def __init__(self, threshold, nb_inp, nb_hid, layer_id='', W=None):
base.__init__(self, layer_id, W=W)
# Hyperparameters
self.threshold = threshold
self.nb_inp = nb_inp
self.nb_hid = nb_hid
# Set optimizable parameter dictionary
init = uniform(low=-1, high=1,
size=(nb_inp, nb_hid)).astype(th.config.floatX)
init = init / np.sqrt((init**2).sum(0))[None, :]
self.W = W if W else th.shared(init, name='W' + self.layer_id)
def __init__(self,
act,
nb_inp,
nb_hid,
init='gaussian',
normalize=True,
layer_id=None,
W=None,
b_enc=None,
b_opp=None):
base.__init__(self, layer_id, W=W, b_enc=b_enc, b_opp=b_opp)
# Hyperparameters
self.act = act
self.nb_inp = nb_inp
self.nb_hid = nb_hid
# Parameters
if not self.W:
if init == 'gaussian':
w = np.random.normal(0, 0.1, (nb_inp, nb_hid))
elif init == 'uniform':
w = np.random.uniform(-6 / sqrt(nb_inp + nb_hid),
6 / sqrt(nb_inp + nb_hid),
(nb_inp, nb_hid))
if normalize:
w /= np.sqrt((w**2).sum(0))[None, :]
self.W = shared_x(w, name='W' + self.layer_id)
self.scale = shared_x(1, name='scale' + self.layer_id)
if not self.b_enc:
self.b_enc = shared_x(np.zeros(nb_hid),
name='b_enc' + self.layer_id)
if not self.b_opp:
self.b_opp = shared_x(np.zeros(nb_hid),
name='b_opp' + self.layer_id)
# For debug
self.__init_W = shared_x(self.W.get_value())
def __init__(self,
act,
nb_inp,
nb_hid,
init='gaussian',
normalize='True',
layer_id=None,
scale=None,
W=None,
b=None,
b_dec=None):
base.__init__(self, layer_id, b=b, b_dec=b_dec, scale=scale)
# Hyperparameters
self.act = act
self.nb_inp = nb_inp
self.nb_hid = nb_hid
# Parameters
if not W:
if init == 'gaussian':
w = np.random.normal(0, 0.1, (nb_inp, nb_hid))
elif init == 'uniform':
w = np.random.uniform(-6 / sqrt(nb_inp + nb_hid),
6 / sqrt(nb_inp + nb_hid),
(nb_inp, nb_hid))
if normalize:
w /= np.sqrt((w**2).sum(0))[None, :]
self.W = shared_x(w, name='W' + self.layer_id)
if not self.scale:
self.scale = shared_x(1, name='scale' + self.layer_id)
if not self.b:
self.b = shared_x(np.zeros(nb_hid), name='b' + self.layer_id)
if not self.b_dec:
self.b_dec = shared_x(np.zeros(nb_inp),
name='b_dec' + self.layer_id)
def __init__(self,
dae_act,
gdae_act,
nb_inp,
nb_hid,
nb_fac,
nb_map,
shared,
divide,
layer_id='',
W=None,
B=None,
FACX=None,
FACY=None,
M_enc=None,
M_dec=None,
B_map=None,
B_dec=None):
base.__init__(self,
layer_id,
W=W,
B=B,
FACX=FACX,
FACY=FACY,
M_enc=M_enc,
B_map=B_map,
B_dec=B_dec)
# Hyperparameters
self.nb_inp = nb_inp
self.nb_hid = nb_hid
self.nb_fac = nb_fac
self.nb_map = nb_map
self.shared = shared
self.divide = divide
self.dae_act = dae_act
self.gdae_act = gdae_act
# Pre-calculation for conciseness
daef = 4. if dae_act == T.nnet.sigmoid else 1.
gdaef = 4. if gdae_act == T.nnet.sigmoid else 1.
if W == None and nb_hid > 0:
self.W = shared_x(uniform(-daef * sqrt(6. / nb_inp),
daef * sqrt(6. / nb_inp),
size=(nb_inp, nb_hid)),
name='W{}'.format(layer_id))
if B == None and nb_hid > 0:
self.B = shared_x(np.zeros(nb_hid), name='B{}'.format(layer_id))
if FACX is None and nb_fac > 0:
self.FAC_X = shared_x(uniform(-gdaef * sqrt(6. /
(nb_inp + nb_fac)),
gdaef * sqrt(6. / (nb_inp + nb_fac)),
size=(nb_inp, nb_fac)),
name='FACX{}'.format(layer_id))
if FACY is None and nb_fac > 0:
self.FAC_Y = self.FAC_X if shared else shared_x(
uniform(-gdaef * sqrt(6. / (nb_inp + nb_fac)),
gdaef * sqrt(6. / (nb_inp + nb_fac)),
size=(nb_inp, nb_fac)),
name='FACX{}'.format(layer_id))
if M_enc is None and nb_fac > 0 and nb_map > 0:
self.M_enc = shared_x(uniform(-gdaef * sqrt(6. /
(nb_fac + nb_map)),
gdaef * sqrt(6. / (nb_fac + nb_map)),
size=(nb_fac, nb_map)),
name='M_enc{}'.format(layer_id))
self.M_dec = shared_x(uniform(-gdaef * sqrt(6. /
(nb_fac + nb_map)),
gdaef * sqrt(6. / (nb_fac + nb_map)),
size=(nb_map, nb_fac)),
name='M_dec{}'.format(layer_id))
self.BMAP = th.shared(
np.zeros(nb_map, dtype=th.config.floatX),
borrow=True,
name='mapping_unit_biases{}'.format(layer_id))
self.BVIS = th.shared(np.zeros(nb_inp, dtype=th.config.floatX),
borrow=True,
name='decoder_unit_biases{}'.format(layer_id))