def __init__(self,
name,
x,
attn1,
attn2,
mask,
x_in,
attn1_in,
attn2_in,
n_out,
path=None,
init_func=normal_weight):
SuperLayer.__init__(self, name, path)
if path is None:
self.params['Wx'] = to_shared(init_func((x_in, n_out)),
self.name + '_Wx')
self.params['Wv'] = to_shared(init_func((n_out, )),
self.name + '_Wv')
self.params['Wa_1'] = to_shared(init_func((attn1_in, n_out)),
self.name + '_Wa_1')
self.params['Wa_2'] = to_shared(init_func((attn2_in, n_out)),
self.name + '_Wa_2')
self.params['b'] = to_shared(init_bias((n_out, )),
self.name + '_b')
self.output = self.stream(x, attn1, attn2, mask)
def __init__(self, name, x, n_in, n_out, path=None, activation=T.tanh, init_func=normal_weight):
SuperLayer.__init__(self, name, path)
self.activation = activation
if path is None:
self.params['W'] = to_shared(init_func((n_in, n_out)), self.name + '_W')
self.params['b'] = to_shared(init_bias((n_out,)), self.name + '_b')
self.output = self.stream(x)
def __init__(self,
name,
x,
n_voc,
ndim,
init_value=None,
path=None,
init_func=normal_weight):
SuperLayer.__init__(self, name, path)
if path is None:
if init_value == None:
self.params['E'] = to_shared(init_func((n_voc, ndim)),
self.name + '_E')
else:
self.params['E'] = to_shared(init_value, self.name + '_E')
self.output = self.stream(x)
def __init__(self,
name,
image,
filter_shape,
image_shape,
path=None,
init_func=normal_weight):
SuperLayer.__init__(self, name, path)
self.filter_shape = filter_shape
self.image_shape = image_shape
if path is None:
self.params['W'] = to_shared(init_func(filter_shape),
self.name + '_W')
self.params['b'] = to_shared(init_bias((filter_shape[0], )),
self.name + '_b')
self.output = self.stream(image)
def __init__(self,
name,
x,
mask,
n_in,
hidden_size,
n_out,
path=None,
init_func=normal_weight):
SuperLayer.__init__(self, name, path)
if path is None:
self.params['Wi'] = to_shared(init_func((n_in, hidden_size)),
self.name + '_Wi')
self.params['Wf'] = to_shared(init_func((n_in, hidden_size)),
self.name + '_Wf')
self.params['Wo'] = to_shared(init_func((n_in, hidden_size)),
self.name + '_Wo')
self.params['Wc'] = to_shared(init_func((n_in, hidden_size)),
self.name + '_Wc')
self.params['Ui'] = to_shared(
init_func((hidden_size, hidden_size)), self.name + '_Ui')
self.params['Uf'] = to_shared(
init_func((hidden_size, hidden_size)), self.name + '_Uf')
self.params['Uo'] = to_shared(
init_func((hidden_size, hidden_size)), self.name + '_Uo')
self.params['Uc'] = to_shared(
init_func((hidden_size, hidden_size)), self.name + '_Uc')
self.params['bi'] = to_shared(init_bias((n_out, )),
self.name + '_bi')
self.params['bf'] = to_shared(init_bias((n_out, )),
self.name + '_bf')
self.params['bo'] = to_shared(init_bias((n_out, )),
self.name + '_bo')
self.params['bc'] = to_shared(init_bias((n_out, )),
self.name + '_bc')
self.output = self.stream(x, mask)
def __init__(self, name, x, rate, is_train, path=None):
SuperLayer.__init__(self, name, None)
self.output = self.stream(x, rate, is_train)
def __init__(self, name, x, pooling_size, path=None):
SuperLayer.__init__(self, name, path)
self.pooling_size = pooling_size
self.output = self.stream(x)