def __init__(self, num_features=93, num_classes=9):
"""
"""
self.prefix = 'nl_simp'
self.num_features = num_features
self.num_classes = num_classes
self.layers = [('input', InputLayer),
('dropouti', DropoutLayer),
('dense0', DenseLayer),
('dropout0', DropoutLayer),
('dense1', DenseLayer),
('dropout1', DropoutLayer),
('dense2', DenseLayer),
('dropout2', DropoutLayer),
('dense3', DenseLayer),
('dropout3', DropoutLayer),
('output', DenseLayer)]
self.early_stopping = EarlyStopping(patience=25)
self.nn = My_NeuralNet(layers=self.layers,
input_shape=(None, self.num_features),
dropouti_p=0.12,
dense0_num_units=900,
dense0_W=HeNormal(),
dense0_nonlinearity = LeakyRectify(leakiness=0.002),
dropout0_p=0.35,
dense1_num_units=600,
dense1_W=HeNormal(),
dense1_nonlinearity = LeakyRectify(leakiness=0.002),
dropout1_p=0.2,
dense2_num_units=400,
dense2_W=HeNormal(),
dense2_nonlinearity = LeakyRectify(leakiness=0.002),
dropout2_p=0.1,
dense3_num_units=300,
dense3_W=HeNormal(),
dense3_nonlinearity = LeakyRectify(leakiness=0.002),
dropout3_p=0.1,
output_num_units=num_classes,
output_nonlinearity=softmax,
update=adagrad,
update_learning_rate=shared(float32(0.01)),
batch_iterator_train=BatchIterator(batch_size=512),
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=0.01,
stop=0.005),
# AdjustVariable('update_momentum', start=0.1, stop=0.0),
self.early_stopping,
],
eval_size=0.,
verbose=1,
max_epochs=300
)
self.nn2 = deepcopy(self.nn)
def __init__(self, num_features=93, num_classes=9):
"""
"""
self.prefix = 'nl_2lv_cal'
self.num_features = num_features
self.num_classes = num_classes
self.layers0 = [('input', InputLayer),
('dropoutn', DropoutLayer),
('dense0', DenseLayer),
('dropout0', DropoutLayer),
('dense1', DenseLayer),
('dropout1', DropoutLayer),
('dense2', DenseLayer),
('dropout2', DropoutLayer),
('dense3', DenseLayer),
('dropout3', DropoutLayer),
('output', DenseLayer)]
self.early_stopping0 = EarlyStopping(patience=20)
self.nn0 = My_NeuralNet(layers=self.layers0,
input_shape=(None, self.num_features),
dropoutn_p=0.12,
dense0_num_units=900,
dense0_W=HeNormal(),
dense0_nonlinearity = LeakyRectify(leakiness=0.002),
dropout0_p=0.35,
dense1_num_units=600,
dense1_W=HeNormal(),
dense1_nonlinearity = LeakyRectify(leakiness=0.002),
dropout1_p=0.2,
dense2_num_units=400,
dense2_W=HeNormal(),
dense2_nonlinearity = LeakyRectify(leakiness=0.002),
dropout2_p=0.1,
dense3_num_units=300,
dense3_W=HeNormal(),
dense3_nonlinearity = LeakyRectify(leakiness=0.002),
dropout3_p=0.1,
output_num_units=2,
output_nonlinearity=softmax,
update=adagrad,
update_learning_rate=shared(float32(0.01)),
batch_iterator_train=BatchIterator(batch_size=256),
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=0.01, stop=0.001),
# AdjustVariable('update_momentum', start=0.1, stop=0.0),
self.early_stopping0,
],
eval_size=0,
verbose=1,
max_epochs=36
)
self.layers1 = [('input', InputLayer),
('dropoutn', DropoutLayer),
('dense0', DenseLayer),
# ('gaussian0', GaussianNoiseLayer),
('dropout0', DropoutLayer),
('dense1', DenseLayer),
# ('gaussian1', GaussianNoiseLayer),
('dropout1', DropoutLayer),
('dense2', DenseLayer),
('dropout2', DropoutLayer),
# ('dense3', DenseLayer),
# ('dropout3', DropoutLayer),
('output', DenseLayer)]
self.early_stopping1 = EarlyStopping(patience=20)
self.nn1 = My_NeuralNet(layers=self.layers1,
input_shape=(None, self.num_features),
dropoutn_p=0.18,
dense0_num_units=900,
dense0_W=HeNormal(),
dense0_nonlinearity = LeakyRectify(leakiness=0.002),
dropout0_p=0.35,
dense1_num_units=600,
dense1_W=HeNormal(),
dense1_nonlinearity = LeakyRectify(leakiness=0.002),
dropout1_p=0.15,
dense2_num_units=400,
dense2_W=HeNormal(),
dense2_nonlinearity = LeakyRectify(leakiness=0.002),
dropout2_p=0.05,
output_num_units=3,
output_nonlinearity=softmax,
update=adagrad,
update_learning_rate=shared(float32(0.01)),
batch_iterator_train=BatchIterator(batch_size=256),
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=0.01, stop=0.001),
# AdjustVariable('update_momentum', start=0.1, stop=0.0),
self.early_stopping1,
],
eval_size=0,
verbose=1,
max_epochs=60
)
self.layers2 = [('input', InputLayer),
('dropoutn', DropoutLayer),
('dense0', DenseLayer),
# ('gaussian0', GaussianNoiseLayer),
('dropout0', DropoutLayer),
('dense1', DenseLayer),
# ('gaussian1', GaussianNoiseLayer),
('dropout1', DropoutLayer),
('dense2', DenseLayer),
('dropout2', DropoutLayer),
# ('dense3', DenseLayer),
# ('dropout3', DropoutLayer),
('output', DenseLayer)]
self.early_stopping2 = EarlyStopping(patience=20)
self.nn2 = My_NeuralNet(layers=self.layers2,
input_shape=(None, self.num_features),
dropoutn_p=0.13,
dense0_num_units=900,
dense0_W=HeNormal(),
dense0_nonlinearity = LeakyRectify(leakiness=0.002),
dropout0_p=0.35,
dense1_num_units=600,
dense1_W=HeNormal(),
dense1_nonlinearity = LeakyRectify(leakiness=0.002),
dropout1_p=0.15,
dense2_num_units=400,
dense2_W=HeNormal(),
dense2_nonlinearity = LeakyRectify(leakiness=0.002),
dropout2_p=0.05,
output_num_units=6,
output_nonlinearity=softmax,
update=adagrad,
update_learning_rate=shared(float32(0.01)),
batch_iterator_train=BatchIterator(batch_size=256),
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=0.01, stop=0.001),
# AdjustVariable('update_momentum', start=0.1, stop=0.0),
self.early_stopping2,
],
# eval_size=0.2,
eval_size=0,
verbose=1,
max_epochs=31
)
