def sample(draws=500,
model=None,
warmup_steps=None,
num_chains=1,
kernel='nuts'):
"""Markov-chain Monte Carlo sampling.
Sampling should be run within the context of a model or the model should be passed as an argument `model` explicitly.
Number of samples is given by `draws` which defaults to `500`.
Warm-up steps are assumed to be 30% of sample count.
MCMC kernel can be selected by setting `kernel`. `hmc` and `nuts` are available.
`pmpyro.inference.sample` returns a trace of samples
"""
# get model from context
if model is None:
model = Context.get_context()
stfn = model.stfn # get stochastic function from model
data = model.args # get data
# make nuts kernel
kernels = {'nuts': NUTS(stfn, adapt_step_size=True), 'hmc': HMC(stfn)}
# if not num_chains: # figure out number of chains
# num_chains = max(os.cpu_count() -1, 2)
if not warmup_steps: # figure out warm-up steps
warmup_steps = int(0.3 * draws)
# run MCMC
mcmc = MCMC(kernels[kernel],
num_samples=draws,
warmup_steps=warmup_steps,
num_chains=num_chains)
mcmc.run(*data)
# get num samples
num_samples = num_chains * draws
return mcmc.get_samples()
def main():
train_x = torch.linspace(0, 1, 1000).double()
train_y = torch.sin(train_x * (2 * math.pi)).double() + torch.randn(train_x.size()).double() * 0.1
# Use a positive constraint instead of usual GreaterThan(1e-4) so that LogNormal has support over full range.
likelihood = gpytorch.likelihoods.GaussianLikelihood(noise_constraint=gpytorch.constraints.Positive())
model = ExactGPModel(train_x, train_y, likelihood)
# model.mean_module.register_prior("mean_prior", UniformPrior(-1, 1), "constant")
# model.covar_module.base_kernel.register_prior("lengthscale_prior", UniformPrior(0.0, 9.0), "lengthscale")
# # model.covar_module.base_kernel.register_prior("period_length_prior", UniformPrior(0.0, 4.0), "period_length")
# model.covar_module.register_prior("outputscale_prior", UniformPrior(0, 4), "outputscale")
# likelihood.register_prior("noise_prior", UniformPrior(0.0, 0.25), "noise")
mll = gpytorch.mlls.ExactMarginalLogLikelihood(likelihood, model)
def pyro_model(x, y):
priors= {
'covar_module.base_kernel.raw_lengthscale': Normal(0, 2).expand([1, 1]),
'covar_module.raw_outputscale': Normal(0, 2),
'likelihood.noise_covar.raw_noise': Normal(0, 2).expand([1]),
'mean_module.constant': Normal(0, 2),
}
fn = pyro.random_module("model", model, prior=priors)
sampled_model = fn()
output = sampled_model.likelihood(sampled_model(x))
pyro.sample("obs", output, obs=y)
# model.mean_module.constant.data.fill_(0.0)
# model.covar_module.outputscale = 0.5**2
# model.covar_module.base_kernel.lengthscale = 1
# model.likelihood.noise = 0.05**2
model.double()
likelihood.double()
nuts_kernel = NUTS(pyro_model, adapt_step_size=True, jit_compile=False)
hmc_kernel = HMC(pyro_model, step_size=0.1, num_steps=10, adapt_step_size=True,\
init_strategy=pyro.infer.autoguide.initialization.init_to_median(num_samples=20))
mcmc_run = MCMC(nuts_kernel, num_samples=num_samples, warmup_steps=warmup_steps)#, initial_params=initial_params)
return model, likelihood, mll, mcmc_run, train_x, train_y
labels = ["c", "ls","os","noise"]
fig, axes = plt.subplots(nrows=2, ncols=2)
for i in range(4):
if i == 0:
samples = getattr(chain, labels[i]).reshape(-1)
else:
samples = 1/(1+np.exp(-1*getattr(chain, labels[i]).reshape(-1)))
sns.distplot(samples, ax=axes[int(i/2), int(i%2)])
axes[int(i/2)][int(i%2)].legend([labels[i]])
plt.show()
pickle.dump(chain, open("results/test_mcmc.pkl", "wb"))
return
def test_model_with_potential_fn():
init_params = {"z": torch.tensor(0.)}
def potential_fn(params):
return params["z"]
mcmc = MCMC(kernel=HMC(potential_fn=potential_fn),
num_samples=10,
warmup_steps=10,
initial_params=init_params)
mcmc.run()
def sample_posterior_mcmc(self, num_samples, thin=10):
"""
Samples from posterior for true observation q(theta | x0) using MCMC.
:param num_samples: Number of samples to generate.
:param mcmc_method: Which MCMC method to use ['metropolis-hastings', 'slice', 'hmc', 'nuts']
:param thin: Generate (num_samples * thin) samples in total, then select every
'thin' sample.
:return: torch.Tensor of shape [num_samples, parameter_dim]
"""
# Always sample in eval mode.
self._neural_posterior.eval()
if self._mcmc_method == "slice-np":
self.posterior_sampler.gen(20) # Burn-in for 200 samples
samples = torch.Tensor(self.posterior_sampler.gen(num_samples))
else:
if self._mcmc_method == "slice":
kernel = Slice(potential_function=self._potential_function)
elif self._mcmc_method == "hmc":
kernel = HMC(potential_fn=self._potential_function)
elif self._mcmc_method == "nuts":
kernel = NUTS(potential_fn=self._potential_function)
else:
raise ValueError(
"'mcmc_method' must be one of ['slice', 'hmc', 'nuts']."
)
num_chains = mp.cpu_count() - 1
initial_params = self._prior.sample((num_chains,))
sampler = MCMC(
kernel=kernel,
num_samples=(thin * num_samples) // num_chains + num_chains,
warmup_steps=200,
initial_params={"": initial_params},
num_chains=num_chains,
mp_context="spawn",
)
sampler.run()
samples = next(iter(sampler.get_samples().values())).reshape(
-1, self._simulator.parameter_dim
)
samples = samples[::thin][:num_samples]
assert samples.shape[0] == num_samples
# Back to training mode.
self._neural_posterior.train()
return samples
def sample_posterior(self, num_samples, thin=1):
"""
Samples from posterior for true observation q(theta | x0) ~ q(x0 | theta) p(theta)
using most recent likelihood estimate q(x0 | theta) with MCMC.
:param num_samples: Number of samples to generate.
:param thin: Generate (num_samples * thin) samples in total, then select every
'thin' sample.
:return: torch.Tensor of shape [num_samples, parameter_dim]
"""
# Always sample in eval mode.
self._neural_likelihood.eval()
if self._mcmc_method == "slice-np":
self.posterior_sampler.gen(20)
samples = torch.Tensor(self.posterior_sampler.gen(num_samples))
else:
if self._mcmc_method == "slice":
kernel = Slice(potential_function=self._potential_function)
elif self._mcmc_method == "hmc":
kernel = HMC(potential_fn=self._potential_function)
elif self._mcmc_method == "nuts":
kernel = NUTS(potential_fn=self._potential_function)
else:
raise ValueError(
"'mcmc_method' must be one of ['slice', 'hmc', 'nuts']."
)
num_chains = mp.cpu_count() - 1
# TODO: decide on way to initialize chain
initial_params = self._prior.sample((num_chains,))
sampler = MCMC(
kernel=kernel,
num_samples=num_samples // num_chains + num_chains,
warmup_steps=200,
initial_params={"": initial_params},
num_chains=num_chains,
)
sampler.run()
samples = next(iter(sampler.get_samples().values())).reshape(
-1, self._simulator.parameter_dim
)
samples = samples[:num_samples].to(device)
assert samples.shape[0] == num_samples
# Back to training mode.
self._neural_likelihood.train()
return samples
def test_model_with_potential_fn(run_mcmc_cls):
init_params = {"z": torch.tensor(0.0)}
def potential_fn(params):
return params["z"]
run_mcmc_cls(
data=None,
kernel=HMC(potential_fn=potential_fn),
num_samples=10,
warmup_steps=10,
initial_params=init_params,
)
def _train_hmc(self, train_loader, n_samples, warmup, step_size, num_steps,
device):
print("\n == HMC training ==")
pyro.clear_param_store()
num_batches = int(len(train_loader.dataset) / train_loader.batch_size)
batch_samples = int(n_samples / num_batches) + 1
print("\nn_batches=", num_batches, "\tbatch_samples =", batch_samples)
kernel = HMC(self.model, step_size=step_size, num_steps=num_steps)
mcmc = MCMC(kernel=kernel,
num_samples=batch_samples,
warmup_steps=warmup,
num_chains=1)
start = time.time()
for x_batch, y_batch in train_loader:
x_batch = x_batch.to(device)
labels = y_batch.to(device).argmax(-1)
mcmc.run(x_batch, labels)
execution_time(start=start, end=time.time())
self.posterior_predictive = {}
posterior_samples = mcmc.get_samples(n_samples)
state_dict_keys = list(self.basenet.state_dict().keys())
if DEBUG:
print("\n", list(posterior_samples.values())[-1])
for model_idx in range(n_samples):
net_copy = copy.deepcopy(self.basenet)
model_dict = OrderedDict({})
for weight_idx, weights in enumerate(posterior_samples.values()):
model_dict.update(
{state_dict_keys[weight_idx]: weights[model_idx]})
net_copy.load_state_dict(model_dict)
self.posterior_predictive.update({str(model_idx): net_copy})
if DEBUG:
print("\n", weights[model_idx])
self.save()
def monte_carlo(y):
pyro.clear_param_store()
# create a Simple Hamiltonian Monte Carlo kernel with step_size of 0.1
hmc_kernel = HMC(conditioned_model, step_size=.1)
mcmc = MCMC(hmc_kernel, num_samples=500, warmup_steps=100)
# create a Markov Chain Monte Carlo method with:
# the hmc_kernel, 500 samples, and 100 warmup iterations
mcmc.run(model,y)
mcmc.run(model, y)
sample_dict = mcmc.get_samples(num_samples=5000)
plt.figure(figsize=(8, 6))
sns.distplot(sample_dict["p"].numpy())
plt.xlabel("Observed probability value")
plt.ylabel("Observed frequency")
plt.show()
mcmc.summary(prob=0.95)
return sample_dict
def mcmc(model,
obs,
num_samples,
kernel='HMC',
kernel_params={},
mcmc_params={},
sites=['theta']):
# NOTE: requires differentiable model
model_conditioned = partial(model, obs=obs)
if kernel.upper() == 'HMC':
mcmc_kernel = HMC(model_conditioned, **kernel_params)
elif kernel.upper() == 'NUTS':
mcmc_kernel = NUTS(model_conditioned, **kernel_params)
else:
raise NotImplementedError
mcmc = MCMC(mcmc_kernel, num_samples, **mcmc_params)
mcmc_run = mcmc.run()
posterior = pyro.infer.EmpiricalMarginal(mcmc_run, sites=sites)
return posterior
def run(
task: Task,
num_samples: int,
num_observation: Optional[int] = None,
observation: Optional[torch.Tensor] = None,
num_chains: int = 10,
num_warmup: int = 10000,
kernel: str = "slice",
kernel_parameters: Optional[Dict[str, Any]] = None,
thinning: int = 1,
diagnostics: bool = True,
available_cpu: int = 1,
mp_context: str = "fork",
jit_compile: bool = False,
automatic_transforms_enabled: bool = True,
initial_params: Optional[torch.Tensor] = None,
**kwargs: Any,
) -> torch.Tensor:
"""Runs MCMC using Pyro on potential function
Produces `num_samples` while accounting for warmup (burn-in) and thinning.
Note that the actual number of simulations is not controlled for with MCMC since
algorithms are only used as a reference method in the benchmark.
MCMC is run on the potential function, which returns the unnormalized
negative log posterior probability. Note that this requires a tractable likelihood.
Pyro is used to automatically construct the potential function.
Args:
task: Task instance
num_samples: Number of samples to generate from posterior
num_observation: Observation number to load, alternative to `observation`
observation: Observation, alternative to `num_observation`
num_chains: Number of chains
num_warmup: Warmup steps, during which parameters of the sampler are adapted.
Warmup samples are not returned by the algorithm.
kernel: HMC, NUTS, or Slice
kernel_parameters: Parameters passed to kernel
thinning: Amount of thinning to apply, in order to avoid drawing
correlated samples from the chain
diagnostics: Flag for diagnostics
available_cpu: Number of CPUs used to parallelize chains
mp_context: multiprocessing context, only fork might work
jit_compile: Just-in-time (JIT) compilation, can yield significant speed ups
automatic_transforms_enabled: Whether or not to use automatic transforms
initial_params: Parameters to initialize at
Returns:
Samples from posterior
"""
assert not (num_observation is None and observation is None)
assert not (num_observation is not None and observation is not None)
tic = time.time()
log = sbibm.get_logger(__name__)
hook_fn = None
if diagnostics:
log.info(f"MCMC sampling for observation {num_observation}")
tb_writer, tb_close = tb_make_writer(
logger=log,
basepath=
f"tensorboard/pyro_{kernel.lower()}/observation_{num_observation}",
)
hook_fn = tb_make_hook_fn(tb_writer)
if "num_simulations" in kwargs:
warnings.warn(
"`num_simulations` was passed as a keyword but will be ignored, see docstring for more info."
)
# Prepare model and transforms
conditioned_model = task._get_pyro_model(num_observation=num_observation,
observation=observation)
transforms = task._get_transforms(
num_observation=num_observation,
observation=observation,
automatic_transforms_enabled=automatic_transforms_enabled,
)
kernel_parameters = kernel_parameters if kernel_parameters is not None else {}
kernel_parameters["jit_compile"] = jit_compile
kernel_parameters["transforms"] = transforms
log.info("Using kernel: {name}({parameters})".format(
name=kernel,
parameters=",".join([f"{k}={v}"
for k, v in kernel_parameters.items()]),
))
if kernel.lower() == "nuts":
mcmc_kernel = NUTS(model=conditioned_model, **kernel_parameters)
elif kernel.lower() == "hmc":
mcmc_kernel = HMC(model=conditioned_model, **kernel_parameters)
elif kernel.lower() == "slice":
mcmc_kernel = Slice(model=conditioned_model, **kernel_parameters)
else:
raise NotImplementedError
if initial_params is not None:
site_name = "parameters"
initial_params = {site_name: transforms[site_name](initial_params)}
else:
initial_params = None
mcmc_parameters = {
"num_chains": num_chains,
"num_samples": thinning * num_samples,
"warmup_steps": num_warmup,
"available_cpu": available_cpu,
"initial_params": initial_params,
}
log.info("Calling MCMC with: MCMC({name}_kernel, {parameters})".format(
name=kernel,
parameters=",".join([f"{k}={v}" for k, v in mcmc_parameters.items()]),
))
mcmc = MCMC(mcmc_kernel, hook_fn=hook_fn, **mcmc_parameters)
mcmc.run()
toc = time.time()
log.info(f"Finished MCMC after {toc-tic:.3f} seconds")
log.info(f"Automatic transforms {mcmc.transforms}")
log.info(f"Apply thinning of {thinning}")
mcmc._samples = {
"parameters": mcmc._samples["parameters"][:, ::thinning, :]
}
num_samples_available = (mcmc._samples["parameters"].shape[0] *
mcmc._samples["parameters"].shape[1])
if num_samples_available < num_samples:
warnings.warn("Some samples will be included multiple times")
samples = mcmc.get_samples(
num_samples=num_samples,
group_by_chain=False)["parameters"].squeeze()
else:
samples = mcmc.get_samples(
group_by_chain=False)["parameters"].squeeze()
idx = torch.randperm(samples.shape[0])[:num_samples]
samples = samples[idx, :]
assert samples.shape[0] == num_samples
if diagnostics:
mcmc.summary()
tb_ess(tb_writer, mcmc)
tb_r_hat(tb_writer, mcmc)
tb_marginals(tb_writer, mcmc)
tb_acf(tb_writer, mcmc)
tb_posteriors(tb_writer, mcmc)
tb_plot_posterior(tb_writer, samples, tag="posterior/final")
tb_close()
return samples
plt.subplot(2, 2, 4)
plt.hist(alpha.numpy(), bins=30, density=True)
plt.xlabel('alpha')
plt.ylabel('density')
# In[9]:
pyro.clear_param_store()
# Set random seed for reproducibility.
pyro.set_rng_seed(1)
# Set up HMC sampler.
kernel = HMC(dp_sb_gmm,
step_size=0.01,
trajectory_length=1,
adapt_step_size=False,
adapt_mass_matrix=False,
jit_compile=True)
hmc = MCMC(kernel, num_samples=500, warmup_steps=500)
hmc.run(y, 10)
# 06:13 for marginalized version.
# Get posterior samples
hmc_posterior_samples = hmc.get_samples()
hmc_posterior_samples['eta'] = stickbreak(hmc_posterior_samples['v'])
# In[10]:
pyro.clear_param_store()
def localnews(INFERENCE):
device = torch.device('cpu')
torch.set_default_tensor_type(torch.DoubleTensor)
if torch.cuda.is_available():
device = torch.device('cuda')
torch.set_default_tensor_type(torch.cuda.DoubleTensor)
# preprocess data
data = pd.read_csv("data/localnews.csv", index_col=[0])
N = data.station_id.unique().shape[0]
data.date = data.date.apply(
lambda x: datetime.datetime.strptime(x, '%m/%d/%Y').date())
# data = data[(data.date<=datetime.date(2017, 9, 5)) & (data.date>=datetime.date(2017, 8, 25))]
# data = data[data.station_id.isin([1345,3930])]
ds = data.t.to_numpy().reshape((-1, 1))
ohe = OneHotEncoder()
ohe = LabelEncoder()
X = data.drop(columns=[
"station_id", "date", "national_politics", "sinclair2017", "post",
"affiliation", "callsign", "t"
]).to_numpy().reshape(-1, ) # , "weekday","affiliation","callsign"
Group = data.sinclair2017.to_numpy().reshape(-1, 1)
ohe.fit(X)
X = ohe.transform(X)
station_le = LabelEncoder()
ids = data.station_id.to_numpy().reshape(-1, )
station_le.fit(ids)
ids = station_le.transform(ids)
# weekday/day/unit effects and time trend
X = np.concatenate((X.reshape(-1, 1), ds, ids.reshape(-1, 1), Group, ds),
axis=1)
# numbers of dummies for each effect
X_max_v = [np.max(X[:, i]).astype(int) for i in range(X.shape[1] - 2)]
Y = data.national_politics.to_numpy()
T0 = data[data.date == datetime.date(2017, 9, 1)].t.to_numpy()[0]
train_condition = (data.post != 1) | (data.sinclair2017 != 1)
train_x = torch.Tensor(X[train_condition], device=device).double()
train_y = torch.Tensor(Y[train_condition], device=device).double()
idx = data.sinclair2017.to_numpy()
train_g = torch.from_numpy(idx[train_condition]).to(device)
test_x = torch.Tensor(X).double()
test_y = torch.Tensor(Y).double()
test_g = torch.from_numpy(idx)
# define likelihood
noise_prior = gpytorch.priors.GammaPrior(concentration=1, rate=10)
likelihood = gpytorch.likelihoods.GaussianLikelihood(noise_prior=noise_prior if "MAP" in INFERENCE else None,\
noise_constraint=gpytorch.constraints.Positive())
# likelihood2 = gpytorch.likelihoods.GaussianLikelihood(noise_prior=noise_prior if "MAP" in INFERENCE else None,\
# noise_constraint=gpytorch.constraints.Positive())
model = MultitaskGPModel(test_x,
test_y,
X_max_v,
likelihood,
MAP="MAP" in INFERENCE)
model.drift_t_module.T0 = T0
model2 = MultitaskGPModel(train_x,
train_y,
X_max_v,
likelihood,
MAP="MAP" in INFERENCE)
# model2 = MultitaskGPModel(test_x, test_y, X_max_v, likelihood2, MAP="MAP" in INFERENCE)
# model2.drift_t_module.T0 = T0
model2.double()
# group effects
# model.x_covar_module[0].c2 = torch.var(train_y)
# model.x_covar_module[0].raw_c2.requires_grad = False
# weekday/day/unit effects initialize to 0.05**2
for i in range(len(X_max_v)):
model.x_covar_module[i].c2 = torch.tensor(0.05**2)
# fix unit mean/variance by not requiring grad
model.x_covar_module[-1].raw_c2.requires_grad = False
# model.unit_mean_module.constant.data.fill_(0.12)
# model.unit_mean_module.constant.requires_grad = False
model.group_mean_module.constantvector.data[0].fill_(0.11)
model.group_mean_module.constantvector.data[1].fill_(0.12)
# set precision to double tensors
torch.set_default_tensor_type(torch.DoubleTensor)
train_x, train_y = train_x.to(device), train_y.to(device)
test_x, test_y = test_x.to(device), test_y.to(device)
model.to(device)
likelihood.to(device)
# define Loss for GPs - the marginal log likelihood
mll = gpytorch.mlls.ExactMarginalLogLikelihood(likelihood, model)
if torch.cuda.is_available():
train_x = train_x.cuda()
train_y = train_y.cuda()
model = model.cuda()
likelihood = likelihood.cuda()
if not os.path.isdir("results"):
os.mkdir("results")
transforms = {
'group_index_module.raw_rho':
model.group_index_module.raw_rho_constraint.transform,
'group_t_covar_module.base_kernel.raw_lengthscale':
model.group_t_covar_module.base_kernel.raw_lengthscale_constraint.
transform,
'group_t_covar_module.raw_outputscale':
model.group_t_covar_module.raw_outputscale_constraint.transform,
'unit_t_covar_module.base_kernel.raw_lengthscale':
model.unit_t_covar_module.base_kernel.raw_lengthscale_constraint.
transform,
'unit_t_covar_module.raw_outputscale':
model.unit_t_covar_module.raw_outputscale_constraint.transform,
'likelihood.noise_covar.raw_noise':
model.likelihood.noise_covar.raw_noise_constraint.transform,
'x_covar_module.0.raw_c2':
model.x_covar_module[0].raw_c2_constraint.transform,
'x_covar_module.1.raw_c2':
model.x_covar_module[1].raw_c2_constraint.transform
#'x_covar_module.2.raw_c2': model.x_covar_module[2].raw_c2_constraint.transform
}
priors = {
'group_index_module.raw_rho':
pyro.distributions.Normal(0, 1.5),
'group_t_covar_module.base_kernel.raw_lengthscale':
pyro.distributions.Normal(30, 10).expand([1, 1]),
'group_t_covar_module.raw_outputscale':
pyro.distributions.Normal(-7, 1),
'unit_t_covar_module.base_kernel.raw_lengthscale':
pyro.distributions.Normal(30, 10).expand([1, 1]),
'unit_t_covar_module.raw_outputscale':
pyro.distributions.Normal(-7, 1),
'likelihood.noise_covar.raw_noise':
pyro.distributions.Normal(-7, 1).expand([1]),
'x_covar_module.0.raw_c2':
pyro.distributions.Normal(-7, 1).expand([1]),
'x_covar_module.1.raw_c2':
pyro.distributions.Normal(-7, 1).expand([1])
#'model.x_covar_module.2.raw_c2': pyro.distributions.Normal(-6, 1).expand([1])
}
# plot_pyro_prior(priors, transforms)
def pyro_model(x, y):
fn = pyro.random_module("model", model, prior=priors)
sampled_model = fn()
output = sampled_model.likelihood(sampled_model(x))
pyro.sample("obs", output, obs=y)
if INFERENCE == 'MCMCLOAD':
with open('results/localnews_MCMC.pkl', 'rb') as f:
mcmc_run = pickle.load(f)
mcmc_samples = mcmc_run.get_samples()
print(mcmc_run.summary())
plot_pyro_posterior(mcmc_samples, transforms)
# plot_posterior(mcmc_samples)
return
for k, d in mcmc_samples.items():
mcmc_samples[k] = d[idx]
model.pyro_load_from_samples(mcmc_samples)
visualize_localnews_MCMC(data, train_x, train_y, train_i, test_x, test_y, test_i, model,\
likelihood, T0, station_le, 10)
return
elif INFERENCE == 'MAP':
model.group_index_module._set_rho(0.0)
model.group_t_covar_module.outputscale = 0.05**2
model.group_t_covar_module.base_kernel.lengthscale = 15
model.likelihood.noise_covar.noise = 0.05**2
model.unit_t_covar_module.outputscale = 0.05**2
model.unit_t_covar_module.base_kernel.lengthscale = 30
# weekday/day/unit effects initialize to 0.01**2
for i in range(len(X_max_v)):
model.x_covar_module[i].c2 = torch.tensor(0.05**2)
for name, param in model.drift_t_module.named_parameters():
param.requires_grad = True
model.drift_t_module._set_T1(0.0)
model.drift_t_module._set_T2(5.0)
model.drift_t_module.base_kernel.lengthscale = 30.0
model.drift_t_module.outputscale = 0.05**2
# model.drift_t_module.raw_T1.requires_grad = False
# model.drift_t_module.raw_T2.requires_grad = False
optimizer = torch.optim.LBFGS(model.parameters(),
lr=0.1,
history_size=10,
max_iter=4)
model, likelihood = train(test_x, test_y, model, likelihood, mll,
optimizer, training_iterations)
torch.save(model.state_dict(),
'results/localnews_' + INFERENCE + '_model_state.pth')
return
elif INFERENCE == 'MCMC':
model.group_index_module._set_rho(0.9)
model.group_t_covar_module.outputscale = 0.02**2
model.group_t_covar_module.base_kernel._set_lengthscale(3)
model.likelihood.noise_covar.noise = 0.03**2
model.unit_t_covar_module.outputscale = 0.02**2
model.unit_t_covar_module.base_kernel._set_lengthscale(30)
# weekday/day/unit effects initialize to 0.0**2
for i in range(len(X_max_v) - 1):
model.x_covar_module[i].c2 = torch.tensor(0.01**2)
# model.x_covar_module[i].raw_c2.requires_grad = False
initial_params = {'group_index_module.rho_prior': model.group_index_module.raw_rho.detach(),\
'group_t_covar_module.base_kernel.lengthscale_prior': model.group_t_covar_module.base_kernel.raw_lengthscale.detach(),\
'group_t_covar_module.outputscale_prior': model.group_t_covar_module.raw_outputscale.detach(),\
'unit_t_covar_module.base_kernel.lengthscale_prior': model.unit_t_covar_module.base_kernel.raw_lengthscale.detach(),\
'unit_t_covar_module.outputscale_prior': model.unit_t_covar_module.raw_outputscale.detach(),\
'likelihood.noise_covar.noise_prior': model.likelihood.raw_noise.detach(),
'x_covar_module.0.c2_prior': model.x_covar_module[0].raw_c2.detach(),
'x_covar_module.1.c2_prior': model.x_covar_module[1].raw_c2.detach()}
with gpytorch.settings.fast_computations(
covar_root_decomposition=False, log_prob=False, solves=False):
nuts_kernel = NUTS(pyro_model, adapt_step_size=True, adapt_mass_matrix=True, jit_compile=False,\
init_strategy=pyro.infer.autoguide.initialization.init_to_value(values=initial_params))
hmc_kernel = HMC(pyro_model, step_size=0.1, num_steps=10, adapt_step_size=True,\
init_strategy=pyro.infer.autoguide.initialization.init_to_mean())
mcmc_run = MCMC(nuts_kernel,
num_samples=num_samples,
warmup_steps=warmup_steps)
mcmc_run.run(train_x, train_y)
pickle.dump(mcmc_run, open("results/localnews_MCMC.pkl", "wb"))
# plot_pyro_posterior(mcmc_run.get_samples(), transforms)
return
visualize_localnews_MCMC(data, train_x, train_y, train_g, test_x, test_y, test_i, model,\
likelihood, T0, station_le, num_samples)
else:
model.load_strict_shapes(False)
state_dict = torch.load('results/localnews_MAP_model_state.pth')
model.load_state_dict(state_dict)
model2.load_state_dict(state_dict)
print(
f'Parameter name: rho value = {model.group_index_module.rho.detach().numpy()}'
)
# print(f'Parameter name: unit mean value = {model.unit_mean_module.constant.detach().numpy()}')
print(
f'Parameter name: group ls value = {model.group_t_covar_module.base_kernel.lengthscale.detach().numpy()}'
)
print(
f'Parameter name: group os value = {np.sqrt(model.group_t_covar_module.outputscale.detach().numpy())}'
)
print(
f'Parameter name: unit ls value = {model.unit_t_covar_module.base_kernel.lengthscale.detach().numpy()}'
)
print(
f'Parameter name: unit os value = {np.sqrt(model.unit_t_covar_module.outputscale.detach().numpy())}'
)
print(
f'Parameter name: noise value = {np.sqrt(model.likelihood.noise.detach().numpy())}'
)
print(
f'Parameter name: weekday std value = {np.sqrt(model.x_covar_module[0].c2.detach().numpy())}'
)
print(
f'Parameter name: day std value = {np.sqrt(model.x_covar_module[1].c2.detach().numpy())}'
)
print(
f'Parameter name: unit std value = {np.sqrt(model.x_covar_module[2].c2.detach().numpy())}'
)
print(
f'Parameter name: drift ls value = {model.drift_t_module.base_kernel.lengthscale.detach().numpy()}'
)
print(
f'Parameter name: drift cov os value = {np.sqrt(model.drift_t_module.outputscale.detach().numpy())}'
)
print(
f'Parameter name: drift cov T1 value = {model.drift_t_module.T1.detach().numpy()}'
)
print(
f'Parameter name: drift cov T2 value = {model.drift_t_module.T2.detach().numpy()}'
)
visualize_localnews(data, test_x, test_y, test_g, model, model2,
likelihood, T0, station_le, train_condition)
def main(inference, model, width, n_samples, warmup, init_method, burnin, skip,
metrics_skip, cycles, temperature, momentum, precond_update, lr,
batch_size, load_samples, save_samples, reject_samples, run_id,
log_dir, sampling_decay, progressbar, skip_first, _run, _log):
assert inference in ["SGLD", "HMC", "VerletSGLD", "OurHMC", "HMCReject", "VerletSGLDReject", "SGLDReject"]
assert width > 0
assert n_samples > 0
assert cycles > 0
assert temperature >= 0
data = get_data()
x_train = data.norm.train_X
y_train = data.norm.train_y
x_test = data.norm.test_X
y_test = data.norm.test_y
model = get_model(x_train=x_train, y_train=y_train)
if load_samples is None:
if init_method == "he":
exp_utils.he_initialize(model)
elif init_method == "he_uniform":
exp_utils.he_uniform_initialize(model)
elif init_method == "he_zerobias":
exp_utils.he_zerobias_initialize(model)
elif init_method == "prior":
pass
else:
raise ValueError(f"unknown init_method={init_method}")
else:
state_dict = exp_utils.load_samples(load_samples, idx=-1, keep_steps=False)
model_sd = model.state_dict()
for k in state_dict.keys():
if k not in model_sd:
_log.warning(f"key {k} not in model, ignoring")
del state_dict[k]
elif model_sd[k].size() != state_dict[k].size():
_log.warning(f"key {k} size mismatch, model={model_sd[k].size()}, loaded={state_dict[k].size()}")
state_dict[k] = model_sd[k]
missing_keys = set(model_sd.keys()) - set(state_dict.keys())
_log.warning(f"The following keys were not found in loaded state dict: {missing_keys}")
model_sd.update(state_dict)
model.load_state_dict(model_sd)
del state_dict
del model_sd
if save_samples:
model_saver_fn = (lambda: exp_utils.HDF5ModelSaver(
exp_utils.sneaky_artifact(_run, "samples.pt"), "w"))
else:
@contextlib.contextmanager
def model_saver_fn():
yield None
with exp_utils.HDF5Metrics(
exp_utils.sneaky_artifact(_run, "metrics.h5"), "w") as metrics_saver,\
model_saver_fn() as model_saver:
if inference == "HMC":
_potential_fn = model.get_potential(x_train, y_train, eff_num_data=len(x_train))
kernel = HMC(potential_fn=_potential_fn,
adapt_step_size=False, adapt_mass_matrix=False,
step_size=1e-3, num_steps=32)
mcmc = MCMC(kernel, num_samples=n_samples, warmup_steps=warmup, initial_params=model.params_dict())
else:
if inference == "SGLD":
runner_class = bnn_priors.inference.SGLDRunner
elif inference == "VerletSGLD":
runner_class = bnn_priors.inference.VerletSGLDRunner
elif inference == "OurHMC":
runner_class = bnn_priors.inference.HMCRunner
elif inference == "VerletSGLDReject":
runner_class = bnn_priors.inference_reject.VerletSGLDRunnerReject
elif inference == "HMCReject":
runner_class = bnn_priors.inference_reject.HMCRunnerReject
elif inference == "SGLDReject":
runner_class = bnn_priors.inference_reject.SGLDRunnerReject
assert (n_samples * skip) % cycles == 0
sample_epochs = n_samples * skip // cycles
epochs_per_cycle = warmup + burnin + sample_epochs
if batch_size is None:
batch_size = len(data.norm.train)
# Disable parallel loading for `TensorDataset`s.
num_workers = (0 if isinstance(data.norm.train, t.utils.data.TensorDataset) else 2)
dataloader = t.utils.data.DataLoader(data.norm.train, batch_size=batch_size, shuffle=True, drop_last=False, num_workers=num_workers)
dataloader_test = t.utils.data.DataLoader(data.norm.test, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=num_workers)
mcmc = runner_class(model=model, dataloader=dataloader, dataloader_test=dataloader_test, epochs_per_cycle=epochs_per_cycle,
warmup_epochs=warmup, sample_epochs=sample_epochs, learning_rate=lr,
skip=skip, metrics_skip=metrics_skip, sampling_decay=sampling_decay, cycles=cycles, temperature=temperature,
momentum=momentum, precond_update=precond_update,
metrics_saver=metrics_saver, model_saver=model_saver, reject_samples=reject_samples)
mcmc.run(progressbar=progressbar)
samples = mcmc.get_samples()
samples = {k: v[skip_first:] for k, v in samples.items()}
model.eval()
batch_size = min(batch_size, len(data.norm.test))
dataloader_test = t.utils.data.DataLoader(data.norm.test, batch_size=batch_size)
return evaluate_model(model, dataloader_test, samples)
with pyro.plate('latent_response', N):
eta = pyro.sample('eta', dist.Normal(0, 1))
# Latent function.
f = compute_f(alpha, rho, beta, eta, X)
with pyro.plate('response', N):
pyro.sample('obs', dist.Bernoulli(logits=f), obs=y)
# HMC
pyro.clear_param_store() # clear global parameter cache.
pyro.set_rng_seed(2) # set random number generator seed.
hmc = MCMC(HMC(gpc,
step_size=0.05,
trajectory_length=1,
adapt_step_size=False,
adapt_mass_matrix=False,
jit_compile=True),
num_samples=500,
warmup_steps=500) # sampler setup.
hmc.run(X, y.double()) # run mcmc
hmc_posterior_samples = hmc.get_samples() # get posterior samples.
# NUTS
pyro.clear_param_store()
pyro.set_rng_seed(2)
nuts = MCMC(NUTS(gpc,
target_accept_prob=0.8,
max_tree_depth=10,
jit_compile=True),
num_samples=500,
y_hidden_dist = dist.Exponential(1 / link[i])
if truncation_label[i] == 0:
y_real = pyro.sample("obs_{}".format(i),
y_hidden_dist,
obs = y[i])
else:
truncation_prob = 1 - y_hidden_dist.cdf(y[i])
pyro.sample("truncation_label_{}".format(i),
dist.Bernoulli(truncation_prob),
obs = truncation_label[i])
pyro.clear_param_store()
hmc_kernel = HMC(model,
step_size = 0.1,
num_steps = 4)
mcmc_run = MCMC(hmc_kernel,
num_samples=5,
warmup_steps=1).run(x, y, truncation_label)
marginal_a = EmpiricalMarginal(mcmc_run,
sites="a_model")
posterior_a = [marginal_a.sample() for i in range(50)]
sns.distplot(posterior_a)
"""# Modeling using HMC with Vectorized Data