def f(v):
hessp_beta1 = GNvp(c1, z1, tmp_params1, v)
hessp_beta2 = GNvp(c2, z2, tmp_params2, v)
if step >= 1:
weighted_hessp = ((step - 1) * hessp_beta1 + hessp_beta2) / step
else:
weighted_hessp = hessp_beta2
return weighted_hessp.data / bias_correction2
def f(p):
# pvar = Variable(p.float(), requires_grad=False)
pvar = torch.nn.Parameter(p.float()) #, requires_grad=False)
# vector_to_parameters(pvar, self._params_tmp)
import time
s = time.time()
# c, z, tmp_params = closure(pvar)
c, z, tmp_params = closure([pvar]) #self._params_tmp)
e = time.time()
# print ("Closure time: ", (e-s))
v1 = GNvp(c, z, tmp_params, ng)
# v2 = weighted_fvp_fn(ng)
loss = F.mse_loss(v1, v2)
return float(loss.data)
def f(v):
hessp = GNvp(c, z, params, v)
return hessp.data / bias_correction2