def __init__(self,units,activation='tanh',initializer='glorotuniform',recurrent_initializer='orthogonal',**kwargs):
super(SimpleRNNCell, self).__init__(**kwargs)
self.units = units
self.activator_cls = get_activator(activation).__class__
self.initializer = get_initializer(initializer)
self.recurrent_initializer = get_initializer(recurrent_initializer)
self.__first_initialize=True
def __init__(self,epsilon=1e-3,momentum=0.9,axis=-1,gamma_initializer='ones',beta_initializer='zeros',moving_mean_initializer='zeros', moving_variance_initializer='ones'):
# axis=-1 when input Fully Connected Layers(data shape:(M,N),where M donotes Batch-size,and N represents feature nums)
# axis=1 when input Convolution Layers(data shape:(M,C,H,W),represents Batch-size,Channels,Height,Width,respectively)
self.epsilon=epsilon
self.axis=axis
self.momentum=momentum
self.gamma_initializer=get_initializer(gamma_initializer)
self.beta_initializer=get_initializer(beta_initializer)
self.moving_mean_initializer=get_initializer(moving_mean_initializer)
self.moving_variance_initializer=get_initializer(moving_variance_initializer)
super(BatchNormalization,self).__init__()
def __init__(self,
n_out,
n_in=None,
initializer='Normal',
activation='linear',
kernel_regularizer=None):
self.n_out = n_out
self.n_in = n_in
self.initializer = get_initializer(initializer)
self.activator = get_activator(activation)
self.kernel_regularizer = get_regularizer(kernel_regularizer)
super(Dense, self).__init__()
def __init__(self,
input_dim,
output_dim,
embeddings_initializer='uniform',
mask_zero=False,
input_length=None,
**kwargs):
super(Embedding, self).__init__(**kwargs)
self.input_dim = input_dim
self.output_dim = output_dim
self.initializer = get_initializer(embeddings_initializer)
self.mask_zero = mask_zero
self.input_length = input_length
def __init__(self,
filter_nums,
filter_size,
input_shape=None,
stride=1,
padding='VALID',
activation='linear',
initializer='Normal'):
self.filter_nums = filter_nums
self.filter_size = filter_size
self.input_shape = input_shape
self.stride = stride
self.padding = padding
self.activator = get_activator(activation)
self.initializer = get_initializer(initializer)
super(Conv2D, self).__init__(input_shape=input_shape)