Neuralnet Classifier
The deep learning module is useful to create and manipulate different neural network architectures. The core of this module is the NeuralNet class, which stores the definition of each layer of a neural network and a dictionary of learning parameters.
A NeuralNet object can be obtained from graphlab.deeplearning.create.
The function, selects a default network architecture depending on the input
dataset using simple rules: it creates a 2-layer Perceptron Network for dense
numeric input, and a 1-layer Convolution Network for image data input.
Warning: Due to the high complexity of netualnet models, the default network does not always work out of the box, and you will often need to tweak the architectures a bit to make it work for your problem.
Introductory Example: Digit Recognition on MNIST Data
In this example, we will train a covolutional neural network for digit recognition. We need to make sure all of the images are the same size, since neural nets have fixed input size. We can use the resize function. The setup code to get started is as follows:
import graphlab as gl
# Load the MNIST data (from an S3 bucket)
data = gl.SFrame('http://s3.amazonaws.com/dato-datasets/mnist/sframe/train')
test_data = gl.SFrame('http://s3.amazonaws.com/dato-datasets/mnist/sframe/test')
# Random split the training-data
training_data, validation_data = data.random_split(0.8)
# Make sure all images are of the same size (Required by Neuralnets)
for sf in [training_data, validation_data, test_data]:
sf['image'] = gl.image_analysis.resize(sf['image'], 28, 28, 1)
We will use the builtin NeuralNet architecture for MNIST (a one layer convolutional neuralnet work). The layers of the network can be viewed as follows.
net = gl.deeplearning.get_builtin_neuralnet('mnist')
print "Layers of the network "
print "--------------------------------------------------------"
print net.layers
print "Parameters of the network "
print "--------------------------------------------------------"
print net.params
Layers of the network
--------------------------------------------------------
layer[0]: ConvolutionLayer
padding = 1
stride = 2
random_type = xavier
num_channels = 32
kernel_size = 3
layer[1]: MaxPoolingLayer
stride = 2
kernel_size = 3
layer[2]: FlattenLayer
layer[3]: DropoutLayer
threshold = 0.5
layer[4]: FullConnectionLayer
init_sigma = 0.01
num_hidden_units = 100
layer[5]: SigmoidLayer
layer[6]: FullConnectionLayer
init_sigma = 0.01
num_hidden_units = 10
layer[7]: SoftmaxLayer
Parameters of the network
--------------------------------------------------------
{'batch_size': 100,
'divideby': 255,
'init_random': 'gaussian',
'l2_regularization': 0.0,
'learning_rate': 0.1,
'momentum': 0.9}
We can now train the neural network using the specified network as follows:
model = gl.neuralnet_classifier.create(training_data, target='label',
network = net,
validation_set=validation_data,
metric=['accuracy', 'recall@2'],
max_iterations=3)
Making Predictions
We can now classify the test data, and output the most likely class label. The score corresponds to the learned probability of the testing instance belonging to the predicted class.
Note that this is inherently a multi-class classification problem, so the classify provides the top label predictions for each data point along with a probability/confidence of the class prediction.
predictions = model.classify(test_data)
print predictions
+--------+-------+----------------+
| row_id | class | score |
+--------+-------+----------------+
| 0 | 0 | 0.998417854309 |
| 1 | 0 | 0.999230742455 |
| 2 | 0 | 0.999326109886 |
| 3 | 0 | 0.997855246067 |
| 4 | 0 | 0.997171103954 |
| 5 | 0 | 0.996235311031 |
| 6 | 0 | 0.999143242836 |
| 7 | 0 | 0.999519705772 |
| 8 | 0 | 0.999182283878 |
| 9 | 0 | 0.999905228615 |
| ... | ... | ... |
+--------+-------+----------------+
[10000 rows x 3 columns]
Making Detailed Predictions
We can use the predict_topk interface if we want prediction scores for
each class in the top-k classes (sorted in decreasing order of score).
Predict the top 2 most likely digits
pred_top2 = model.predict_topk(test_data, k=2)
print pred_top2
+--------+-------+-------------------+
| row_id | class | score |
+--------+-------+-------------------+
| 0 | 0 | 0.998417854309 |
| 0 | 6 | 0.000686840794515 |
| 1 | 0 | 0.999230742455 |
| 1 | 2 | 0.000284609268419 |
| 2 | 0 | 0.999326109886 |
| 2 | 8 | 0.000261707202299 |
| 3 | 0 | 0.997855246067 |
| 3 | 8 | 0.00118813838344 |
| 4 | 0 | 0.997171103954 |
| 4 | 6 | 0.00115600414574 |
| ... | ... | ... |
+--------+-------+-------------------+
[20000 rows x 3 columns]
Evaluating the Model
We can evaluate the classifier on the test data. Default metrics are accuracy, and confusion matrix.
result = model.evaluate(test_data)
print "Accuracy : %s" % result['accuracy']
print "Confusion Matrix : \n%s" % result['confusion_matrix']
Accuracy : 0.977599978447
Confusion Matrix :
+--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
| 0 | 0 | 973 |
| 2 | 0 | 4 |
| 4 | 0 | 1 |
| 5 | 0 | 2 |
| 6 | 0 | 9 |
| 7 | 0 | 1 |
| 8 | 0 | 1 |
| 9 | 0 | 3 |
| 1 | 1 | 1122 |
| 2 | 1 | 1 |
| ... | ... | ... |
+--------------+-----------------+-------+
[65 rows x 3 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
Using a Neural Network for Feature Extraction
A previously trained model can be used to extract dense features for a given input. Theextract_features function takes an input dataset, propagates each example through the network, and returns an SArray of dense feature vectors, each of which is the concatenation of all the hidden unit values at layer[layer_id]. These feature vectors can be used as input to train another classifier such as a LogisticClassifier, an SVMClassifier, another NeuralNetClassifier, or BoostedTreesClassifier. Input dataset size must be the same as for the training of the model, except for images which are automatically resized.
If the original network is trained on a large dataset, these deep features can be very powerful. This is especially true in the context of image analysis, where a model trained on the very large ImageNet dataset can learn general purpose features.
In this example, we will build a neural network for classification of digits, then build a generic classifier on top of those extracted features.
# The data is the MNIST digit recognition dataset
data = graphlab.SFrame('http://s3.amazonaws.com/dato-datasets/mnist/sframe/train6k')
net = graphlab.deeplearning.get_builtin_neuralnet('mnist')
m = graphlab.neuralnet_classifier.create(data,
target='label',
network=net,
max_iterations=3)
# Now, let's extract features from the last layer
data['features'] = m.extract_features(data)
# Now, let's build a new classifier on top of extracted features
m = graphlab.classifier.create(data,
features = ['features'],
target='label')
We also provide a model trained on Imagenet.This pre-trained model gives pre-trained features of excellent quality for images, and the way you would use such a model is demonstrated below:
imagenet_path = 'http://s3.amazonaws.com/dato-datasets/deeplearning/imagenet_model_iter45'
imagenet_model = graphlab.load_model(imagenet_path)
data['image'] = graphlab.image_analysis.resize(data['image'], 256, 256, 3)
data['imagenet_features'] = imagenet_model.extract_features(data)
One can also use our feature engineering tools for extracting deep features.