def hmc_propose(chains, adapted_approximation, adapted_scale, maxGradient):
e = 0.25
start_vectors = chains.vectors
startp = chains.logps
p = random.multivariate_normal(
mean=zeros(chains.dimensions), cov=adapted_approximation.inv_orientation, size=chains._nChains
)
start_p = p
p = p - (e / 2) * (-chains.logp_grads)
T = 3
for i in range(T):
chains.propose(chains.vectors + e * vectorsMult(adapted_approximation.orientation, p))
if i != T - 1:
p = p - e * (-chains.logp_grads)
p = p - (e / 2) * (-chains.logp_grads)
p = -p
end_vectors = chains.vectors
chains.propose(start_vectors)
chains.propose(end_vectors)
chains.vectors - start_vectors
# ( (-startp) - (-chains.logps) + sum(start_p**2, axis =1)/2 - sum(p**2, axis =1)/2)
# ( (-startp) - (-chains.logps) + kenergy(start_p, adapted_approximation.orientation) - kenergy(p, adapted_approximation.orientation))
# return sum(p**2, axis =1)/2,sum(start_p**2, axis =1) /2
return kenergy(p, adapted_approximation.orientation), kenergy(start_p, adapted_approximation.orientation)
def hmc_propose(chains, adapted_approximation, adapted_scale, maxGradient):
e = .25
start_vectors = chains.vectors
startp = chains.logps
p = random.multivariate_normal(mean=zeros(chains.dimensions),
cov=adapted_approximation.inv_orientation,
size=chains._nChains)
start_p = p
p = p - (e / 2) * (-chains.logp_grads)
T = 3
for i in range(T):
chains.propose(chains.vectors +
e * vectorsMult(adapted_approximation.orientation, p))
if i != T - 1:
p = p - e * (-chains.logp_grads)
p = p - (e / 2) * (-chains.logp_grads)
p = -p
end_vectors = chains.vectors
chains.propose(start_vectors)
chains.propose(end_vectors)
chains.vectors - start_vectors
#( (-startp) - (-chains.logps) + sum(start_p**2, axis =1)/2 - sum(p**2, axis =1)/2)
#( (-startp) - (-chains.logps) + kenergy(start_p, adapted_approximation.orientation) - kenergy(p, adapted_approximation.orientation))
#return sum(p**2, axis =1)/2,sum(start_p**2, axis =1) /2
return kenergy(p, adapted_approximation.orientation), kenergy(
start_p, adapted_approximation.orientation)
def drift(x, gradient):
return vectorsMult(scaledOrientation, truncate_gradient(x, gradient,adapted_approximation, maxGradient) /2)
