def logistic_logg(v, theta):
(logH, logmH, dlogH, dlogmH) = logistic_logh(np.asarray(range(1, v
+ 1)), theta)
logg = np.zeros((v, ))
dlogg = np.zeros_like(dlogH)
for ii in range(v):
if ii == 0:
logg[ii] = logH[ii]
dlogg[ii, :] = dlogH[ii, :]
else:
logg[ii] = logmH[:ii].sum() + logH[ii]
dlogg[ii, :] = dlogmH[:ii, :].sum(axis=0) + dlogH[ii, :]
# exp(logmG) = 1 - G = 1 - cumsum(g) = 1 - cumsum(exp(logg)), but this is a much
# more numerically stable way to do it that won't underflow for G close to 1.
logmG = np.cumsum(logmH)
dlogmG = np.cumsum(dlogmH, axis=0)
return (logg, logmG, dlogg, dlogmG)
def evaluate(self,v):
hazard = logistic_logh(self.hazard_params)
(logH, logmH, dlogH, dlogmH) = hazard.evaluate(asarray(range(1,v+1)))
logg = zeros((v, ))
dlogg = zeros_like(dlogH)
for ii in range(v):
if ii == 0:
logg[ii] = logH[ii]
dlogg[ii,:] = dlogH[ii,:]
else:
logg[ii] = logmH[:ii].sum() + logH[ii]
dlogg[ii,:] = dlogmH[:ii,:].sum(axis=0) + dlogH[ii,:]
# exp(logmG) = 1 - G = 1 - cumsum(g) = 1 - cumsum(exp(logg)), but this is a much
# more numerically stable way to do it that won't underflow for G close to 1.
logmG = cumsum(logmH)
dlogmG = cumsum(dlogmH, axis=0)
return (logg, logmG, dlogg, dlogmG)
def logistic_logg(v, theta):
(logH, logmH, dlogH, dlogmH) = logistic_logh(np.asarray(range(1, v + 1)),
theta)
logg = np.zeros((v, ))
dlogg = np.zeros_like(dlogH)
for ii in range(v):
if ii == 0:
logg[ii] = logH[ii]
dlogg[ii, :] = dlogH[ii, :]
else:
logg[ii] = logmH[:ii].sum() + logH[ii]
dlogg[ii, :] = dlogmH[:ii, :].sum(axis=0) + dlogH[ii, :]
# exp(logmG) = 1 - G = 1 - cumsum(g) = 1 - cumsum(exp(logg)), but this is a much
# more numerically stable way to do it that won't underflow for G close to 1.
logmG = np.cumsum(logmH)
dlogmG = np.cumsum(dlogmH, axis=0)
return (logg, logmG, dlogg, dlogmG)