class Model2(Model):
def __init__(self, N):
super(Model2, self).__init__(N=N)
self.p_B = Marginal(N)
self.p_A_B = Conditional(N)
def forward(self, inputs):
# Compute the (negative) log-likelihood with the
# decomposition p(x_A, x_B) = p(x_B)p(x_A | x_B)
inputs_A, inputs_B = torch.split(inputs, 1, dim=1)
inputs_A, inputs_B = inputs_A.squeeze(1), inputs_B.squeeze(1)
return self.p_B(inputs_B) + self.p_A_B(inputs_B, inputs_A)
def set_analytical_maximum_likelihood(self, pi_A, pi_B_A):
pi_A_th = torch.from_numpy(pi_A)
pi_B_A_th = torch.from_numpy(pi_B_A)
log_joint = torch.log(pi_A_th.unsqueeze(1)) + torch.log(pi_B_A_th)
log_p_B = torch.logsumexp(log_joint, dim=0)
self.p_B.w.data = log_p_B
self.p_A_B.w.data = log_joint.t() - log_p_B.unsqueeze(1)
def init_parameters(self):
self.p_B.init_parameters()
self.p_A_B.init_parameters()
class Model1(Model):
def __init__(self, N):
super(Model1, self).__init__(N=N)
self.p_A = Marginal(N)
self.p_B_A = Conditional(N)
def forward(self, inputs):
# Compute the (negative) log-likelihood with the
# decomposition p(x_A, x_B) = p(x_A)p(x_B | x_A)
inputs_A, inputs_B = torch.split(inputs, 1, dim=1)
inputs_A, inputs_B = inputs_A.squeeze(1), inputs_B.squeeze(1)
return self.p_A(inputs_A) + self.p_B_A(inputs_A, inputs_B)
def set_analytical_maximum_likelihood(self, pi_A, pi_B_A):
pi_A_th = torch.from_numpy(pi_A)
pi_B_A_th = torch.from_numpy(pi_B_A)
self.p_A.w.data = torch.log(pi_A_th)
self.p_B_A.w.data = torch.log(pi_B_A_th)
def init_parameters(self):
self.p_A.init_parameters()
self.p_B_A.init_parameters()
def __init__(self, N, dtype=None):
Model.__init__(self, N)
ModelB2A.__init__(self, Marginal(N, dtype=dtype),
Conditional(N, dtype=dtype))
def __init__(self, N):
super(Model2, self).__init__(N=N)
self.p_B = Marginal(N)
self.p_A_B = Conditional(N)
def __init__(self, N, dtype=None):
Model.__init__(self, N)
ModelA2B.__init__(self, MixtureMarginal(N, dtype=dtype), Conditional(N, dtype=dtype))