p = StandardNormal((target.size, )).to(args.device)
model_id = get_model_id(args)
state_dict = torch.load(path_check)
pi.load_state_dict(state_dict)
##############
## Training ##
##############
print('Running MCMC...')
time_before = time.time()
samples, rate = metropolis_hastings(
pi=pi,
num_dims=target.size,
num_chains=eval_args.num_chains,
num_samples=eval_args.num_samples,
steps_per_sample=eval_args.steps_per_sample,
burnin_steps=eval_args.burnin_steps,
proposal_scale=eval_args.proposal_scale)
runtime = time.time() - time_before
print('')
print('Projecting...')
num_chains, num_samples, dim = samples.shape
theta = torch.zeros(num_chains, num_samples, dim)
with torch.no_grad():
for i in range(num_samples):
print('{}/{}'.format(i + 1, num_samples), end='\r')
for t in pi.transforms:
samples[:, i], _ = t(samples[:, i])
theta[:, i] = samples[:, i].detach().cpu()
def test_univariate_functions(f, x_0, expected_mean, expected_std):
samples = metropolis_hastings(f, x_0)
np.testing.assert_almost_equal(samples.mean(), expected_mean, decimal=1)
np.testing.assert_almost_equal(samples.std(), expected_std, decimal=1)
def test_raise_exception_for_non_functions():
with pytest.raises(TypeError):
samples = metropolis_hastings(sum([]))
def test_raise_exception_for_wrong_n_samples(n_samples):
with pytest.raises(ValueError):
samples = metropolis_hastings(standard_norm_dist, 0, n_samples)
def test_correctness():
samples = metropolis_hastings(standard_norm_dist, 0, 100)
assert len(samples) == 100
assert isinstance(samples, np.ndarray)
assert np.issubdtype(samples.dtype, np.floating)