained.
In [ ]:
export_model = tf.keras.Sequential([
vectorize_layer,
model,
layers.Activation('sigmoid')
])
export_model.compile(
loss=losses.BinaryCrossentropy(from_logits=False), optimizer="adam", metrics=['accuracy']
)
# Test it with `raw_test_ds`, which yields raw strings
loss, accuracy = export_model.evaluate(raw_test_ds)
print(accuracy)
Including the text preprocessing logic inside your model enables you to export a model for production that simplifies deployment, and reduces the potential for train/test skew.
There is a performance difference to keep in mind when choosing where to apply your TextVectorization layer. Using it outside of your model enables you to do asynchronous CPU processing and buffering of your data when training on GPU. So, if you're training your model on the GPU, you probably want to go with this option to get the best performance while developing your model, then switch to including the TextVectorization layer inside your model when you're ready to prepare for deployment.
Visit this tutorial to learn more about saving models.
This tutorial showed how to train a binary classifier from scratch on the IMDB dataset. As an exercise, you can modify this notebook to train a multiclass classifier to predict the tag of a programming question on Stack Overflow.
We have prepared a dataset for you to use containing the body of several thousand programming questions (for example, "How can sort a dictionary by value in Python?") posted to Stack Overflow. Each of these is labeled with exactly one tag (either Python, CSharp, JavaScript, or Java). Your task is to take a question as input, and predict the appropriate tag, in this case, Python.
The dataset you will work with contains 8,000 questions extracted from the much larger public Stack Overflow dataset on BigQuery, which contains more than 17 million posts.
After downloading the dataset, you will find it has a similar directory structure to the IMDB dataset you worked with previously:
train/
...python/
......0.txt
......1.txt
...javascript/
......0.txt
......1.txt
...csharp/
......0.txt
......1.txt
...java/
......0.txt
......1.txtNote: to increase the difficulty of the classification problem, we have replaced any occurences of the words Python, CSharp, JavaScript, or Java in the programming questions with the word blank (as many questions contain the language they're about).
To complete this exercise, you should modify this notebook to work with the Stack Overflow dataset by making the following modifications:
At the top of your notebook, update the code that downloads the IMDB dataset with code to download the Stack Overflow dataset we have prepreared. As the Stack Overflow dataset has a similar directory structure, you will not need to make many modifications.
Modify the last layer of your model to read Dense(4), as there are now four output classes.
When you compile your model, change the loss to SparseCategoricalCrossentropy. This is the correct loss function to use for a multiclass classification problem, when the labels for each class are integers (in our case, they can be 0, 1, 2, or 3).
Once these changes are complete, you will be able to train a multiclass classifier.
If you get stuck, you can find a solution here.
This tutorial introduced text classification from scratch. To learn more about the text classification workflow in general, we recommend reading this guide from Google Developers.