def test_mlp_classifier_convergence():
# Run until convergence
# assert that network can converge
model = mlp.MLP(
(2, 3, 2), transfers=SoftmaxTransfer(), error_func=CrossEntropyError())
dataset = datasets.get_and()
model.train(*dataset, retries=5, error_break=0.002)
assert validation.get_error(model, *dataset) <= 0.02
def test_mlp_classifier():
# Run for a couple of iterations
# assert that new error is less than original
model = mlp.MLP(
(2, 2, 2), transfers=SoftmaxTransfer(), error_func=CrossEntropyError())
dataset = datasets.get_xor()
error = validation.get_error(model, *dataset)
model.train(*dataset, iterations=20)
assert validation.get_error(model, *dataset) < error
def test_dropout_mlp_classifier_convergence():
# Run until convergence
# assert that network can converge
# Since XOR does not really need dropout, we use high probabilities
model = mlp.DropoutMLP(
(2, 8, 2),
transfers=SoftmaxTransfer(),
error_func=CrossEntropyError(),
input_active_probability=1.0,
hidden_active_probability=0.9)
dataset = datasets.get_and()
# Error break lower than cutoff, since dropout may have different error
# after training
model.train(*dataset, retries=5, error_break=0.002, error_improve_iters=50)
# Dropout sacrifices training accuracy for better generalization
# so we don't worry as much about convergence
assert validation.get_error(model, *dataset) <= 0.1
dataset = datasets.get_iris()
# Make a multilayer perceptron to classify the iris dataset
model = MLP(
# The MLP will take 4 attributes, have 1 hidden layer with 2 neurons,
# and outputs one of 3 classes
(4, 2, 3),
# We will use a softmax output layer for this classification problem
# Because we are only changing the output transfer, we pass a single
# Transfer object. We could customize all transfer layers by passing
# a list of Transfer objects.
transfers=SoftmaxTransfer(),
# Cross entropy error will pair nicely with our softmax output.
error_func=CrossEntropyError(),
# Lets use the quasi-newton BFGS optimizer for this problem
# BFGS requires and n^2 operation, where n is the number of weights,
# but this isn't a problem for our relatively small MLP.
# If we don't want to deal with optimizers, the default
# option will select an appropriate optimizer for us.
optimizer=optimize.BFGS(
# We can even customize the line search method
step_size_getter=optimize.WolfeLineSearch(
# And the initial step size for our line search
initial_step_getter=optimize.FOChangeInitialStep())))
# NOTE: For rapid prototyping, we could quickly implement an MLP as follows
# model = MLP((4, 2, 3))
def test_mlp_jacobian_softmax_out_ce():
_check_jacobian(lambda s1, s2, s3: mlp.MLP(
(s1, s2, s3), transfers=SoftmaxTransfer(), error_func=CrossEntropyError()))
def test_mlp_jacobian_relu_out_ce():
_check_jacobian(lambda s1, s2, s3: mlp.MLP(
(s1, s2, s3), transfers=mlp.ReluTransfer(), error_func=CrossEntropyError()))
def test_mlp_obj_and_obj_jac_match_softmax_out_ce():
_check_obj_and_obj_jac_match(
lambda s1, s2, s3: mlp.MLP(
(s1, s2, s3), transfers=SoftmaxTransfer(), error_func=CrossEntropyError()),
classification=True
)