def full_linear_network():
# Simple Linear unit
n_in = N_full_grid_extractor
n_out = 1
theta = 0.05
layers = [
layer.FullyConnected(n_in, n_out, activation=activation.Linear(),
weight_init=misc.normalised_initialiser(1))
]
return network.Network(layers, theta, loss=activation.MSE())
def two_layer_sigmoid_network():
n_in = N_simple_grid_extractor
n_out1 = n_in // 2
n_out2 = 1
theta = 0.05
layers = [
layer.FullyConnected(n_in, n_out1, activation=activation.Sigmoid(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out1, n_out2, activation=activation.Linear(),
weight_init=misc.normalised_initialiser(1)),
]
return network.Network(layers, theta, loss=activation.MSE())
def full_two_layer_sigmoid_network_sigoutput():
n_in = N_full_grid_extractor
n_out1 = n_in // 2
n_out2 = 1
theta = 0.05
layers = [
layer.FullyConnected(n_in, n_out1, activation=activation.Sigmoid(),
weight_init=misc.zero_initialiser(1)),
layer.FullyConnected(n_out1, n_out2, activation=activation.Sigmoid(),
weight_init=misc.zero_initialiser(1)),
]
return network.Network(layers, theta, loss=activation.XEntropy())
def full4Sigmoid():
n_in = N_full_grid_extractor
n_out1 = n_in * 2 // 3
n_out2 = n_out1 * 2 // 3
n_out3 = n_out2 * 2 // 3
n_out4 = 1
theta = 0.05
layers = [
layer.FullyConnected(n_in, n_out1, activation=activation.Sigmoid(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out1, n_out2, activation=activation.Sigmoid(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out2, n_out3, activation=activation.Sigmoid(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out3, n_out4, activation=activation.Linear(),
weight_init=misc.normalised_initialiser(1)),
]
return network.Network(layers, theta, loss=activation.MSE())
def four_layer_leaky_relu_network():
n_in = N_simple_grid_extractor
n_out1 = n_in * 2 // 3
n_out2 = n_out1 * 2 // 3
n_out3 = n_out2 * 2 // 3
n_out4 = 1
theta = 0.05
layers = [
layer.FullyConnected(n_in, n_out1, activation=activation.LeakyRelu(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out1, n_out2, activation=activation.LeakyRelu(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out2, n_out3, activation=activation.LeakyRelu(),
weight_init=misc.normalised_initialiser(1)),
layer.FullyConnected(n_out3, n_out4, activation=activation.LeakyRelu(),
weight_init=misc.normalised_initialiser(1)),
]
return network.Network(layers, theta, loss=activation.MSE())