def full(scaledparams, x, y, zstates):
params = hyperparameters.descale(scaledparams)
pi = params[0]
mu = params[1]
sigma = params[2]
sigmabg = params[3]
tau = params[4]
nvar = x.shape[0]
nsample = x.shape[1]
constpi = math.pi
marglik_k = 0
lmlzlist = list()
logk = 0
sigma2 = sigma * sigma
sigmabg2 = sigmabg * sigmabg
muz0 = np.zeros(nvar)
sigz0 = np.repeat(sigmabg2, nvar)
for z in zstates:
nz = len(z)
muz = muz0.copy()
muz[z] = mu
sigz = sigz0.copy()
sigz[z] = sigma2
# P(z | theta)
log_probz = nz * np.log(pi) + (nvar - nz) * np.log(1 - pi)
# N(y | Mz, Sz)
#Sz = np.einsum('ki, k, kj -> ij', x, sigz, x)
Sz = mat3mul(x.T, np.diag(sigz), x)
Sz[np.diag_indices_from(Sz)] += 1 / tau
Mz = np.einsum('ij, i -> j', x, muz)
logdetS = np.linalg.slogdet(Sz)[1]
invS = np.linalg.inv(Sz)
y_minus_M = y - Mz
nterm = np.einsum('i, ij, j', y_minus_M, invS, y_minus_M)
log_normz = -0.5 * (logdetS + (nsample * np.log(2 * constpi)) + nterm)
# Combine the log values
lmlzlist.append(log_probz + log_normz)
lmlzarr = np.array(lmlzlist)
logk = np.max(lmlzarr)
marglik_k = np.sum(np.exp(lmlzarr - logk))
logmarglik = logk + np.log(marglik_k)
return lmlzlist, -logmarglik
def raw_comps(scaledparams, x, y, zstates):
params = hyperparameters.descale(scaledparams)
pi = params[0]
mu = params[1]
sigma = params[2]
sigmabg = params[3]
tau = params[4]
logmarglik, der, pz, l = iterative_inverse(pi,
mu,
sigma,
sigmabg,
tau,
x,
y,
zstates,
grad=False)
return l
def func_grad(scaledparams, x, y, zstates):
params = hyperparameters.descale(scaledparams)
pi = params[0]
mu = params[1]
sigma = params[2]
sigmabg = params[3]
tau = params[4]
logmarglik, der, pz, l = iterative_inverse(pi, mu, sigma, sigmabg, tau, x,
y, zstates)
der = hyperparameters.gradscale(params, der)
#delta = 0.0001
#newparams = scaledparams.copy()
#newparams[4] += delta
#newtau = hyperparameters.descale(newparams)[4]
#newm, dummy1, dummy2 = iterative_inverse(pi, mu, sigma, sigmabg, newtau, x, y, zstates)
#tau_grad = (newm - logmarglik) / delta
#sigmatau = np.sqrt(1 / tau)
#print ("Params here are: {:g} {:g} {:g} {:g} {:g}".format(pi, mu, sigma, sigmabg, sigmatau))
#print ("Derivative of tau by brute force is {:f}".format(tau_grad))
#print ("Derivative of tau from equation is {:f}".format(der[4]))
return -logmarglik, -der
def model_exp(scaledparams, x, y, zstates):
params = hyperparameters.descale(scaledparams)
success, logmarglik, zprob, grad, zexp = czcompgrad(params, x, y, zstates, get_grad=False, get_exp=True)
zexp = zexp.reshape(len(zstates), x.shape[0])
return zprob, zexp
def prob_comps(scaledparams, x, y, zstates):
params = hyperparameters.descale(scaledparams)
success, logmarglik, zprob, grad, zexp = czcompgrad(params, x, y, zstates, get_grad=False)
return zprob
def func_grad(scaledparams, x, y, zstates):
params = hyperparameters.descale(scaledparams)
success, logmarglik, zprob, grad, zexp = czcompgrad(params, x, y, zstates, get_grad=True)
grad = hyperparameters.gradscale(params, grad)
return success, -logmarglik, -grad
def params(self):
return hyperparameters.descale(self._params)