def bdmc(model, loader, forward_schedule, n_sample):
"""Bidirectional Monte Carlo.
Backward schedule is set to be the reverse of the forward schedule.
Args:
model (vae.VAE): VAE model
loader (iterator): iterator to loop over pairs of Variables; the first
entry being `x`, the second being `z` sampled from the *true*
posterior `p(z|x)`
forward_schedule: forward temperature schedule
n_sample (int): number of importance samples
Returns:
two lists for forward and backward bounds on batches of data
"""
# iterator is exhaustible in py3, so need duplicate
loader_forward, loader_backward = itertools.tee(loader, 2)
# forward chain
forward_logws = ais.ais_trajectory(
model,
loader_forward,
forward=True,
schedule=forward_schedule,
n_sample=n_sample,
device=device,
)
# backward chain
backward_schedule = torch.flip(forward_schedule, dims=(0,)).contiguous()
backward_logws = ais.ais_trajectory(
model,
loader_backward,
forward=False,
schedule=backward_schedule,
n_sample=n_sample,
device=device,
)
upper_bounds = []
lower_bounds = []
for i, (forward, backward) in enumerate(zip(forward_logws, backward_logws)):
lower_bounds.append(forward.mean().detach().item())
upper_bounds.append(backward.mean().detach().item())
upper_bounds = float(np.mean(upper_bounds))
lower_bounds = float(np.mean(lower_bounds))
print(
f"Average bounds on simulated data: lower {lower_bounds:.4f}, upper {upper_bounds:.4f}"
)
return forward_logws, backward_logws
def bdmc(model, loader, forward_schedule, n_sample,
device=torch.device("cpu")):
"""Bidirectional Monte Carlo. Backward schedule is set to be the reverse of
the forward schedule.
Args:
model (vae.VAE): VAE model
loader (iterator): iterator to loop over pairs of Variables; the first
entry being `x`, the second being `z` sampled from the *true*
posterior `p(z|x)`
forward_schedule (list or numpy.ndarray): forward temperature schedule
Returns:
Two lists for forward and backward bounds on batchs of data
"""
# iterator is exhaustable in py3, so need duplicate
load, load_ = itertools.tee(loader, 2)
# forward chain
forward_logws = ais.ais_trajectory(
model,
load,
forward=True,
schedule=forward_schedule,
n_sample=n_sample,
device=device,
)
# backward chain
backward_schedule = np.flip(forward_schedule, axis=0)
backward_logws = ais.ais_trajectory(
model,
load_,
forward=False,
schedule=backward_schedule,
n_sample=n_sample,
device=device,
)
upper_bounds = []
lower_bounds = []
for i, (forward, backward) in enumerate(zip(forward_logws,
backward_logws)):
lower_bounds.append(forward.mean())
upper_bounds.append(backward.mean())
upper_bounds = np.mean(upper_bounds)
lower_bounds = np.mean(lower_bounds)
print('Average bounds on simulated data: lower %.4f, upper %.4f' %
(lower_bounds, upper_bounds))
return forward_logws, backward_logws
def forward_ais(model,
loader,
forward_schedule=np.linspace(0., 1., 500),
n_sample=100):
"""Bidirectional Monte Carlo. Integrate forward and backward AIS.
The backward schedule is the reverse of the forward.
Args:
model (vae.VAE): VAE model
loader (iterator): iterator to loop over pairs of Variables; the first
entry being `x`, the second being `z` sampled from the true
posterior `p(z|x)`
forward_schedule (list or numpy.ndarray): forward temperature schedule;
backward schedule is used as its reverse
Returns:
Two lists for forward and backward bounds on batchs of data
"""
# iterator is exhaustable in py3, so need duplicate
load, load_ = itertools.tee(loader, 2)
# forward chain
forward_logws = ais_trajectory(model,
load,
mode='forward',
schedule=forward_schedule,
n_sample=n_sample)
lower_bounds = []
for forward in forward_logws:
lower_bounds.append(forward.mean())
lower_bounds = np.mean(lower_bounds)
return forward_logws