def test_InceptionTime_hyperparameters(self):
""" Network depth from range [5,5] should be 5.
Maximum kernal size from range [12, 12] should be 12.
Minimum filter number from range [32, 32] should be 32. """
custom_settings = get_default()
x_shape = (None, 20, 3)
kwargs = {
'IT_min_network_depth': 5,
'IT_max_network_depth': 5,
'IT_min_max_kernel_size': 10,
'IT_max_max_kernel_size': 10,
'IT_min_filters_number': 32,
'IT_max_filters_number': 32
}
# Replace default parameters with input
for key, value in kwargs.items():
if key in custom_settings:
custom_settings[key] = value
model_type = InceptionTime(x_shape, None, **custom_settings)
hyperparams = model_type.generate_hyperparameters()
assert hyperparams.get('network_depth') == 5, 'Wrong network depth'
assert hyperparams.get('max_kernel_size') == 10, 'Wrong kernel'
assert hyperparams.get('filters_number') == 32, 'Wrong filter number'
def test_InceptionTime_starts_with_batchnorm(self):
""" InceptionTime models should always start with a batch normalization layer. """
model_type = InceptionTime((None, 20, 3), 2)
model = model_type.create_model(16)
assert 'BatchNormalization' in str(type(
model.layers[1])), 'Wrong layer type.'
def test_InceptionTime_first_inception_module(self):
""" Test layers of first inception module. """
model_type = InceptionTime((None, 20, 3), 2)
model = model_type.create_model(16)
assert 'Conv1D' or 'Convolution1D' in str(type(model.layers[2])), 'Wrong layer type.'
assert 'MaxPooling1D' in str(type(model.layers[3])), 'Wrong layer type.'
assert 'Concatenate' in str(type(model.layers[8])), 'Wrong layer type.'
def test_InceptionTime_depth(self):
""" InceptionTime model should have depth (number of residual modules) as defined by user. """
depths = 3
model_type = InceptionTime((None, 20, 3), 2)
model = model_type.create_model(16, network_depth=depths)
concat_layers = [str(type(layer)) for layer in model.layers if 'concatenate' in str(type(layer)).lower()]
assert len(concat_layers) == depths, 'Wrong number of inception modules (network depths).'
def test_InceptionTime_first_module_dim(self):
""""The output shape throughout the first residual module should be (None, nr_timesteps, min_filters_number)"""
min_filters_number = 16
model_type = InceptionTime((None, 30, 5), 2)
model = model_type.create_model(min_filters_number)
secondConvlayer = model.layers[5]
firstConcatlayer = model.layers[8]
assert secondConvlayer.output_shape == (None, 30, min_filters_number)
assert firstConcatlayer.output_shape == (None, 30, min_filters_number * 4)
def test_InceptionTime_metrics(self):
"""InceptionTime model should be compiled with the metrics that we give it"""
metrics = ['accuracy', 'mae']
x_shape = (None, 20, 3)
nr_classes = 2
X_train, y_train = generate_train_data(x_shape, nr_classes)
model_type = InceptionTime(x_shape, nr_classes, metrics=metrics)
model = model_type.create_model(16)
model.fit(X_train, y_train)
model_metrics = [m.name for m in model.metrics]
for metric in metrics:
assert metric in model_metrics