def run():
training_set, evaluation_set = dat.get_data_sets()
sample = next(training_set)
n_pixels = np.prod(sample.shape)
printer = Printer(input_shape=sample.shape)
N_NODES = [24]
n_nodes = N_NODES + [n_pixels]
layers = []
layers.append(RangeNormalization(training_set))
for i_layer in range(len(n_nodes)):
new_layer = Dense(
n_nodes[i_layer],
activation_function=Tanh(),
# initializer=Glorot(),
# initializer=He(),
initializer=Uniform(scale=3),
previous_layer=layers[-1],
optimizer=Momentum(),
)
new_layer.add_regularizer(L1())
new_layer.add_regularizer(Limit(4.0))
layers.append(new_layer)
layers.append(Difference(layers[-1], layers[0]))
autoencoder = ANN(
layers=layers,
error_function=Sqr,
printer=printer,
)
msg = """
Running autoencoder demo
on Nordic Runes data set.
Find performance history plots,
model parameter report,
and neural network visualizations
in the {} directory.
""".format(autoencoder.reports_path)
print(msg)
autoencoder.train(training_set)
autoencoder.evaluate(evaluation_set)
def initialize(
activation_function=Tanh,
initializer=Glorot,
learning_rate=1e-4,
n_nodes_00=79,
n_nodes_0=23,
n_nodes_1=9,
n_nodes_2=23,
n_nodes_3=79,
patch_size=11,
**kwargs,
):
training_set, tuning_set, evaluation_set = ldr.get_data_sets(
patch_size=patch_size)
sample = next(training_set)
n_pixels = np.prod(sample.shape)
# n_nodes_dense = [n_nodes_00, n_nodes_0, n_nodes_1, n_nodes_2, n_nodes_3]
n_nodes_dense = [n_nodes_00, n_nodes_1, n_nodes_3]
# n_nodes_dense = [n_nodes_1]
n_nodes_dense = [n for n in n_nodes_dense if n > 0]
n_nodes = n_nodes_dense + [n_pixels]
layers = []
layers.append(RangeNormalization(training_set))
for i_layer in range(len(n_nodes)):
new_layer = Dense(n_nodes[i_layer],
activation_function=activation_function,
initializer=initializer,
previous_layer=layers[-1],
optimizer=Momentum(
learning_rate=learning_rate,
momentum_amount=.9,
))
layers.append(new_layer)
layers.append(Difference(layers[-1], layers[0]))
autoencoder = ANN(
layers=layers,
error_function=Sqr,
n_iter_train=5e4,
n_iter_evaluate=1e4,
n_iter_evaluate_hyperparameters=9,
verbose=False,
)
return autoencoder, training_set, tuning_set
def initialize(
limit=None,
L1_param=None,
L2_param=None,
learning_rate=None,
momentum=.9,
**kwargs,
):
training_set, tuning_set, evaluation_set = ldr.get_data_sets()
sample = next(training_set)
n_pixels = np.prod(sample.shape)
N_NODES = [33]
n_nodes = N_NODES + [n_pixels]
layers = []
layers.append(RangeNormalization(training_set))
for i_layer in range(len(n_nodes)):
new_layer = Dense(
n_nodes[i_layer],
activation_function=Tanh,
initializer=Glorot(),
previous_layer=layers[-1],
optimizer=Momentum(
learning_rate=learning_rate,
momentum_amount=momentum,
),
)
if limit is not None:
new_layer.add_regularizer(Limit(limit))
if L1_param is not None:
new_layer.add_regularizer(L1(L1_param))
if L2_param is not None:
new_layer.add_regularizer(L2(L2_param))
layers.append(new_layer)
layers.append(Difference(layers[-1], layers[0]))
autoencoder = ANN(
layers=layers,
error_function=Sqr,
n_iter_train=5e4,
n_iter_evaluate=1e4,
verbose=False,
)
return autoencoder, training_set, tuning_set
def run():
training_set, tuning_set, evaluation_set = ldr.get_data_sets()
sample = next(training_set)
n_pixels = np.prod(sample.shape)
printer = Printer(input_shape=sample.shape)
N_NODES = [64, 36, 24, 36, 64]
# N_NODES = [64]
n_nodes = N_NODES + [n_pixels]
layers = []
layers.append(RangeNormalization(training_set))
for i_layer in range(len(n_nodes)):
new_layer = Dense(
n_nodes[i_layer],
activation_function=Tanh,
initializer=Glorot(),
previous_layer=layers[-1],
optimizer=Momentum(),
)
# new_layer.add_regularizer(L1())
new_layer.add_regularizer(Limit(4.0))
layers.append(new_layer)
layers.append(Difference(layers[-1], layers[0]))
autoencoder = ANN(
layers=layers,
error_function=Sqr,
printer=printer,
)
msg = """
Running autoencoder on images of the surface of Mars.
Find performance history plots, model parameter report,
and neural network visualizations in the directory
{}
""".format(autoencoder.reports_path)
print(msg)
autoencoder.train(training_set)
autoencoder.evaluate(tuning_set)
autoencoder.evaluate(evaluation_set)
def initialize():
training_set, tuning_set, evaluation_set = ldr.get_data_sets()
sample = next(training_set)
n_pixels = np.prod(sample.shape)
printer = Printer(input_shape=sample.shape)
N_NODES = [33]
n_nodes = N_NODES + [n_pixels]
layers = []
layers.append(RangeNormalization(training_set))
for i_layer in range(len(n_nodes)):
new_layer = Dense(
n_nodes[i_layer],
activation_function=Tanh,
initializer=Glorot(),
previous_layer=layers[-1],
optimizer=SGD(),
)
new_layer.add_regularizer(Limit(4.0))
layers.append(new_layer)
layers.append(Difference(layers[-1], layers[0]))
autoencoder = ANN(
layers=layers,
error_function=Sqr,
n_iter_train=N_ITER_TRAIN,
n_iter_evaluate=N_ITER_EVALUATE,
printer=printer,
verbose=False,
)
return autoencoder, training_set, tuning_set
def train(
image_path,
activation_function=Tanh,
initializer=Glorot,
learning_rate=1e-4,
n_nodes_0=79,
n_nodes_1=9,
n_nodes_2=79,
):
"""
Train an autoencoder to represent image patches more economically.
image_path: str, a path to the directory containing the images
that are to be compressed. If this is a relative path, it needs to be
relative to the directory from which this module is run.
activation_function: one of the classes available in
cottonwood/core/activation_functions.py
As of this writing, {Tanh, Sigmoid, ReLU}
initializer: one of the classes available in
cottonwood/core/initializers.py
As of this writing, {Glorot, He}
learning_rate: float, the learning rate for the Momentum optimizers
that gets called during backpropagation. Feasible values will probably
be between 1e-5 and 1e-3.
n_nodes_x: int, the number of nodes in layer x. Layer 1 is
the narrowest layer, and its node activities
are used as the representation
of the compressed patch.
returns a trained autoencoder
"""
training_patches = ldr.get_training_data(patch_size, image_path)
sample = next(training_patches)
printer = Printer(input_shape=sample.shape)
n_pixels = np.prod(sample.shape)
n_nodes_dense = [n_nodes_0, n_nodes_1, n_nodes_2]
n_nodes = n_nodes_dense + [n_pixels]
printer = Printer(input_shape=sample.shape)
layers = []
layers.append(RangeNormalization(training_patches))
for i_layer in range(len(n_nodes)):
new_layer = Dense(n_nodes[i_layer],
activation_function=activation_function(),
initializer=initializer(),
previous_layer=layers[-1],
optimizer=Momentum(
learning_rate=learning_rate,
momentum_amount=.9,
))
layers.append(new_layer)
layers.append(Difference(layers[-1], layers[0]))
autoencoder = ANN(
layers=layers,
error_function=Sqr,
n_iter_train=5e4,
n_iter_evaluate=1e4,
n_iter_evaluate_hyperparameters=9,
printer=printer,
verbose=True,
viz_interval=1e4,
)
autoencoder.train(training_patches)
return autoencoder