def __init__(self,
n_in,
n_hidden,
n_out,
activation=T.tanh,
inner_activation=T.nnet.sigmoid,
output_type='real',
batch_size=200):
self.activation = activation
self.inner_activation = inner_activation
self.output_type = output_type
self.batch_size = batch_size
self.n_hidden = n_hidden
self.W_i = shared_common(gloroat_uniform((n_in, n_hidden)))
self.U_i = shared_common(ortho_weight(n_hidden))
self.b_i = shared_zeros((n_hidden, ))
self.W_f = shared_common(gloroat_uniform((n_in, n_hidden)))
self.U_f = shared_common(ortho_weight(n_hidden))
self.b_f = shared_zeros((n_hidden, ))
self.W_c = shared_common(gloroat_uniform((n_in, n_hidden)))
self.U_c = shared_common(ortho_weight(n_hidden))
self.b_c = shared_zeros((n_hidden, ))
self.W_o = shared_common(gloroat_uniform((n_in, n_hidden)))
self.U_o = shared_common(ortho_weight(n_hidden))
self.b_o = shared_zeros((n_hidden, ))
self.params = [
self.W_i, self.U_i, self.b_i, self.W_c, self.U_c, self.b_c,
self.W_f, self.U_f, self.b_f, self.W_o, self.U_o, self.b_o
]
def __init__(self, rng, n_in, n_out, tensor_num=3, activation=T.tanh):
self.tensor_num = tensor_num
self.W = []
for i in range(tensor_num):
self.W.append(shared_common(ortho_weight(100)))
self.activation = activation
self.hidden_layer = HiddenLayer2(rng, tensor_num * 5 * n_in, n_out)
self.params = self.W + self.hidden_layer.params
def __init__(self,
rng,
n_in,
n_out,
W=None,
b=None,
activation=T.tanh,
hidden_size=100):
self.W = shared_common(ortho_weight(hidden_size), 'W')
self.activation = activation
self.hidden_layer = HiddenLayer2(rng, 2 * 5 * n_in, n_out)
self.params = [self.W] + self.hidden_layer.params
def __init__(self,
rng,
n_in,
n_out,
W=None,
b=None,
activation=T.tanh,
hidden_size=100):
self.W = shared_common(ortho_weight(hidden_size))
self.activation = activation
self.conv_layer = LeNetConvPoolLayer(rng,
filter_shape=(8, 1, 3, 3),
image_shape=(200, 1, 50, 50),
poolsize=(3, 3),
non_linear='relu')
self.hidden_layer = HiddenLayer2(rng, 2048, n_out)
self.params = self.conv_layer.params + self.hidden_layer.params
def __init__(self,
input_l,
input_r,
n_in,
n_hidden,
n_out,
activation=T.tanh,
output_type='real',
batch_size=200,
input_lm=None,
input_rm=None):
if input_lm is None:
input_lm = shared_ones((batch_size, 20))
if input_rm is None:
input_rm = shared_ones((batch_size, 20))
self.activation = activation
self.output_type = output_type
self.W = shared_common(ortho_weight(n_hidden), 'W')
self.W_in = shared_common(gloroat_uniform(n_in, n_hidden), 'W_in')
self.h0 = shared_zeros((batch_size, n_hidden), 'h0')
self.bh = shared_zeros((batch_size, n_hidden), 'bh')
self.params = [self.W, self.W_in, self.bh]
def step(x_t, mask, h_tm1):
h_tm1 = mask * h_tm1
h_t = T.tanh(
T.dot(x_t, self.W_in) + T.dot(h_tm1, self.W) + self.bh)
return h_t
self.h_l, _ = scan_dimshuffle(step, input_l, input_lm,
shared_zeros((batch_size, n_hidden)))
self.h_r, _ = scan_dimshuffle(step, input_r, input_rm,
shared_zeros((batch_size, n_hidden)))
self.h_l = self.h_l.dimshuffle(1, 0, 2)
self.h_r = self.h_r.dimshuffle(1, 0, 2)