def v_to_h(self, V, do_sampling=True):
""" Compute the data representation in the hidden space """
m = V.shape[0]
if len(V.shape) == 1: V.shape = (1, m)
# print V.shape, self.W.shape, self.hbias.shape
H = Activation.sigmoid_function( np.dot(np.c_[np.ones(m), V], np.vstack((self.hbias, self.W)) ) )
if do_sampling:
H = self.sample(H)
return H
def h_to_v(self, H, do_sampling=False, pois_N=None):
""" Restore visual data using hidden space representation.
Returns the data representation in the original space.
"""
m = H.shape[0]
pre_distr = np.dot( np.c_[np.ones(m), H], np.vstack((self.vbias, self.W.T)) )
if self.mode == 'gaus-bin':
return pre_distr
elif self.mode == 'pois-bin':
e = np.exp(pre_distr)
if pois_N is None:
return ( e.T / np.sum(e, axis=1) ).T
else:
return ( e.T / ((1. / pois_N) * np.sum(e, axis=1)) ).T
V = Activation.sigmoid_function(pre_distr)
if do_sampling:
V = self.sample(V)
return V