def compute_bethe_average_energy_MLP(self, factor_beliefs, factor_potentials, batch_factors, debug=False):
'''
Equation (37) in:
https://www.cs.princeton.edu/courses/archive/spring06/cos598C/papers/YedidaFreemanWeiss2004.pdf
'''
assert(factor_potentials.shape == factor_beliefs.shape), (factor_potentials.shape, factor_beliefs.shape)
if debug:
print()
print('!!!!!!!')
print("debugging compute_bethe_average_energy")
print("torch.exp(factor_beliefs):", torch.exp(factor_beliefs))
print("neg_inf_to_zero(factor_potentials):", neg_inf_to_zero(factor_potentials))
pooled_fac_beleifPotentials = global_add_pool(torch.exp(factor_beliefs)*neg_inf_to_zero(factor_potentials), batch_factors)
#keep 1st dimension for # of factors, but flatten remaining dimensions for belief_repeats and each factor
pooled_fac_beleifPotentials = pooled_fac_beleifPotentials.view(pooled_fac_beleifPotentials.shape[0], -1)
if debug:
factor_beliefs_shape = factor_beliefs.shape
pooled_fac_beleifPotentials_orig = torch.sum((torch.exp(factor_beliefs)*neg_inf_to_zero(factor_potentials)).view(factor_beliefs_shape[0], -1), dim=0)
print("original pooled_fac_beleifPotentials_orig:", pooled_fac_beleifPotentials_orig)
print("pooled_fac_beleifPotentials:", pooled_fac_beleifPotentials)
print("factor_beliefs.shape:", factor_beliefs.shape)
print("pooled_fac_beleifPotentials_orig.shape:", pooled_fac_beleifPotentials_orig.shape)
print("pooled_fac_beleifPotentials.shape:", pooled_fac_beleifPotentials.shape)
print("(torch.exp(factor_beliefs)*neg_inf_to_zero(factor_potentials)).view(factor_beliefs_shape[0], -1).shape:", (torch.exp(factor_beliefs)*neg_inf_to_zero(factor_potentials)).view(factor_beliefs_shape[0], -1).shape)
return pooled_fac_beleifPotentials #negate and sum to get average bethe energy
def compute_bethe_entropy_MLP(self, factor_beliefs, var_beliefs, numVars, var_degrees, batch_factors, batch_vars, debug=False):
'''
Equation (38) in:
https://www.cs.princeton.edu/courses/archive/spring06/cos598C/papers/YedidaFreemanWeiss2004.pdf
'''
pooled_fac_beliefs = -global_add_pool(torch.exp(factor_beliefs)*neg_inf_to_zero(factor_beliefs), batch_factors)
#keep 1st dimension for # of factors, but flatten remaining dimensions for belief_repeats and each factor
pooled_fac_beliefs = pooled_fac_beliefs.view(pooled_fac_beliefs.shape[0], -1)
if debug:
factor_beliefs_shape = factor_beliefs.shape
pooled_fac_beliefs_orig = -torch.sum((torch.exp(factor_beliefs)*neg_inf_to_zero(factor_beliefs)).view(factor_beliefs_shape[0], -1), dim=0)
print("pooled_fac_beliefs_orig:", pooled_fac_beliefs_orig)
print("pooled_fac_beliefs:", pooled_fac_beliefs)
var_beliefs_shape = var_beliefs.shape
assert(var_beliefs_shape[0] == var_degrees.shape[0])
pooled_var_beliefs = global_add_pool(torch.exp(var_beliefs)*neg_inf_to_zero(var_beliefs)*(var_degrees.float() - 1).view(var_degrees.shape[0], 1, 1), batch_vars)
#keep 1st dimension for # of factors, but flatten remaining dimensions for belief_repeats and variable states
pooled_var_beliefs = pooled_var_beliefs.view(pooled_var_beliefs.shape[0], -1)
if debug:
pooled_var_beliefs_orig = torch.sum(torch.exp(var_beliefs)*neg_inf_to_zero(var_beliefs)*(var_degrees.float() - 1).view(var_beliefs_shape[0], -1), dim=0)
print("pooled_var_beliefs_orig:", pooled_var_beliefs_orig)
print("pooled_var_beliefs:", pooled_var_beliefs)
# sleep(SHAPECHECK)
return pooled_fac_beliefs, pooled_var_beliefs
def compute_bethe_average_energy(factor_beliefs, factor_potentials, debug=False):
'''
Equation (37) in:
https://www.cs.princeton.edu/courses/archive/spring06/cos598C/papers/YedidaFreemanWeiss2004.pdf
'''
assert(factor_potentials.shape == factor_beliefs.shape)
if debug:
print()
print('!!!!!!!')
print("debugging compute_bethe_average_energy")
print("torch.exp(factor_beliefs):", torch.exp(factor_beliefs))
print("neg_inf_to_zero(factor_potentials):", neg_inf_to_zero(factor_potentials))
bethe_average_energy = -torch.sum(torch.exp(factor_beliefs)*neg_inf_to_zero(factor_potentials)) #elementwise multiplication, then sum
# print("bethe_average_energy:", bethe_average_energy)
return bethe_average_energy
def compute_bethe_entropy(self, factor_beliefs, var_beliefs):
'''
Equation (38) in:
https://www.cs.princeton.edu/courses/archive/spring06/cos598C/papers/YedidaFreemanWeiss2004.pdf
'''
bethe_entropy = -torch.sum(torch.exp(factor_beliefs)*neg_inf_to_zero(factor_beliefs)) #elementwise multiplication, then sum
assert(var_beliefs.shape == torch.Size([self.numVars, 2])), (var_beliefs.shape, [self.numVars, 2])
# sum_{x_i} b_i(x_i)*ln(b_i(x_i))
inner_sum = torch.einsum('ij,ij->i', [torch.exp(var_beliefs), neg_inf_to_zero(var_beliefs)])
# sum_{i=1}^N (d_i - 1)*inner_sum
outer_sum = torch.sum((self.var_degrees.float() - 1) * inner_sum)
# outer_sum = torch.einsum('i,i->', [self.var_degrees - 1, inner_sum])
bethe_entropy += outer_sum
# print("bethe_entropy:", bethe_entropy)
return bethe_entropy