def __init__(self, **hyper_params):
super(NADE, self).__init__()
self.register_hyper_param('n_X', help='no. observed binary variables')
self.register_hyper_param('n_Y',
help='no. conditioning binary variables')
self.register_hyper_param('n_hid', help='no. latent binary variables')
self.register_hyper_param('unroll_scan', default=1)
self.register_model_param('b',
help='visible bias',
default=lambda: np.zeros(self.n_X))
self.register_model_param('c',
help='hidden bias',
default=lambda: np.zeros(self.n_hid))
self.register_model_param(
'Ub',
help='cond. weights Ub',
default=lambda: default_weights(self.n_Y, self.n_X))
self.register_model_param(
'Uc',
help='cond. weights Uc',
default=lambda: default_weights(self.n_Y, self.n_hid))
self.register_model_param(
'W',
help='encoder weights',
default=lambda: default_weights(self.n_X, self.n_hid))
self.register_model_param(
'V',
help='decoder weights',
default=lambda: default_weights(self.n_hid, self.n_X))
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(DSBN, self).__init__()
self.register_hyper_param('n_X',
help='no. lower-layer binary variables')
self.register_hyper_param('n_Y',
help='no. upper-layer binary variables')
self.register_hyper_param('n_D', help='no. deterministic units')
self.register_hyper_param('non_lin',
default='sigmoid',
help="nonlinearity for deterministic layer")
# Sigmoid Belief Layer
self.register_model_param('a',
help='deterministic bias',
default=lambda: -np.ones(self.n_D))
self.register_model_param(
'U',
help='deterministic weights',
default=lambda: default_weights(self.n_Y, self.n_D))
self.register_model_param('b',
help='stochastic bias',
default=lambda: -np.ones(self.n_X))
self.register_model_param(
'W',
help='stochastic weights',
default=lambda: default_weights(self.n_D, self.n_X))
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(NADETop, self).__init__()
self.register_hyper_param('n_X', help='no. observed binary variables')
self.register_hyper_param('n_hid', help='no. latent binary variables')
self.register_hyper_param('unroll_scan', default=1)
self.register_model_param('b', help='visible bias', default=lambda: np.zeros(self.n_X))
self.register_model_param('c', help='hidden bias' , default=lambda: np.zeros(self.n_hid))
self.register_model_param('W', help='encoder weights', default=lambda: default_weights(self.n_X, self.n_hid) )
self.register_model_param('V', help='decoder weights', default=lambda: default_weights(self.n_hid, self.n_X) )
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(DARN, self).__init__()
# Hyper parameters
self.register_hyper_param('n_X', help='no. binary variables')
self.register_hyper_param('n_Y', help='no. conditioning binary variables')
self.register_hyper_param('unroll_scan', default=1)
# Model parameters
self.register_model_param('b', help='sigmoid(b)-bias ', default=lambda: np.zeros(self.n_X))
self.register_model_param('W', help='weights (triangular)', default=lambda: default_weights(self.n_X, self.n_X) )
self.register_model_param('U', help='cond. weights U', default=lambda: default_weights(self.n_Y, self.n_X) )
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(DiagonalGaussian, self).__init__()
self.register_hyper_param('n_X',
help='no. lower-layer binary variables')
self.register_hyper_param('n_Y',
help='no. upper-layer binary variables')
self.register_hyper_param('n_hid', help='no. hidden variables')
self.register_hyper_param('log_sigma2_min',
help='log(sigma**2) cutoff',
default=-np.inf)
self.register_hyper_param('final_tanh', help='', default=False)
self.register_hyper_param('initial_tanh', help='', default=False)
self.set_hyper_params(hyper_params)
# sanitize hyper params
if isinstance(self.n_hid, int):
self.n_hid = (self.n_hid, )
if isinstance(self.n_hid, list):
self.n_hid = tuple(self.n_hid)
self.n_layers = len(self.n_hid)
hidden_size = self.n_hid + (self.n_Y, )
for i in reversed(xrange(self.n_layers)):
W_name = "W%d" % i
b_name = "b%d" % i
n_upper = hidden_size[i + 1]
n_lower = hidden_size[i]
def create_W_init(n_upper, n_lower):
return lambda: default_weights(n_upper, n_lower)
def create_b_init(n_lower):
return lambda: np.zeros(n_lower)
self.register_model_param(W_name,
default=create_W_init(n_upper, n_lower))
self.register_model_param(b_name, default=create_b_init(n_lower))
# lowers layer parametrization... higher layers will be parameterized in self.setup()
self.register_model_param(
'W_mu', default=lambda: default_weights(hidden_size[0], self.n_X))
self.register_model_param(
'W_ls', default=lambda: default_weights(hidden_size[0], self.n_X))
self.register_model_param('b_mu', default=lambda: np.zeros(self.n_X))
self.register_model_param('b_ls', default=lambda: np.zeros(self.n_X))
def __init__(self, **hyper_params):
super(DSBN, self).__init__()
self.register_hyper_param('n_X', help='no. lower-layer binary variables')
self.register_hyper_param('n_Y', help='no. upper-layer binary variables')
self.register_hyper_param('n_D', help='no. deterministic units')
self.register_hyper_param('non_lin', default='sigmoid', help="nonlinearity for deterministic layer")
# Sigmoid Belief Layer
self.register_model_param('a', help='deterministic bias', default=lambda: -np.ones(self.n_D))
self.register_model_param('U', help='deterministic weights', default=lambda: default_weights(self.n_Y, self.n_D) )
self.register_model_param('b', help='stochastic bias', default=lambda: -np.ones(self.n_X))
self.register_model_param('W', help='stochastic weights', default=lambda: default_weights(self.n_D, self.n_X) )
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(SBN, self).__init__()
self.register_hyper_param('n_X', help='no. lower-layer binary variables')
self.register_hyper_param('n_Y', help='no. upper-layer binary variables')
# Sigmoid Belief Layer
self.register_model_param('b', help='P lower-layer bias', default=lambda: -np.ones(self.n_X))
self.register_model_param('W', help='P weights', default=lambda: default_weights(self.n_Y, self.n_X) )
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(DARN, self).__init__()
# Hyper parameters
self.register_hyper_param('n_X', help='no. binary variables')
self.register_hyper_param('n_Y',
help='no. conditioning binary variables')
self.register_hyper_param('unroll_scan', default=1)
# Model parameters
self.register_model_param('b',
help='sigmoid(b)-bias ',
default=lambda: np.zeros(self.n_X))
self.register_model_param(
'W',
help='weights (triangular)',
default=lambda: default_weights(self.n_X, self.n_X))
self.register_model_param(
'U',
help='cond. weights U',
default=lambda: default_weights(self.n_Y, self.n_X))
self.set_hyper_params(hyper_params)
def __init__(self, **hyper_params):
super(GibbsSBN, self).__init__()
# Hyper parameters
self.register_hyper_param('n_X',
help='no. lower-layer binary variables')
self.register_hyper_param('n_Y',
help='no. upper-layer binary variables')
self.register_hyper_param('unroll_scan', default=1)
# Model parameters
self.register_model_param('b',
help='P lower-layer bias',
default=lambda: -np.ones(self.n_X))
self.register_model_param(
'W',
help='P weights',
default=lambda: default_weights(self.n_Y, self.n_X))
self.set_hyper_params(hyper_params)
def create_W_init(n_upper, n_lower):
return lambda: default_weights(n_upper, n_lower)
