Advanced Usage

Reusing Deep Features

Often it is necessary to train and evaluate more than one instance of a model on the same dataset. For example this is required by both K-Folds Cross Validation and Hyperparameter Tuning. In such cases much of the work across different runs can be reused.

Sound Classification is done using a three stage process. The work done in the first two of these stages (signal preprocessing and VGGish feature extraction) will be identical across runs and can be reused.

The code below use the ESC-10 dataset from the Introductory Example to do 5-folds cross validation:

import turicreate as tc

data = tc.load_sframe('./ESC-10')

# Calculate the deep features just once.
data['deep_features'] = tc.sound_classifier.get_deep_features(data['audio'])

accuracies = []
for cur_fold in data['fold'].unique():
    test_set = data.filter_by(cur_fold, 'fold')
    train_set = data.filter_by(cur_fold, 'fold', exclude=True)

    model = tc.sound_classifier.create(train_set,
                                       target='category',
                                       feature='deep_features')

    metrics = model.evaluate(test_set)
    accuracies.append(metrics['accuracy'])

print("Accuracies: {}".format(accuracies))

This allows us to perform 5-fold cross validation more than twice as fast. For larger datasets the time savings will be even greater.

Tune Custom Neural Network Configuration

The custom neural network used by the Sound Classifier is made up of a series of dense layers. Using more layers or more units in each layer can have a significant affect on accuracy. This will also affect the size of your model.

The custom_layer_sizes parameter allows you specify how many layers and the number of units in each layer. The default values for this parameter is [100, 100] which corresponds to two dense layers with a 100 units each.

Using a smaller number of overall units will result in a smaller model, potentially with minimal affects on accuracy. If you have a large amount of training data, you should get better accuracy by using more layers and/or more units.

The code below tries several different neural network configurations and reports the validation accuracy for each one:

import turicreate as tc

data = tc.load_sframe('./ESC-10')

# Calculate the deep features just once.
data['deep_features'] = tc.sound_classifier.get_deep_features(data['audio'])

# Try several different neural network configurations
models = []
network_configurations = ([100, 100], [100], [1000, 1000], [100, 100, 100])
for cur_hyper_parameter in network_configurations:
    cur_model = tc.sound_classifier.create(data, target='category',
                                             custom_layer_sizes=cur_hyper_parameter,
                                             feature='deep_features')
    models.append(cur_model)

# Report results
for m, p in zip(models, network_configurations):
    print("{}, {}".format(p, m.validation_accuracy))