def __init__(self, x_dim, h_dim, t_dim):
super(VAE_BKDG, self).__init__()
self.x_dim = x_dim
self.h_dim = h_dim
self.t_dim = t_dim
d_dim = int(x_dim/t_dim)
self.d_dim = d_dim
l_dim = int(d_dim * (d_dim+1)/2)
self.l_dim = l_dim
z_dim = t_dim * l_dim
self.z_dim = z_dim
# feature
self.fc0 = nn.Linear(x_dim, h_dim)
self.fc1 = nn.Linear(h_dim, h_dim)
# encode
self.fc21 = nn.Linear(h_dim, z_dim)
self.fc22 = nn.Linear(h_dim, int(t_dim*(t_dim+1)/2))
# transform
self.fc2 = nn.Linear(z_dim, h_dim)
self.fc3 = nn.Linear(h_dim, h_dim)
# decode
self.fc41 = nn.Linear(h_dim, x_dim)
self.fc42 = nn.Linear(h_dim, x_dim)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
self.tanh = nn.Tanh()
t = torch.linspace(0,2,steps=t_dim+1); t = t[1:]
self.K = Variable(torch.exp(-torch.pow(t.unsqueeze(1)-t.unsqueeze(0),2)/2/2) + 1e-4*torch.eye(t_dim))
self.Kh = torch.potrf(self.K)
self.iK = torch.potri(self.Kh)
def __init__(self, x_dim, h_dim, z_dim):
super(VAE, self).__init__()
self.x_dim = x_dim # ND
self.h_dim = h_dim
self.z_dim = z_dim # ND^*
self.t_dim = np.int(x_dim / (2 * z_dim / x_dim - 1)) # N
# feature
self.fc0 = nn.Linear(x_dim, h_dim)
# encode
self.fc21 = nn.Linear(h_dim, z_dim)
self.fc22 = nn.Linear(h_dim, np.int(self.t_dim * (self.t_dim + 1) / 2))
# self.fc23 = nn.Linear(h_dim, z_dim)
# transform
self.fc3 = nn.Linear(z_dim, h_dim)
# decode
self.fc41 = nn.Linear(h_dim, x_dim)
self.fc42 = nn.Linear(h_dim, x_dim)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
self.tanh = nn.Tanh()
# problem-specific parameters
self.D = np.int(self.z_dim / self.x_dim * 2 - 1)
self.N = np.int(self.x_dim / self.D)
# GP kernel
t = torch.linspace(0, 2, steps=self.N + 1)
t = t[1:]
self.K = Variable(
torch.exp(-torch.pow(t.unsqueeze(1) - t.unsqueeze(0), 2) / 2 / 2) +
1e-4 * torch.eye(self.N))
self.Kh = torch.potrf(self.K)
# self.iK = Variable(torch.inverse(self.K.data))
self.iK = torch.potri(self.Kh)
def __init__(self, x_dim, h_dim, t_dim):
super(VGP, self).__init__()
self.x_dim = x_dim
self.h_dim = h_dim
self.t_dim = t_dim
d_dim = int(x_dim / t_dim)
self.d_dim = d_dim
d2h_dim = int(d_dim * (d_dim + 1) / 2)
self.d2h_dim = d2h_dim
z_dim = t_dim * d2h_dim
self.z_dim = z_dim
# source-target dimensions of GP draws
f_in = 20
self.f_in = f_in
f_out = 100
self.f_out = f_out
# encode 1: x -> s,t (variational data)
self.fc1 = nn.Linear(x_dim, h_dim)
self.fc11 = nn.Linear(h_dim, f_in)
self.fc12 = nn.Linear(h_dim, f_out)
# encode 2: f -> z
self.fc2 = nn.Linear(f_out, h_dim)
self.fc21 = nn.Linear(h_dim, z_dim)
self.fc22 = nn.Linear(h_dim, int(t_dim * (t_dim + 1) / 2))
# encode 3: x, z -> r
self.fc3 = nn.Linear(x_dim + z_dim, h_dim)
self.fc31 = nn.Linear(h_dim, f_in + f_out)
self.fc32 = nn.Linear(h_dim, f_in + f_out)
# decode: z -> x
self.fc4 = nn.Linear(z_dim, h_dim)
self.fc41 = nn.Linear(h_dim, x_dim)
self.fc42 = nn.Linear(h_dim, x_dim)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
self.tanh = nn.Tanh()
# GP kernel
t = torch.linspace(0, 2, steps=t_dim + 1)
t = t[1:]
self.K = Variable(
torch.exp(-torch.pow(t.unsqueeze(1) - t.unsqueeze(0), 2) / 2 / 2) +
1e-4 * torch.eye(t_dim))
self.Kh = torch.potrf(self.K)
# self.iK = Variable(torch.inverse(self.K.data))
self.iK = torch.potri(self.Kh)
def estep(self, data):
"""
Compute log observation probabilities under GMM.
Args:
data: N x D tensor of points
Returns:
logobs: N x K tensor of log probabilities (for each point on each
cluster)
"""
# Constants
N, D = data.shape
K = self._sigma.shape[0]
logobs = -0.5 * ptu.ones((N, K)) * D * np.log(2*np.pi) # Constant
for i in range(K):
mu, sigma = self._mu[i], self._sigma[i]
L = torch.potri(sigma, upper=False) # Cholesky decomposition
logobs[:, i] -= torch.sum(torch.log(torch.))
def potri_compat(var):
return Variable(torch.potri(
var.data)) if torch.__version__ < '0.3.0' else torch.potri(var)
def psd_inv(mat): u = T.cholesky(mat, upper=False) return T.potri(u, upper=False)
def _get_precision(self):
if self._l is None:
self._get_chol()
self._precision = Variable(torch.potri(self._l.data))