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.
|
|
|
This notebook classifies movie reviews as positive or negative using the text of the review. This is an example of binary classification, an important and widely applicable kind of machine learning problem.
We will demonstrate the use of graph regularization in this notebook by building a graph from the given input. The general recipe for building a graph-regularized model using the Neural Structured Learning (NSL) framework when the input does not contain an explicit graph is as follows:
Note: We expect that it would take readers about 1 hour to go through this tutorial.
In [0]:
!pip install --quiet neural-structured-learning
!pip install --quiet tensorflow-hub
In [0]:
import matplotlib.pyplot as plt
import numpy as np
import neural_structured_learning as nsl
import tensorflow as tf
import tensorflow_hub as hub
# Resets notebook state
tf.keras.backend.clear_session()
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 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 this tutorial, we will use a preprocessed version of the IMDB dataset.
The IMDB dataset comes packaged with TensorFlow. It has already been preprocessed such 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 (or uses a cached copy if it has already been downloaded):
In [0]:
imdb = tf.keras.datasets.imdb
(pp_train_data, pp_train_labels), (pp_test_data, pp_test_labels) = (
imdb.load_data(num_words=10000))
The argument num_words=10000 keeps the top 10,000 most frequently occurring words in the training data. The rare words are discarded to keep the size of the vocabulary manageable.
In [0]:
print('Training entries: {}, labels: {}'.format(
len(pp_train_data), len(pp_train_labels)))
training_samples_count = len(pp_train_data)
The text of reviews have been converted to integers, where each integer represents a specific word in a dictionary. Here's what the first review looks like:
In [0]:
print(pp_train_data[0])
Movie reviews may be different lengths. The below code shows the number of words in the first and second reviews. Since inputs to a neural network must be the same length, we'll need to resolve this later.
In [0]:
len(pp_train_data[0]), len(pp_train_data[1])
In [0]:
def build_reverse_word_index():
# A dictionary mapping words to an integer index
word_index = imdb.get_word_index()
# The first indices are reserved
word_index = {k: (v + 3) for k, v in word_index.items()}
word_index['<PAD>'] = 0
word_index['<START>'] = 1
word_index['<UNK>'] = 2 # unknown
word_index['<UNUSED>'] = 3
return dict((value, key) for (key, value) in word_index.items())
reverse_word_index = build_reverse_word_index()
def decode_review(text):
return ' '.join([reverse_word_index.get(i, '?') for i in text])
Now we can use the decode_review function to display the text for the first review:
In [0]:
decode_review(pp_train_data[0])
In [0]:
!mkdir -p /tmp/imdb
We will use pretrained Swivel embeddings to create embeddings in the
tf.train.Example format for each sample in the input. We will store the
resulting embeddings in the TFRecord format along with an additional feature
that represents the ID of each sample. This is important and will allow us match
sample embeddings with corresponding nodes in the graph later.
In [0]:
pretrained_embedding = 'https://tfhub.dev/google/tf2-preview/gnews-swivel-20dim/1'
hub_layer = hub.KerasLayer(
pretrained_embedding, input_shape=[], dtype=tf.string, trainable=True)
In [0]:
def _int64_feature(value):
"""Returns int64 tf.train.Feature."""
return tf.train.Feature(int64_list=tf.train.Int64List(value=value.tolist()))
def _bytes_feature(value):
"""Returns bytes tf.train.Feature."""
return tf.train.Feature(
bytes_list=tf.train.BytesList(value=[value.encode('utf-8')]))
def _float_feature(value):
"""Returns float tf.train.Feature."""
return tf.train.Feature(float_list=tf.train.FloatList(value=value.tolist()))
def create_embedding_example(word_vector, record_id):
"""Create tf.Example containing the sample's embedding and its ID."""
text = decode_review(word_vector)
# Shape = [batch_size,].
sentence_embedding = hub_layer(tf.reshape(text, shape=[-1,]))
# Flatten the sentence embedding back to 1-D.
sentence_embedding = tf.reshape(sentence_embedding, shape=[-1])
features = {
'id': _bytes_feature(str(record_id)),
'embedding': _float_feature(sentence_embedding.numpy())
}
return tf.train.Example(features=tf.train.Features(feature=features))
def create_embeddings(word_vectors, output_path, starting_record_id):
record_id = int(starting_record_id)
with tf.io.TFRecordWriter(output_path) as writer:
for word_vector in word_vectors:
example = create_embedding_example(word_vector, record_id)
record_id = record_id + 1
writer.write(example.SerializeToString())
return record_id
# Persist TF.Example features containing embeddings for training data in
# TFRecord format.
create_embeddings(pp_train_data, '/tmp/imdb/embeddings.tfr', 0)
Now that we have the sample embeddings, we will use them to build a similarity graph, i.e, nodes in this graph will correspond to samples and edges in this graph will correspond to similarity between pairs of nodes.
Neural Structured Learning provides a graph building library to build a graph based on sample embeddings. It uses cosine similarity as the similarity measure to compare embeddings and build edges between them. It also allows us to specify a similarity threshold, which can be used to discard dissimilar edges from the final graph. In this example, using 0.99 as the similarity threshold, we end up with a graph that has 445,327 bi-directional edges.
In [0]:
nsl.tools.build_graph(['/tmp/imdb/embeddings.tfr'],
'/tmp/imdb/graph_99.tsv',
similarity_threshold=0.99)
Note: Graph quality and by extension, embedding quality, are very important for graph regularization. While we have used Swivel embeddings in this notebook, using BERT embeddings for instance, will likely capture review semantics more accurately. We encourage users to use embeddings of their choice and as appropriate to their needs.
We create sample features for our problem using the tf.train.Example format
and persist them in the TFRecord format. Each sample will include the
following three features:
In [0]:
def create_example(word_vector, label, record_id):
"""Create tf.Example containing the sample's word vector, label, and ID."""
features = {
'id': _bytes_feature(str(record_id)),
'words': _int64_feature(np.asarray(word_vector)),
'label': _int64_feature(np.asarray([label])),
}
return tf.train.Example(features=tf.train.Features(feature=features))
def create_records(word_vectors, labels, record_path, starting_record_id):
record_id = int(starting_record_id)
with tf.io.TFRecordWriter(record_path) as writer:
for word_vector, label in zip(word_vectors, labels):
example = create_example(word_vector, label, record_id)
record_id = record_id + 1
writer.write(example.SerializeToString())
return record_id
# Persist TF.Example features (word vectors and labels) for training and test
# data in TFRecord format.
next_record_id = create_records(pp_train_data, pp_train_labels,
'/tmp/imdb/train_data.tfr', 0)
create_records(pp_test_data, pp_test_labels, '/tmp/imdb/test_data.tfr',
next_record_id)
Since we have the sample features and the synthesized graph, we can generate the augmented training data for Neural Structured Learning. The NSL framework provides a library to combine the graph and the sample features to produce the final training data for graph regularization. The resulting training data will include original sample features as well as features of their corresponding neighbors.
In this tutorial, we consider undirected edges and use a maximum of 3 neighbors per sample to augment training data with graph neighbors.
In [0]:
nsl.tools.pack_nbrs(
'/tmp/imdb/train_data.tfr',
'',
'/tmp/imdb/graph_99.tsv',
'/tmp/imdb/nsl_train_data.tfr',
add_undirected_edges=True,
max_nbrs=3)
We are now ready to build a base model without graph regularization. In order to build this model, we can either use embeddings that were used in building the graph, or we can learn new embeddings jointly along with the classification task. For the purpose of this notebook, we will do the latter.
In [0]:
NBR_FEATURE_PREFIX = 'NL_nbr_'
NBR_WEIGHT_SUFFIX = '_weight'
We will use an instance of HParams to inclue various hyperparameters and
constants used for training and evaluation. We briefly describe each of them
below:
num_classes: There are 2 classes -- positive and negative.
max_seq_length: This is the maximum number of words considered from each movie review in this example.
vocab_size: This is the size of the vocabulary considered for this example.
distance_type: This is the distance metric used to regularize the sample with its neighbors.
graph_regularization_multiplier: This controls the relative weight of the graph regularization term in the overall loss function.
num_neighbors: The number of neighbors used for graph regularization.
This value has to be less than or equal to the max_nbrs argument used
above when invoking nsl.tools.pack_nbrs.
num_fc_units: The number of units in the fully connected layer of the neural network.
train_epochs: The number of training epochs.
batch_size: Batch size used for training and evaluation.
eval_steps: The number of batches to process before deeming evaluation
is complete. If set to None, all instances in the test set are evaluated.
In [0]:
class HParams(object):
"""Hyperparameters used for training."""
def __init__(self):
### dataset parameters
self.num_classes = 2
self.max_seq_length = 256
self.vocab_size = 10000
### neural graph learning parameters
self.distance_type = nsl.configs.DistanceType.L2
self.graph_regularization_multiplier = 0.1
self.num_neighbors = 2
### model architecture
self.num_embedding_dims = 16
self.num_lstm_dims = 64
self.num_fc_units = 64
### training parameters
self.train_epochs = 10
self.batch_size = 128
### eval parameters
self.eval_steps = None # All instances in the test set are evaluated.
HPARAMS = HParams()
The reviews—the arrays of integers—must be converted to tensors before being fed into the neural network. This conversion can be done a couple of ways:
Convert the arrays into vectors of 0s and 1s indicating word occurrence,
similar to a one-hot encoding. For example, the sequence [3, 5] would become a 10000-dimensional vector that is all zeros except for indices 3 and 5, which are ones. Then, make this the first layer in our network—a Dense layer—that can handle floating point vector data. This approach is memory intensive, though, requiring a num_words * num_reviews size matrix.
Alternatively, we can pad the arrays so they all have the same length, then
create an integer tensor of shape max_length * num_reviews. We can use an
embedding layer capable of handling this shape as the first layer in our
network.
In this tutorial, we will use the second approach.
Since the movie reviews must be the same length, we will use the pad_sequence
function defined below to standardize the lengths.
In [0]:
def make_dataset(file_path, training=False):
"""Creates a `tf.data.TFRecordDataset`.
Args:
file_path: Name of the file in the `.tfrecord` format containing
`tf.train.Example` objects.
training: Boolean indicating if we are in training mode.
Returns:
An instance of `tf.data.TFRecordDataset` containing the `tf.train.Example`
objects.
"""
def pad_sequence(sequence, max_seq_length):
"""Pads the input sequence (a `tf.SparseTensor`) to `max_seq_length`."""
pad_size = tf.maximum([0], max_seq_length - tf.shape(sequence)[0])
padded = tf.concat(
[sequence.values,
tf.fill((pad_size), tf.cast(0, sequence.dtype))],
axis=0)
# The input sequence may be larger than max_seq_length. Truncate down if
# necessary.
return tf.slice(padded, [0], [max_seq_length])
def parse_example(example_proto):
"""Extracts relevant fields from the `example_proto`.
Args:
example_proto: An instance of `tf.train.Example`.
Returns:
A pair whose first value is a dictionary containing relevant features
and whose second value contains the ground truth labels.
"""
# The 'words' feature is a variable length word ID vector.
feature_spec = {
'words': tf.io.VarLenFeature(tf.int64),
'label': tf.io.FixedLenFeature((), tf.int64, default_value=-1),
}
# We also extract corresponding neighbor features in a similar manner to
# the features above during training.
if training:
for i in range(HPARAMS.num_neighbors):
nbr_feature_key = '{}{}_{}'.format(NBR_FEATURE_PREFIX, i, 'words')
nbr_weight_key = '{}{}{}'.format(NBR_FEATURE_PREFIX, i,
NBR_WEIGHT_SUFFIX)
feature_spec[nbr_feature_key] = tf.io.VarLenFeature(tf.int64)
# We assign a default value of 0.0 for the neighbor weight so that
# graph regularization is done on samples based on their exact number
# of neighbors. In other words, non-existent neighbors are discounted.
feature_spec[nbr_weight_key] = tf.io.FixedLenFeature(
[1], tf.float32, default_value=tf.constant([0.0]))
features = tf.io.parse_single_example(example_proto, feature_spec)
# Since the 'words' feature is a variable length word vector, we pad it to a
# constant maximum length based on HPARAMS.max_seq_length
features['words'] = pad_sequence(features['words'], HPARAMS.max_seq_length)
if training:
for i in range(HPARAMS.num_neighbors):
nbr_feature_key = '{}{}_{}'.format(NBR_FEATURE_PREFIX, i, 'words')
features[nbr_feature_key] = pad_sequence(features[nbr_feature_key],
HPARAMS.max_seq_length)
labels = features.pop('label')
return features, labels
dataset = tf.data.TFRecordDataset([file_path])
if training:
dataset = dataset.shuffle(10000)
dataset = dataset.map(parse_example)
dataset = dataset.batch(HPARAMS.batch_size)
return dataset
train_dataset = make_dataset('/tmp/imdb/nsl_train_data.tfr', True)
test_dataset = make_dataset('/tmp/imdb/test_data.tfr')
A 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.
We will use a bi-directional LSTM as our base model in this tutorial.
In [0]:
# This function exists as an alternative to the bi-LSTM model used in this
# notebook.
def make_feed_forward_model():
"""Builds a simple 2 layer feed forward neural network."""
inputs = tf.keras.Input(
shape=(HPARAMS.max_seq_length,), dtype='int64', name='words')
embedding_layer = tf.keras.layers.Embedding(HPARAMS.vocab_size, 16)(inputs)
pooling_layer = tf.keras.layers.GlobalAveragePooling1D()(embedding_layer)
dense_layer = tf.keras.layers.Dense(16, activation='relu')(pooling_layer)
outputs = tf.keras.layers.Dense(1, activation='sigmoid')(dense_layer)
return tf.keras.Model(inputs=inputs, outputs=outputs)
def make_bilstm_model():
"""Builds a bi-directional LSTM model."""
inputs = tf.keras.Input(
shape=(HPARAMS.max_seq_length,), dtype='int64', name='words')
embedding_layer = tf.keras.layers.Embedding(HPARAMS.vocab_size,
HPARAMS.num_embedding_dims)(
inputs)
lstm_layer = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(HPARAMS.num_lstm_dims))(
embedding_layer)
dense_layer = tf.keras.layers.Dense(
HPARAMS.num_fc_units, activation='relu')(
lstm_layer)
outputs = tf.keras.layers.Dense(1, activation='sigmoid')(dense_layer)
return tf.keras.Model(inputs=inputs, outputs=outputs)
# Feel free to use an architecture of your choice.
model = make_bilstm_model()
model.summary()
The layers are effectively stacked sequentially to build the classifier:
Input layer which takes the integer-encoded
vocabulary.Embedding layer, which 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).Dense)
layer with 64 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, and excluding the Embedding layer. 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.
In [0]:
model.compile(
optimizer='adam', loss='binary_crossentropy', 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 a fraction of 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 this tutorial, we take roughly 10% of the initial training samples (10% of 25000) as labeled data for training and the remaining as validation data. Since the initial train/test split was 50/50 (25000 samples each), the effective train/validation/test split we now have is 5/45/50.
Note that 'train_dataset' has already been batched and shuffled.
In [0]:
validation_fraction = 0.9
validation_size = int(validation_fraction *
int(training_samples_count / HPARAMS.batch_size))
print(validation_size)
validation_dataset = train_dataset.take(validation_size)
train_dataset = train_dataset.skip(validation_size)
In [0]:
history = model.fit(
train_dataset,
validation_data=validation_dataset,
epochs=HPARAMS.train_epochs,
verbose=1)
In [0]:
results = model.evaluate(test_dataset, steps=HPARAMS.eval_steps)
print(results)
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)
# "-r^" is for solid red line with triangle markers.
plt.plot(epochs, loss, '-r^', label='Training loss')
# "-b0" is for solid blue line with circle markers.
plt.plot(epochs, val_loss, '-bo', label='Validation loss')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(loc='best')
plt.show()
In [0]:
plt.clf() # clear figure
plt.plot(epochs, acc, '-r^', label='Training acc')
plt.plot(epochs, val_acc, '-bo', label='Validation acc')
plt.title('Training and validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(loc='best')
plt.show()
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.
We are now ready to try graph regularization using the base model that we built
above. We will use the GraphRegularization wrapper class provided by the
Neural Structured Learning framework to wrap the base (bi-LSTM) model to include
graph regularization. The rest of the steps for training and evaluating the
graph-regularized model are similar to that of the base model.
To assess the incremental benefit of graph regularization, we will create a new
base model instance. This is because model has already been trained for a few
iterations, and reusing this trained model to create a graph-regularized model
will not be a fair comparison for model.
In [0]:
# Build a new base LSTM model.
base_reg_model = make_bilstm_model()
In [0]:
# Wrap the base model with graph regularization.
graph_reg_config = nsl.configs.make_graph_reg_config(
max_neighbors=HPARAMS.num_neighbors,
multiplier=HPARAMS.graph_regularization_multiplier,
distance_type=HPARAMS.distance_type,
sum_over_axis=-1)
graph_reg_model = nsl.keras.GraphRegularization(base_reg_model,
graph_reg_config)
graph_reg_model.compile(
optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
In [0]:
graph_reg_history = graph_reg_model.fit(
train_dataset,
validation_data=validation_dataset,
epochs=HPARAMS.train_epochs,
verbose=1)
In [0]:
graph_reg_results = graph_reg_model.evaluate(test_dataset, steps=HPARAMS.eval_steps)
print(graph_reg_results)
In [0]:
graph_reg_history_dict = graph_reg_history.history
graph_reg_history_dict.keys()
There are five entries in total in the dictionary: training loss, training accuracy, training graph loss, validation loss, and validation accuracy. We can plot them all together for comparison. Note that the graph loss is only computed during training.
In [0]:
acc = graph_reg_history_dict['accuracy']
val_acc = graph_reg_history_dict['val_accuracy']
loss = graph_reg_history_dict['loss']
graph_loss = graph_reg_history_dict['graph_loss']
val_loss = graph_reg_history_dict['val_loss']
epochs = range(1, len(acc) + 1)
plt.clf() # clear figure
# "-r^" is for solid red line with triangle markers.
plt.plot(epochs, loss, '-r^', label='Training loss')
# "-gD" is for solid green line with diamond markers.
plt.plot(epochs, graph_loss, '-gD', label='Training graph loss')
# "-b0" is for solid blue line with circle markers.
plt.plot(epochs, val_loss, '-bo', label='Validation loss')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(loc='best')
plt.show()
In [0]:
plt.clf() # clear figure
plt.plot(epochs, acc, '-r^', label='Training acc')
plt.plot(epochs, val_acc, '-bo', label='Validation acc')
plt.title('Training and validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(loc='best')
plt.show()
Semi-supervised learning and more specifically, graph regularization in the context of this tutorial, can be really powerful when the amount of training data is small. The lack of training data is compensated by leveraging similarity among the training samples, which is not possible in traditional supervised learning.
We define supervision ratio as the ratio of training samples to the total number of samples which includes training, validation, and test samples. In this notebook, we have used a supervision ratio of 0.05 (i.e, 5% of the labeled data) for training both the base model as well as the graph-regularized model. We illustrate the impact of the supervision ratio on model accuracy in the cell below.
In [0]:
# Accuracy values for both the Bi-LSTM model and the feed forward NN model have
# been precomputed for the following supervision ratios.
supervision_ratios = [0.3, 0.15, 0.05, 0.03, 0.02, 0.01, 0.005]
model_tags = ['Bi-LSTM model', 'Feed Forward NN model']
base_model_accs = [[84, 84, 83, 80, 65, 52, 50], [87, 86, 76, 74, 67, 52, 51]]
graph_reg_model_accs = [[84, 84, 83, 83, 65, 63, 50],
[87, 86, 80, 75, 67, 52, 50]]
plt.clf() # clear figure
fig, axes = plt.subplots(1, 2)
fig.set_size_inches((12, 5))
for ax, model_tag, base_model_acc, graph_reg_model_acc in zip(
axes, model_tags, base_model_accs, graph_reg_model_accs):
# "-r^" is for solid red line with triangle markers.
ax.plot(base_model_acc, '-r^', label='Base model')
# "-gD" is for solid green line with diamond markers.
ax.plot(graph_reg_model_acc, '-gD', label='Graph-regularized model')
ax.set_title(model_tag)
ax.set_xlabel('Supervision ratio')
ax.set_ylabel('Accuracy(%)')
ax.set_ylim((25, 100))
ax.set_xticks(range(len(supervision_ratios)))
ax.set_xticklabels(supervision_ratios)
ax.legend(loc='best')
plt.show()
It can be observed that as the superivision ratio decreases, model accuracy also decreases. This is true for both the base model and for the graph-regularized model, regardless of the model architecture used. However, notice that the graph-regularized model performs better than the base model for both the architectures. In particular, for the Bi-LSTM model, when the supervision ratio is 0.01, the accuracy of the graph-regularized model is ~20% higher than that of the base model. This is primarily because of semi-supervised learning for the graph-regularized model, where structural similarity among training samples is used in addition to the training samples themselves.
We have demonstrated the use of graph regularization using the Neural Structured Learning (NSL) framework even when the input does not contain an explicit graph. We considered the task of sentiment classification of IMDB movie reviews for which we synthesized a similarity graph based on review embeddings. We encourage users to experiment further by varying hyperparameters, the amount of supervision, and by using different model architectures.