def __init__(self,
rank,
filters,
kernel_size,
strides=1,
padding='valid',
dilation_rate=1,
kernel_factory="xavier_uniform",
bias_factory='zeros',
activation="linear",
kernel_regulariser=None,
bias_regulariser=None,
**kwargs):
self.rank = rank
self.filters = filters
self.kernel_size = normalize_tuple(kernel_size, rank, 'kernel_size')
self.strides = normalize_tuple(strides, rank, 'strides')
self.padding = normalize_padding(padding)
self.dilation_rate = normalize_tuple(dilation_rate, rank,
'dilation_rate')
self.activation = get_activation(activation)
self.kernel_factory = get_weightfactory(kernel_factory)
self.kernel_regulariser = get_regulariser(kernel_regulariser)
self.bias_factory = get_weightfactory(bias_factory)
self.bias_regulariser = get_regulariser(bias_regulariser)
super(Conv, self).__init__(**kwargs)
def __init__(self, n, activation="tanh", h_activation="hard_sigmoid",
weight_factory="xavier_uniform", h_factory="orthogonal",
w_regulariser=None, h_regulariser=None, b_regulariser=None,
w_dropout=0.,
h_dropout=0., **kwargs):
self.n = n
self.w_factory = get_weightfactory(weight_factory)
self.h_factory = get_weightfactory(h_factory)
self.activation = get_activation(activation)
self.h_activation = get_activation(h_activation)
self.w_regulariser = get_regulariser(w_regulariser)
self.b_regulariser = get_regulariser(b_regulariser)
self.h_regulariser = get_regulariser(h_regulariser)
self.dropout_w = min(1., max(0., w_dropout))
self.dropout_h = min(1., max(0., h_dropout))
self.state_spec = [InputDetail(shape=(None, self.n)),
InputDetail(shape=(None, self.n))]
super(GRULayer, self).__init__(**kwargs)
def __init__(self, n, weight_factory='xavier_uniform',
h_factory='orthogonal', activation='tanh',
w_regulariser=None, h_regulariser=None, b_regulariser=None,
w_dropout=0., h_dropout=0., **kwargs):
self.n = n
self.w_factory = get_weightfactory(weight_factory)
self.h_factory = get_weightfactory(h_factory)
self.activation = get_activation(activation)
self.w_regulariser = get_regulariser(w_regulariser)
self.b_regulariser = get_regulariser(b_regulariser)
self.h_regulariser = get_regulariser(h_regulariser)
self.dropout_w = min(1., max(0., w_dropout))
self.dropout_h = min(1., max(0., h_dropout))
self.state_spec = InputDetail(shape=(None, self.n))
Recurrent.__init__(self, **kwargs)
def __init__(self,
n,
weight_factory='xavier_uniform',
activation='linear',
weights=None,
w_regulariser=None,
b_regulariser=None,
input_shape=None,
**kwargs):
self.weightFactory = get_weightfactory(weight_factory)
self.activation = get_activation(activation)
self.w_regulariser = get_regulariser(w_regulariser)
self.b_regulariser = get_regulariser(b_regulariser)
if (input_shape is not None):
self.input_dim = input_shape[1]
else:
self.input_dim = None
self.n = n
if (input_shape is not None):
kwargs['input_shape'] = input_shape
Layer.__init__(self, **kwargs)
def __init__(self, activation, **kwargs):
self.activation = get_activation(activation)
Layer.__init__(self, **kwargs)