def compute_forward_vec(self, seq_potential):
""" compute alpha matrix
Args:
seq_potential: computed potential at every time step using :func:`self.compute_potential`.
It has the shape (seqlen,tagset_size,tagset_size)
"""
# print("we are in compute_forward_vec")
y_codebook = self.y_codebook
T = seq_potential.size(0)
offset = 1
feat_score = seq_potential
if (self.transmatrix_flag):
feat_score = seq_potential + self.y_transparams
if (
self.stopstate_symb
): # when _STOP_ state is defined a transition matrix is defined too
offset = 2
num_rows = T + offset
# create alpha matrix
alpha = torch.Tensor(num_rows, self.tagset_size).fill_(
self.neginf).type(self.fdtype)
alpha[0, y_codebook[self.startstate_symb]] = 0.
alpha = autograd.Variable(alpha)
# print("alpha matrix: \n", alpha)
for t in range(T):
score = alpha[t, :].view(-1, 1) + feat_score[t, :, :]
log_score = logsumexp_var(score, dim=0)
alpha[t + 1, :] = log_score
if (self.stopstate_symb):
t = T
score = alpha[t, :].view(-1, 1) + self.y_transparams
log_score = logsumexp_var(score, dim=0)
alpha[t + 1, :] = log_score
Z = alpha[t + 1,
y_codebook[self.stopstate_symb]] # partition function
else:
t = T
Z = logsumexp_var(alpha[t, :].view(-1, 1),
dim=0)[0] # compute partition function Z
# print("alpha \n", alpha)
# print("Z: ", Z)
return (Z, alpha)
def compute_backward_vec(self, seq_potential):
""" compute beta matrix
Args:
seq_potential: computed potential at every time step using :func:`self.compute_potential`.
It has the shape (seqlen,tagset_size,tagset_size)
"""
# print("we are in compute_backward_vec")
y_codebook = self.y_codebook
T = seq_potential.size(0)
offset = 1
feat_score = seq_potential
if (self.transmatrix_flag):
feat_score = seq_potential + self.y_transparams
if (self.stopstate_symb):
offset = 2
num_rows = T + offset
# create beta matrix
beta = torch.Tensor(num_rows, self.tagset_size).fill_(
self.neginf).type(self.fdtype)
if (self.stopstate_symb):
beta[-1, y_codebook[self.stopstate_symb]] = 0
beta = autograd.Variable(beta)
t = T
score = beta[t + 1, :].view(1, -1) + self.y_transparams
log_score = logsumexp_var(score, dim=1)
beta[t, :] = log_score
else:
beta[-1, :] = 0
beta = autograd.Variable(beta)
# print("beta matrix: \n", beta)
for t in reversed(range(0, T)):
score = beta[t + 1, :].view(1, -1) + feat_score[t, :, :]
log_score = logsumexp_var(score, dim=1)
beta[t, :] = log_score
Z = beta[0, y_codebook[self.startstate_symb]]
# print("beta \n", beta)
# print("Z: ", Z)
return (Z, beta)