def parse_properties(self, n_in, n_hidden, n_out, det_activations, stoch_activations,
stoch_n_hidden):
self.n_hidden = np.array(n_hidden)
self.n_out = n_out
self.n_in = n_in
self.stoch_n_hidden = [np.array(i) for i in stoch_n_hidden]
self.det_activation_names = det_activations
self.det_activation, self.det_activation_prime = parse_activations(det_activations)
self.stoch_activation_names = stoch_activations
self.stoch_activation, self.stoch_activation_prime = parse_activations(stoch_activations)
def parse_properties(self, n_in, n_hidden, n_out, det_activations,
stoch_activations, stoch_n_hidden):
self.n_hidden = np.array(n_hidden)
self.n_out = n_out
self.n_in = n_in
self.stoch_n_hidden = [np.array(i) for i in stoch_n_hidden]
self.det_activation_names = det_activations
self.det_activation, self.det_activation_prime = parse_activations(
det_activations)
self.stoch_activation_names = stoch_activations
self.stoch_activation, self.stoch_activation_prime = parse_activations(
stoch_activations)
def __init__(self, n_in, n_hidden, n_out, activation, bias_init, prng):
"""Defines the basics of a Vanilla Recurrent Neural Network.
:param n_in: integer defining the number of input units.
:param n_hidden: list of integers defining the number of hidden units per layer.
:param activation: list of size len(n_hidden) + 1 defining the activation function per layer.
:param bias_init: list with bias initialization for [layers biases] + [output bias].
:param prng: random number generator.
"""
self.prng = prng
self.n_in = n_in
assert type(
self.n_in) is IntType, "n_in must be an integer: {0!r}".format(
self.n_in)
assert type(
n_hidden) is ListType, "n_hidden must be a list: {0!r}".format(
n_hidden)
self.n_hidden = np.array(n_hidden)
self.activation_list = activation
assert type(
self.activation_list) is ListType, "activation must be a list:\
{0!r}".format(
self.activation_list)
assert len(self.n_hidden) + 1 == len(self.activation_list),\
"Activation list must have len(n_hidden) + 1 values. Activation: {0!r}, n_hidden: \
{1!r}" .format(self.activation_list, self.n_hidden)
self.bias_init = bias_init
assert type(self.bias_init
) is ListType, "biases initilizaition must be a list: \
{0!r}".format(
self.bias_init)
assert len(self.bias_init) == len(n_hidden) + 1,\
"Bias initialization list must have len(n_hidden) + 1 values. Bias list: {0!r}, n_hidden: \
{1!r}" .format(self.bias_init, self.n_hidden)
self.activation = parse_activations(self.activation_list)
self.type = 'VanillaRNN'
self.opt = {
'type': self.type,
'n_in': self.n_in,
'n_hidden': self.n_hidden,
'activation': self.activation_list,
'bias_init': self.bias_init
}
self.defined = False
logging.info(
'RNN loaded. Type: {0}, input layer: {1}, hidden layers: {2} '
'activation: {3}'.format(self.type, self.n_in, self.n_hidden,
self.activation_list))
self.n_out = n_out
self.initialize_weights()
self.generative_network()
def __init__(self, n_in, n_hidden, n_out, activation, bias_init, prng):
self.prng = prng
self.n_in = n_in
assert type(
self.n_in) is IntType, "n_in must be an integer: {0!r}".format(
self.n_in)
assert type(
n_hidden) is ListType, "n_hidden must be a list: {0!r}".format(
n_hidden)
self.n_hidden = np.array(n_hidden)
self.activation_list = activation
assert type(
self.activation_list) is ListType, "activation must be a list:\
{0!r}".format(
self.activation_list)
assert len(self.n_hidden) + 1 == len(self.activation_list),\
"Activation list must have len(n_hidden) + 1 values. Activation: {0!r}, n_hidden: \
{1!r}" .format(self.activation_list, self.n_hidden)
self.bias_init = bias_init
assert type(self.bias_init
) is ListType, "biases initilizaition must be a list: \
{0!r}".format(
self.bias_init)
assert len(self.bias_init) == len(n_hidden) + 1,\
"Bias initialization list must have len(n_hidden) + 1 values. Bias list: {0!r}, n_hidden: \
{1!r}" .format(self.bias_init, self.n_hidden)
self.activation = parse_activations(self.activation_list)
self.n_out = n_out
assert type(
self.n_out) is IntType, "n_out must be an int: {0!r}".format(
self.n_out)
self.forward_rnn = HiddenRNN(n_in, n_hidden, activation[:-1],
bias_init[:-1], prng)
self.backward_rnn = HiddenRNN(n_in, n_hidden, activation[:-1],
bias_init[:-1], prng)
self.type = 'BidirectionalRNN'
self.opt = {
'type': self.type,
'n_in': self.n_in,
'n_hidden': self.n_hidden,
'n_out': self.n_out,
'activation': self.activation_list,
'bias_init': self.bias_init
}
self.initialize_weights()
self.complete_defined = False
self.define_complete_network()
def __init__(self, n_in, n_hidden, n_out, activation, bias_init, prng):
"""Defines the basics of a Vanilla Recurrent Neural Network.
:param n_in: integer defining the number of input units.
:param n_hidden: list of integers defining the number of hidden units per layer.
:param activation: list of size len(n_hidden) + 1 defining the activation function per layer.
:param bias_init: list with bias initialization for [layers biases] + [output bias].
:param prng: random number generator.
"""
self.prng = prng
self.n_in = n_in
assert type(self.n_in) is IntType, "n_in must be an integer: {0!r}".format(self.n_in)
assert type(n_hidden) is ListType, "n_hidden must be a list: {0!r}".format(n_hidden)
self.n_hidden = np.array(n_hidden)
self.activation_list = activation
assert type(self.activation_list) is ListType, "activation must be a list:\
{0!r}".format(self.activation_list)
assert len(self.n_hidden) + 1 == len(self.activation_list),\
"Activation list must have len(n_hidden) + 1 values. Activation: {0!r}, n_hidden: \
{1!r}".format(self.activation_list, self.n_hidden)
self.bias_init = bias_init
assert type(self.bias_init) is ListType, "biases initilizaition must be a list: \
{0!r}".format(self.bias_init)
assert len(self.bias_init) == len(n_hidden) + 1,\
"Bias initialization list must have len(n_hidden) + 1 values. Bias list: {0!r}, n_hidden: \
{1!r}".format(self.bias_init, self.n_hidden)
self.activation = parse_activations(self.activation_list)
self.type = 'VanillaRNN'
self.opt = {'type': self.type, 'n_in': self.n_in, 'n_hidden': self.n_hidden,
'activation': self.activation_list,
'bias_init': self.bias_init}
self.defined = False
logging.info('RNN loaded. Type: {0}, input layer: {1}, hidden layers: {2} '
'activation: {3}'.format(self.type, self.n_in, self.n_hidden, self.activation_list))
self.n_out = n_out
self.initialize_weights()
self.generative_network()
def __init__(self, n_in, n_hidden, n_out, activation, bias_init, prng):
self.prng = prng
self.n_in = n_in
assert type(self.n_in) is IntType, "n_in must be an integer: {0!r}".format(self.n_in)
assert type(n_hidden) is ListType, "n_hidden must be a list: {0!r}".format(n_hidden)
self.n_hidden = np.array(n_hidden)
self.activation_list = activation
assert type(self.activation_list) is ListType, "activation must be a list:\
{0!r}".format(self.activation_list)
assert len(self.n_hidden) + 1 == len(self.activation_list),\
"Activation list must have len(n_hidden) + 1 values. Activation: {0!r}, n_hidden: \
{1!r}".format(self.activation_list, self.n_hidden)
self.bias_init = bias_init
assert type(self.bias_init) is ListType, "biases initilizaition must be a list: \
{0!r}".format(self.bias_init)
assert len(self.bias_init) == len(n_hidden) + 1,\
"Bias initialization list must have len(n_hidden) + 1 values. Bias list: {0!r}, n_hidden: \
{1!r}".format(self.bias_init, self.n_hidden)
self.activation = parse_activations(self.activation_list)
self.n_out = n_out
assert type(self.n_out) is IntType, "n_out must be an int: {0!r}".format(self.n_out)
self.forward_rnn = HiddenRNN(n_in, n_hidden, activation[:-1], bias_init[:-1], prng)
self.backward_rnn = HiddenRNN(n_in, n_hidden, activation[:-1], bias_init[:-1], prng)
self.type = 'BidirectionalRNN'
self.opt = {'type': self.type, 'n_in': self.n_in, 'n_hidden': self.n_hidden,
'n_out': self.n_out, 'activation': self.activation_list,
'bias_init': self.bias_init}
self.initialize_weights()
self.complete_defined = False
self.define_complete_network()
