def test_dropout_mlp():
# Run for a couple of iterations
# assert that new error is less than original
model = mlp.DropoutMLP((2, 8, 2))
dataset = datasets.get_and()
error = validation.get_error(model, *dataset)
model.train(*dataset, iterations=20)
assert validation.get_error(model, *dataset) < error
def test_dropout_mlp_classifier():
# Run for a couple of iterations
# assert that new error is less than original
model = mlp.DropoutMLP(
(2, 8, 2), transfers=SoftmaxTransfer(), error_func=CrossEntropyError())
dataset = datasets.get_and()
error = validation.get_error(model, *dataset)
model.train(*dataset, iterations=20)
assert validation.get_error(model, *dataset) < error
def test_dropout_mlp_dropout():
model = mlp.DropoutMLP(
(2, 4, 3), input_active_probability=0.5, hidden_active_probability=0.5)
# Only bias and active neurons should not be 0
model.train_step([[1, 1]], [[1, 1, 1]])
# Should still have DropoutTransfers (until activation outside of training)
_validate_weight_inputs(model._weight_inputs[0],
model._input_transfer._active_neurons)
for weight_inputs, transfer_func in zip(model._weight_inputs[1:-1],
model._transfers[:-1]):
_validate_weight_inputs(weight_inputs, transfer_func._active_neurons)
def test_dropout_mlp_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), input_active_probability=1.0, hidden_active_probability=0.9)
dataset = datasets.get_and() # Easier and dataset for lienar output
# 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
def test_dropout_mlp_post_training():
# Post training should happen on first activate after training (train step),
# and not more than once, unless training begins again
model = mlp.DropoutMLP(
(2, 4, 3), input_active_probability=0.5, hidden_active_probability=0.5)
# Train, modifying active neurons and weights
model.train_step([[1, 1]], [[1, 1, 1]])
pre_procedure_weights = copy.deepcopy(model._weight_matrices)
# Should call post_training procedure after activate
model.activate([1, 1])
_validate_post_training(model, pre_procedure_weights)
# Weights should not change after another activate
model.activate([1, 1])
_validate_post_training(model, pre_procedure_weights)
