def test_array_param(self):
# Check that the default initialization is finite.
default_init = ArrayParam()
self.assertTrue(np.isfinite(default_init.get()).all())
lb = -0.1
ub = 5.2
shape = (3, 2)
val = np.random.random(shape) * (ub - lb) + lb
bad_val_ub = np.abs(val) + ub
bad_val_lb = lb - np.abs(val)
ap = ArrayParam('test', shape, lb=lb - 0.001, ub=ub + 0.001)
# Check setting.
ap_init = ArrayParam('test',
shape,
lb=lb - 0.001,
ub=ub + 0.001,
val=val)
np_test.assert_array_almost_equal(val, ap_init.get())
self.assertRaises(ValueError, ap.set, val[-1, :])
# Bounds checking is disabled for now until it can be made optional.
#self.assertRaises(ValueError, ap.set, bad_val_ub)
#self.assertRaises(ValueError, ap.set, bad_val_lb)
ap.set(val)
# Check size.
self.assertEqual(np.product(shape), ap.vector_size())
self.assertEqual(np.product(shape), ap.free_size())
# Check getting and free parameters.
np_test.assert_array_almost_equal(val, ap.get())
val_free = ap.get_free()
ap.set(np.full(shape, 0.))
ap.set_free(val_free)
np_test.assert_array_almost_equal(val, ap.get())
val_vec = ap.get_vector()
ap.set(np.full(shape, 0.))
ap.set_vector(val_vec)
np_test.assert_array_almost_equal(val, ap.get())
def dcov(x, y):
dnx = d_n(x)
dny = d_n(y)
denom = np.product(dnx.shape)
dc = (dnx * dny).sum() / denom
dvx = (dnx**2).sum() / denom
dvy = (dny**2).sum() / denom
dr = dc / (np.sqrt(dvx) * np.sqrt(dvy))
return dc #, dr, dvx, dvy
def unpackage_params(d, shapes=[]):
flat_params = unpack(d)
params = []
idx = 0
for W_shape, b_shape in shapes:
W_len = np.product(W_shape)
b_len = np.product(b_shape)
W = (flat_params[idx:idx + W_len]).reshape(*W_shape)
idx += W_len
b = (flat_params[idx:idx + b_len]).reshape(*b_shape)
idx += b_len
params.append((W, b))
return params
def loss_function_kernel(self, weights, x_in, ind):
"""Compute loss function with or without penalty factor."""
net = self.mlp.network_output(x_in, weights)
ref = self.reference_solution(ind)
p_factor = 1.0
is_boundary = (np.product(x_in) < 1.0E-12) or \
(abs(x_in[0] - 1.0) < 1.0E-12) or \
(abs(x_in[1] - 1.0) < 1.0E-12)
if is_boundary:
p_factor = self.penalty
return ((ref - net) * p_factor)**2
def place_control_points(self,
method=[],
domain=[],
step_size=[],
growths=[]):
"""Places control points within a given domain."""
[
self.mlp.assert_input_length(input_list)
for input_list in [method, domain, step_size, growths]
]
print("\nPlacing control points for {} input variables.".format(
len(domain)))
print("------------------------------------------------")
current_var = [domain[i][0] + step_size[i] for i in range(len(domain))]
points = [[domain[i][0]] for i in range(len(domain))]
lengths = []
for i, method_i in enumerate(method):
if method_i == 'uniform':
while (current_var[i] + 1.0E-6) < domain[i][1]:
points[i].append(current_var[i])
current_var[i] += step_size[i]
else:
while (current_var[i] + 1.0E-6) < domain[i][1]:
points[i].append(current_var[i])
current_var[i] *= growths[i]
points[i].append(domain[i][1])
lengths.append(len(points[i]))
total = np.product(np.asarray(lengths))
all_points = []
assert self.mlp.number_of_inputs <= 3
if self.mlp.number_of_inputs == 1:
all_points = np.asarray(points[0])
elif self.mlp.number_of_inputs == 2:
for p0 in points[0]:
for p1 in points[1]:
all_points.append(np.asarray([p0, p1]))
elif self.mlp.number_of_inputs == 3:
for p0 in points[0]:
for p1 in points[1]:
for p2 in points[2]:
all_points.append(np.asarray([p0, p1, p2]))
self.control_points = np.asarray(all_points)
self.loss_control_points = self.control_points
self.loss_indices = range(len(self.control_points))
print("Control points per input parameter: {}".format(lengths))
print("The total number of control points is {}".format(total))
def __init__(self,
shape,
lb=-float("inf"),
ub=float("inf"),
default_validate=True,
free_default=None):
"""
Parameters
-------------
shape: `tuple` of `int`
The shape of the array.
lb: `float`
The (inclusive) lower bound for the entries of the array.
ub: `float`
The (inclusive) upper bound for the entries of the array.
default_validate: `bool`, optional
Whether or not the array is checked by default to lie within the
specified bounds.
free_default: `bool`, optional
Whether the pattern is free by default.
"""
self.default_validate = default_validate
self._shape = tuple(shape)
self._lb = lb
self._ub = ub
assert lb >= -float('inf')
assert ub <= float('inf')
if lb >= ub:
raise ValueError(
'Upper bound ub must strictly exceed lower bound lb')
free_flat_length = flat_length = int(np.product(self._shape))
super().__init__(flat_length,
free_flat_length,
free_default=free_default)
# Cache arrays of indices for flat_indices
# TODO: not sure this is a good idea or much of a speedup.
self.__free_folded_indices = self.fold(np.arange(
self.flat_length(free=True), dtype=int),
validate_value=False,
free=False)
self.__nonfree_folded_indices = self.fold(np.arange(
self.flat_length(free=False), dtype=int),
validate_value=False,
free=False)
def flatten(x):
return np.reshape(x, np.product(x.shape))
def vector_size(self):
return np.product(self.__shape)
def free_size(self):
return np.product(self.__free_shape)
def initialize_weights(self, method):
"""Intialize weights and biases.
Biases are always initialized with zero.
The weights can be chosen randomly or with the Xavier-He
initialization. The MLP has one input layer, at least one
hidden layers, and one output layer. All layers except for
the output have exactly one bias unit. All biases and weights
(input and output) are stored in the array 'weights'.
Parameters
----------
method - string : Defines the initialization method;
Can be either 'random' or 'xavier_he'
Set/Return
----------
weights - array-like : 2D array containing all weights and biases;
It has the shape [rows, columns] with
rows = rows_input + rows_hidden + rows_output
(further explanation in the code)
columns = neurons_per_layer (all hidden layers
have the same number of neurons)
[0, :] accesses the input bias
[1, :] accesses the weights between the first
input neuron and all neurons of the first hidden
layer
...
[number_of_inputs + 1, :] accesses the bias of
the first hidden layer
[number_of_inputs + 2, :] accesses the weights
between the first neuron of the first hidden
layer and all neurons of the second hidden layer
...
[(number_of_inputs + 1) * (layer - 1), :]
accesses the bias of 'layer' with layer=1 for
the first hidden layer and so on
"""
# each input is connected to each neuron in the first hidden layer;
# plus one for the bias
input_rows = self.number_of_inputs + 1
# each neuron in the hidden layer is connected to each neuron in the
# next layer; plus one for the bias
hidden_rows = self.neurons_per_layer + 1
# +1 for the weights between last hidden layer and output neuron
rows = input_rows + hidden_rows * (self.hidden_layers - 1) + 1
weights = np.zeros([rows, self.neurons_per_layer]).astype(np.float64)
np.random.seed(1111)
if method == 'random':
signs = self.random_signs(rows, self.neurons_per_layer)
weights = np.multiply(np.random.rand(rows, self.neurons_per_layer),
signs)
elif method == 'xavier_he':
stdv_in = np.sqrt(6 / self.number_of_inputs)
stdv_rest = np.sqrt(6 / self.neurons_per_layer)
# input weights
signs = self.random_signs(self.number_of_inputs,
self.neurons_per_layer)
weights[1:self.number_of_inputs + 1] = np.multiply(
np.random.rand(self.number_of_inputs, self.neurons_per_layer),
signs) * stdv_in
# weights of hidden layers
start_hidden = self.number_of_inputs + 1
for layer in range(self.hidden_layers - 1):
signs = self.random_signs(self.neurons_per_layer,
self.neurons_per_layer)
start = start_hidden + (self.neurons_per_layer + 1) * layer + 1
end = start + self.neurons_per_layer
weights[start:end] = np.multiply(
np.random.rand(self.neurons_per_layer,
self.neurons_per_layer), signs) * stdv_rest
# output weights
signs = self.random_signs(1, self.neurons_per_layer)
weights[-1] = np.multiply(
np.random.rand(1, self.neurons_per_layer), signs) * stdv_rest
else:
print(
"Could not find method {} for initialization.".format(method))
print("Possible options are 'random', 'xavier_he'")
print("\nInitialized weights and biases for MLP with {} parameters.\n".
format(rows * self.neurons_per_layer))
print("MLP structure:")
print("--------------")
print("Input features: {}".format(self.number_of_inputs))
print("Hidden layers: {}".format(self.hidden_layers))
print("Neurons per layer: {}".format(self.neurons_per_layer))
print("Number of weights: {}".format(np.product(weights.shape)))
return weights
def free_size(self):
return int(np.product(self.__shape))
