def main(args):
assert 11_000_000 >= args.num_datapoints, "11,000,000 data points in the Higgs dataset"
# full dataset takes hours for plain hmc!
if args.dataset == 'higgs':
_, fetch = load_dataset(HIGGS, shuffle=False, num_datapoints=args.num_datapoints)
data, obs = fetch()
else:
data, obs = (np.random.normal(size=(10, 28)), np.ones(10))
hmcecs_key, hmc_key = random.split(random.PRNGKey(args.rng_seed))
# choose inner_kernel
if args.inner_kernel == 'hmc':
inner_kernel = HMC(model)
else:
inner_kernel = NUTS(model)
start = time.time()
losses, hmcecs_samples = run_hmcecs(hmcecs_key, args, data, obs, inner_kernel)
hmcecs_runtime = time.time() - start
start = time.time()
hmc_samples = run_hmc(hmc_key, args, data, obs, inner_kernel)
hmc_runtime = time.time() - start
summary_plot(losses, hmc_samples, hmcecs_samples, hmc_runtime, hmcecs_runtime)
def run_inference(model, capture_history, sex, rng_key, args):
if args.algo == "NUTS":
kernel = NUTS(model)
elif args.algo == "HMC":
kernel = HMC(model)
mcmc = MCMC(kernel, args.num_warmup, args.num_samples, num_chains=args.num_chains,
progress_bar=False if "NUMPYRO_SPHINXBUILD" in os.environ else True)
mcmc.run(rng_key, capture_history, sex)
mcmc.print_summary()
return mcmc.get_samples()
def run_inference(model, at_bats, hits, rng_key, args):
if args.algo == "NUTS":
kernel = NUTS(model)
elif args.algo == "HMC":
kernel = HMC(model)
elif args.algo == "SA":
kernel = SA(model)
mcmc = MCMC(kernel, args.num_warmup, args.num_samples, num_chains=args.num_chains,
progress_bar=False if (
"NUMPYRO_SPHINXBUILD" in os.environ or args.disable_progbar) else True)
mcmc.run(rng_key, at_bats, hits)
return mcmc.get_samples()
def run_inference(
model: Callable,
at_bats: jnp.ndarray,
hits: jnp.ndarray,
rng_key: jnp.ndarray,
*,
num_warmup: int = 1500,
num_samples: int = 3000,
num_chains: int = 1,
algo_name: str = "NUTS",
) -> Dict[str, jnp.ndarray]:
if algo_name == "NUTS":
kernel = NUTS(model)
elif algo_name == "HMC":
kernel = HMC(model)
elif algo_name == "SA":
kernel = SA(model)
else:
raise ValueError("Unknown algorithm name")
mcmc = MCMC(kernel, num_warmup, num_samples, num_chains)
mcmc.run(rng_key, at_bats, hits)
return mcmc.get_samples()
def benchmark_hmc(args, features, labels):
rng_key = random.PRNGKey(1)
start = time.time()
# a MAP estimate at the following source
# https://github.com/google/edward2/blob/master/examples/no_u_turn_sampler/logistic_regression.py#L117
ref_params = {
"coefs":
jnp.array([
+2.03420663e00,
-3.53567265e-02,
-1.49223924e-01,
-3.07049364e-01,
-1.00028366e-01,
-1.46827862e-01,
-1.64167881e-01,
-4.20344204e-01,
+9.47479829e-02,
-1.12681836e-02,
+2.64442056e-01,
-1.22087866e-01,
-6.00568838e-02,
-3.79419506e-01,
-1.06668741e-01,
-2.97053963e-01,
-2.05253899e-01,
-4.69537191e-02,
-2.78072730e-02,
-1.43250525e-01,
-6.77954629e-02,
-4.34899796e-03,
+5.90927452e-02,
+7.23133609e-02,
+1.38526391e-02,
-1.24497898e-01,
-1.50733739e-02,
-2.68872194e-02,
-1.80925727e-02,
+3.47936489e-02,
+4.03552800e-02,
-9.98773426e-03,
+6.20188080e-02,
+1.15002751e-01,
+1.32145107e-01,
+2.69109547e-01,
+2.45785132e-01,
+1.19035013e-01,
-2.59744357e-02,
+9.94279515e-04,
+3.39266285e-02,
-1.44057125e-02,
-6.95222765e-02,
-7.52013028e-02,
+1.21171586e-01,
+2.29205526e-02,
+1.47308692e-01,
-8.34354162e-02,
-9.34122875e-02,
-2.97472421e-02,
-3.03937674e-01,
-1.70958012e-01,
-1.59496680e-01,
-1.88516974e-01,
-1.20889175e00,
])
}
if args.algo == "HMC":
step_size = jnp.sqrt(0.5 / features.shape[0])
trajectory_length = step_size * args.num_steps
kernel = HMC(
model,
step_size=step_size,
trajectory_length=trajectory_length,
adapt_step_size=False,
dense_mass=args.dense_mass,
)
subsample_size = None
elif args.algo == "NUTS":
kernel = NUTS(model, dense_mass=args.dense_mass)
subsample_size = None
elif args.algo == "HMCECS":
subsample_size = 1000
inner_kernel = NUTS(
model,
init_strategy=init_to_value(values=ref_params),
dense_mass=args.dense_mass,
)
# note: if num_blocks=100, we'll update 10 index at each MCMC step
# so it took 50000 MCMC steps to iterative the whole dataset
kernel = HMCECS(inner_kernel,
num_blocks=100,
proxy=HMCECS.taylor_proxy(ref_params))
elif args.algo == "SA":
# NB: this kernel requires large num_warmup and num_samples
# and running on GPU is much faster than on CPU
kernel = SA(model,
adapt_state_size=1000,
init_strategy=init_to_value(values=ref_params))
subsample_size = None
elif args.algo == "FlowHMCECS":
subsample_size = 1000
guide = AutoBNAFNormal(model, num_flows=1, hidden_factors=[8])
svi = SVI(model, guide, numpyro.optim.Adam(0.01), Trace_ELBO())
svi_result = svi.run(random.PRNGKey(2), 2000, features, labels)
params, losses = svi_result.params, svi_result.losses
plt.plot(losses)
plt.show()
neutra = NeuTraReparam(guide, params)
neutra_model = neutra.reparam(model)
neutra_ref_params = {"auto_shared_latent": jnp.zeros(55)}
# no need to adapt mass matrix if the flow does a good job
inner_kernel = NUTS(
neutra_model,
init_strategy=init_to_value(values=neutra_ref_params),
adapt_mass_matrix=False,
)
kernel = HMCECS(inner_kernel,
num_blocks=100,
proxy=HMCECS.taylor_proxy(neutra_ref_params))
else:
raise ValueError(
"Invalid algorithm, either 'HMC', 'NUTS', or 'HMCECS'.")
mcmc = MCMC(kernel,
num_warmup=args.num_warmup,
num_samples=args.num_samples)
mcmc.run(rng_key,
features,
labels,
subsample_size,
extra_fields=("accept_prob", ))
print("Mean accept prob:",
jnp.mean(mcmc.get_extra_fields()["accept_prob"]))
mcmc.print_summary(exclude_deterministic=False)
print("\nMCMC elapsed time:", time.time() - start)
def test_pickle_discrete_hmc(kernel):
mcmc = MCMC(kernel(HMC(bernoulli_model)), num_warmup=10, num_samples=10)
mcmc.run(random.PRNGKey(0))
pickled_mcmc = pickle.loads(pickle.dumps(mcmc))
test_util.check_close(mcmc.get_samples(), pickled_mcmc.get_samples())