def fit_with_hmc(model,
observed_time_series,
num_results=100,
num_warmup_steps=50,
num_leapfrog_steps=15,
initial_state=None,
initial_step_size=None,
chain_batch_shape=(),
num_variational_steps=150,
variational_optimizer=None,
seed=None,
name=None):
"""Draw posterior samples using Hamiltonian Monte Carlo (HMC).
Markov chain Monte Carlo (MCMC) methods are considered the gold standard of
Bayesian inference; under suitable conditions and in the limit of infinitely
many draws they generate samples from the true posterior distribution. HMC [1]
uses gradients of the model's log-density function to propose samples,
allowing it to exploit posterior geometry. However, it is computationally more
expensive than variational inference and relatively sensitive to tuning.
This method attempts to provide a sensible default approach for fitting
StructuralTimeSeries models using HMC. It first runs variational inference as
a fast posterior approximation, and initializes the HMC sampler from the
variational posterior, using the posterior standard deviations to set
per-variable step sizes (equivalently, a diagonal mass matrix). During the
warmup phase, it adapts the step size to target an acceptance rate of 0.75,
which is thought to be in the desirable range for optimal mixing [2].
Args:
model: An instance of `StructuralTimeSeries` representing a
time-series model. This represents a joint distribution over
time-series and their parameters with batch shape `[b1, ..., bN]`.
observed_time_series: `float` `Tensor` of shape
`concat([sample_shape, model.batch_shape, [num_timesteps, 1]]) where
`sample_shape` corresponds to i.i.d. observations, and the trailing `[1]`
dimension may (optionally) be omitted if `num_timesteps > 1`. May
optionally be an instance of `tfp.sts.MaskedTimeSeries`, which includes
a mask `Tensor` to specify timesteps with missing observations.
num_results: Integer number of Markov chain draws.
Default value: `100`.
num_warmup_steps: Integer number of steps to take before starting to
collect results. The warmup steps are also used to adapt the step size
towards a target acceptance rate of 0.75.
Default value: `50`.
num_leapfrog_steps: Integer number of steps to run the leapfrog integrator
for. Total progress per HMC step is roughly proportional to
`step_size * num_leapfrog_steps`.
Default value: `15`.
initial_state: Optional Python `list` of `Tensor`s, one for each model
parameter, representing the initial state(s) of the Markov chain(s). These
should have shape `concat([chain_batch_shape, param.prior.batch_shape,
param.prior.event_shape])`. If `None`, the initial state is set
automatically using a sample from a variational posterior.
Default value: `None`.
initial_step_size: Python `list` of `Tensor`s, one for each model parameter,
representing the step size for the leapfrog integrator. Must
broadcast with the shape of `initial_state`. Larger step sizes lead to
faster progress, but too-large step sizes make rejection exponentially
more likely. If `None`, the step size is set automatically using the
standard deviation of a variational posterior.
Default value: `None`.
chain_batch_shape: Batch shape (Python `tuple`, `list`, or `int`) of chains
to run in parallel.
Default value: `[]` (i.e., a single chain).
num_variational_steps: Python `int` number of steps to run the variational
optimization to determine the initial state and step sizes.
Default value: `200`.
variational_optimizer: Optional `tf.train.Optimizer` instance to use in
the variational optimization. If `None`, defaults to
`tf.train.AdamOptimizer(0.1)`.
Default value: `None`.
seed: Python integer to seed the random number generator.
name: Python `str` name prefixed to ops created by this function.
Default value: `None` (i.e., 'fit_with_hmc').
Returns:
samples: Python `list` of `Tensors` representing posterior samples of model
parameters, with shapes `[concat([[num_results], chain_batch_shape,
param.prior.batch_shape, param.prior.event_shape]) for param in
model.parameters]`.
kernel_results: A (possibly nested) `tuple`, `namedtuple` or `list` of
`Tensor`s representing internal calculations made within the HMC sampler.
#### Examples
Assume we've built a structural time-series model:
```python
day_of_week = tfp.sts.Seasonal(
num_seasons=7,
observed_time_series=observed_time_series,
name='day_of_week')
local_linear_trend = tfp.sts.LocalLinearTrend(
observed_time_series=observed_time_series,
name='local_linear_trend')
model = tfp.sts.Sum(components=[day_of_week, local_linear_trend],
observed_time_series=observed_time_series)
```
To draw posterior samples using HMC under default settings:
```python
samples, kernel_results = tfp.sts.fit_with_hmc(model, observed_time_series)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
samples_, kernel_results_ = sess.run((samples, kernel_results))
print("acceptance rate: {}".format(
np.mean(kernel_results_.inner_results.is_accepted, axis=0)))
print("posterior means: {}".format(
{param.name: np.mean(param_draws, axis=0)
for (param, param_draws) in zip(model.parameters, samples_)}))
```
We can also run multiple chains. This may help diagnose convergence issues
and allows us to exploit vectorization to draw samples more quickly, although
warmup still requires the same number of sequential steps.
```python
from matplotlib import pylab as plt
samples, kernel_results = tfp.sts.fit_with_hmc(
model, observed_time_series, chain_batch_shape=[10])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
samples_, kernel_results_ = sess.run((samples, kernel_results))
print("acceptance rate: {}".format(
np.mean(kernel_results_.inner_results.is_accepted, axis=0)))
# Plot the sampled traces for each parameter. If the chains have mixed, their
# traces should all cover the same region of state space, frequently crossing
# over each other.
for (param, param_draws) in zip(model.parameters, samples_):
if param.prior.event_shape.ndims > 0:
print("Only plotting traces for scalar parameters, skipping {}".format(
param.name))
continue
plt.figure(figsize=[10, 4])
plt.title(param.name)
plt.plot(param_draws)
plt.ylabel(param.name)
plt.xlabel("HMC step")
# Combining the samples from multiple chains into a single dimension allows
# us to easily pass sampled parameters to downstream forecasting methods.
combined_samples_ = [np.reshape(param_draws,
[-1] + list(param_draws.shape[2:]))
for param_draws in samples_]
```
For greater flexibility, you may prefer to implement your own sampler using
the TensorFlow Probability primitives in `tfp.mcmc`. The following recipe
constructs a basic HMC sampler, using a `TransformedTransitionKernel` to
incorporate constraints on the parameter space.
```python
transformed_hmc_kernel = mcmc.TransformedTransitionKernel(
inner_kernel=mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=model.joint_log_prob(observed_time_series),
step_size=step_size,
num_leapfrog_steps=num_leapfrog_steps,
step_size_update_fn=tfp.mcmc.make_simple_step_size_update_policy(
num_adaptation_steps=num_adaptation_steps),
state_gradients_are_stopped=True,
seed=seed),
bijector=[param.bijector for param in model.parameters])
# Initialize from a Uniform[-2, 2] distribution in unconstrained space.
initial_state = [tfp.sts.sample_uniform_initial_state(
param, return_constrained=True) for param in model.parameters]
samples, kernel_results = tfp.mcmc.sample_chain(
kernel=transformed_hmc_kernel,
num_results=num_results,
current_state=initial_state,
num_burnin_steps=num_warmup_steps)
```
#### References
[1]: Radford Neal. MCMC Using Hamiltonian Dynamics. _Handbook of Markov Chain
Monte Carlo_, 2011. https://arxiv.org/abs/1206.1901
[2] M.J. Betancourt, Simon Byrne, and Mark Girolami. Optimizing The
Integrator Step Size for Hamiltonian Monte Carlo.
https://arxiv.org/abs/1411.6669
"""
with tf.compat.v1.name_scope(
name, 'fit_with_hmc', values=[observed_time_series]) as name:
seed = tfd.SeedStream(seed, salt='StructuralTimeSeries_fit_with_hmc')
# Initialize state and step sizes from a variational posterior if not
# specified.
if initial_step_size is None or initial_state is None:
# To avoid threading variational distributions through the training
# while loop, we build our own copy here. `make_template` ensures
# that our variational distributions share the optimized parameters.
def make_variational():
return build_factored_variational_loss(
model, observed_time_series,
init_batch_shape=chain_batch_shape, seed=seed())
make_variational = tf.compat.v1.make_template('make_variational',
make_variational)
_, variational_distributions = make_variational()
minimize_op = _minimize_in_graph(
build_loss_fn=lambda: make_variational()[0], # return just the loss.
num_steps=num_variational_steps,
optimizer=variational_optimizer)
with tf.control_dependencies([minimize_op]):
if initial_state is None:
initial_state = [tf.stop_gradient(d.sample())
for d in variational_distributions.values()]
# Set step sizes using the unconstrained variational distribution.
if initial_step_size is None:
initial_step_size = [
transformed_q.distribution.stddev()
for transformed_q in variational_distributions.values()]
# Multiple chains manifest as an extra param batch dimension, so we need to
# add a corresponding batch dimension to `observed_time_series`.
observed_time_series = sts_util.pad_batch_dimension_for_multiple_chains(
observed_time_series, model, chain_batch_shape=chain_batch_shape)
# When the initial step size depends on a variational optimization, we
# can't initialize step size variables before the optimization runs.
# Instead we initialize with a dummy value of the appropriate
# shape, then wrap the HMC chain with `control_dependencies` to ensure the
# variational step sizes are assigned before HMC actually runs.
step_size = [
tf.compat.v1.get_variable(
initializer=tf.zeros_like(
sample_uniform_initial_state(
param,
init_sample_shape=chain_batch_shape,
return_constrained=False)),
name='{}_step_size'.format(param.name),
trainable=False,
use_resource=True)
for (param, ss) in zip(model.parameters, initial_step_size)
]
step_size_init_op = tf.group([
tf.compat.v1.assign(ss, initial_ss)
for (ss, initial_ss) in zip(step_size, initial_step_size)
])
# Run HMC to sample from the posterior on parameters.
with tf.control_dependencies([step_size_init_op]):
samples, kernel_results = mcmc.sample_chain(
num_results=num_results,
current_state=initial_state,
num_burnin_steps=num_warmup_steps,
kernel=mcmc.TransformedTransitionKernel(
inner_kernel=mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=model.joint_log_prob(observed_time_series),
step_size=step_size,
num_leapfrog_steps=num_leapfrog_steps,
step_size_update_fn=mcmc.make_simple_step_size_update_policy(
num_adaptation_steps=int(num_warmup_steps * 0.8),
decrement_multiplier=0.1,
increment_multiplier=0.1),
state_gradients_are_stopped=True,
seed=seed()),
bijector=[param.bijector for param in model.parameters]),
parallel_iterations=1 if seed is not None else 10)
return samples, kernel_results
def run_vip_hmc_continuous(model_config,
num_samples=2000,
burnin=1000,
use_iaf_posterior=False,
num_leapfrog_steps=4,
num_adaptation_steps=500,
num_optimization_steps=2000,
num_mc_samples=32,
tau=1.,
do_sample=True,
description='',
experiments_dir=''):
tf.reset_default_graph()
if use_iaf_posterior:
# IAF posterior doesn't give us stddevs for step sizes for HMC (we could
# extract them by sampling but I haven't implemented that), and we mostly
# care about it for ELBOs anyway.
do_sample = False
init_val_loc = tf.placeholder('float', shape=())
init_val_scale = tf.placeholder('float', shape=())
(learnable_parameters,
learnable_parametrisation, _) = ed_transforms.make_learnable_parametrisation(
init_val_loc=init_val_loc, init_val_scale=init_val_scale, tau=tau)
def model_vip(*params):
with ed.interception(learnable_parametrisation):
return model_config.model(*params)
log_joint_vip = ed.make_log_joint_fn(model_vip)
with ed.tape() as model_tape:
_ = model_vip(*model_config.model_args)
param_shapes = collections.OrderedDict()
target_vip_kwargs = {}
for param in model_tape.keys():
if param not in model_config.observed_data.keys():
param_shapes[param] = model_tape[param].shape
else:
target_vip_kwargs[param] = model_config.observed_data[param]
def target_vip(*param_args):
i = 0
for param in model_tape.keys():
if param not in model_config.observed_data.keys():
target_vip_kwargs[param] = param_args[i]
i = i + 1
return log_joint_vip(*model_config.model_args, **target_vip_kwargs)
full_kwargs = collections.OrderedDict(model_config.observed_data.items())
full_kwargs['parameterisation'] = collections.OrderedDict()
for k in learnable_parameters.keys():
full_kwargs['parameterisation'][k] = learnable_parameters[k]
if use_iaf_posterior:
elbo = util.get_iaf_elbo(
target_vip,
num_mc_samples=num_mc_samples,
param_shapes=param_shapes)
variational_parameters = {}
else:
elbo, variational_parameters = util.get_mean_field_elbo(
model_vip,
target_vip,
num_mc_samples=num_mc_samples,
model_args=model_config.model_args,
vi_kwargs=full_kwargs)
vip_step_size_approx = util.get_approximate_step_size(
variational_parameters, num_leapfrog_steps)
##############################################################################
best_elbo = None
model_dir = os.path.join(experiments_dir,
str(description + '_' + model_config.model.__name__))
if not tf.gfile.Exists(model_dir):
tf.gfile.MakeDirs(model_dir)
saver = tf.train.Saver()
dir_save = os.path.join(model_dir, 'saved_params_{}'.format(gen_id()))
if not tf.gfile.Exists(dir_save):
tf.gfile.MakeDirs(dir_save)
best_lr = None
best_init_loc = None
best_init_scale = None
learning_rate_ph = tf.placeholder(shape=[], dtype=tf.float32)
learning_rate = tf.Variable(learning_rate_ph, trainable=False)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train = optimizer.minimize(-elbo)
init = tf.global_variables_initializer()
learning_rates = [0.003, 0.01, 0.01, 0.1, 0.003, 0.01]
if use_iaf_posterior:
learning_rates = [3e-5, 1e-4, 3e-4, 1e-4]
start_time = time.time()
for learning_rate_val in learning_rates:
for init_loc in [0.]: #, 10., -10.]:
for init_scale in [init_loc]:
timeline = []
with tf.Session() as sess:
init.run(feed_dict={init_val_loc: init_loc,
init_val_scale: init_scale,
learning_rate_ph: learning_rate_val})
this_timeline = []
for i in range(num_optimization_steps):
_, e = sess.run([train, elbo])
if np.isnan(e):
util.print('got NaN in ELBO optimization, stopping...')
break
this_timeline.append(e)
this_elbo = np.mean(this_timeline[-100:])
info_str = ('finished cVIP optimization with elbo {} vs '
'best ELBO {}'.format(this_elbo, best_elbo))
util.print(info_str)
if best_elbo is None or best_elbo < this_elbo:
best_elbo = this_elbo
timeline = this_timeline
vals = sess.run(list(learnable_parameters.values()))
learned_reparam = collections.OrderedDict(
zip(learnable_parameters.keys(), vals))
vals = sess.run(list(variational_parameters.values()))
learned_variational_params = collections.OrderedDict(
zip(variational_parameters.keys(), vals))
util.print('learned params {}'.format(learned_reparam))
util.print('learned variational params {}'.format(
learned_variational_params))
_ = saver.save(sess, dir_save)
best_lr = learning_rate
best_init_loc = init_loc
best_init_scale = init_scale
vi_time = time.time() - start_time
util.print('BEST: LR={}, init={}, {}'.format(best_lr, best_init_loc,
best_init_scale))
util.print('ELBO: {}'.format(best_elbo))
to_centered = model_config.make_to_centered(**learned_reparam)
results = collections.OrderedDict()
results['elbo'] = best_elbo
with tf.Session() as sess:
saver.restore(sess, dir_save)
results['vp'] = learned_variational_params
if do_sample:
vip_step_size_init = sess.run(vip_step_size_approx)
vip_step_size = [tf.get_variable(
name='step_size_vip'+str(i),
initializer=np.array(vip_step_size_init[i], dtype=np.float32),
use_resource=True, # For TFE compatibility.
trainable=False) for i in range(len(vip_step_size_init))]
kernel_vip = mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=target_vip,
step_size=vip_step_size,
num_leapfrog_steps=num_leapfrog_steps,
step_size_update_fn=mcmc.make_simple_step_size_update_policy(
num_adaptation_steps=num_adaptation_steps, target_rate=0.85))
states, kernel_results_vip = mcmc.sample_chain(
num_results=num_samples,
num_burnin_steps=burnin,
current_state=[
tf.zeros(param_shapes[param]) for param in param_shapes.keys()
],
kernel=kernel_vip,
num_steps_between_results=1)
states_vip = transform_mcmc_states(states, to_centered)
init_again = tf.global_variables_initializer()
init_again.run(feed_dict={
init_val_loc: best_init_loc, init_val_scale: best_init_scale,
learning_rate_ph: 1.0}) # learning rate doesn't matter for HMC.
ess_vip = tfp.mcmc.effective_sample_size(states_vip)
start_time = time.time()
samples, is_accepted, ess, ss_vip, log_accept_ratio = sess.run(
(states_vip, kernel_results_vip.is_accepted, ess_vip,
kernel_results_vip.extra.step_size_assign,
kernel_results_vip.log_accept_ratio))
sampling_time = time.time() - start_time
results['samples'] = collections.OrderedDict()
results['is_accepted'] = is_accepted
results['acceptance_rate'] = np.sum(is_accepted) * 100. / float(
num_samples)
results['ess'] = ess
results['sampling_time'] = sampling_time
results['log_accept_ratio'] = log_accept_ratio
results['step_size'] = [s[0] for s in ss_vip]
i = 0
for param in param_shapes.keys():
results['samples'][param] = samples[i]
i = i + 1
# end if
results['parameterisation'] = collections.OrderedDict()
i = 0
for param in param_shapes.keys():
name_a = param[:-5] + 'a'
name_b = param[:-5] + 'b'
try:
results['parameterisation'][name_a] = learned_reparam[name_a]
results['parameterisation'][name_b] = learned_reparam[name_b]
except KeyError:
continue
i = i + 1
results['elbo_timeline'] = timeline
results['vi_time'] = vi_time
results['init_pos'] = best_init_loc
return results
def _run_hmc(target, param_shapes, transform=noop, step_size_init=0.1,
num_samples=2000, burnin=1000, num_adaptation_steps=500,
num_leapfrog_steps=4):
g = tf.Graph()
with g.as_default():
step_size = [tf.get_variable(
name='step_size'+str(i),
initializer=np.array(step_size_init[i], dtype=np.float32),
use_resource=True, # For TFE compatibility.
trainable=False) for i in range(len(step_size_init))]
kernel = mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=target,
step_size=step_size,
num_leapfrog_steps=num_leapfrog_steps,
step_size_update_fn=mcmc.make_simple_step_size_update_policy(
num_adaptation_steps=num_adaptation_steps, target_rate=0.85))
states, kernel_results = mcmc.sample_chain(
num_results=num_samples,
num_burnin_steps=burnin,
current_state=[
tf.zeros(param_shapes[param]) for param in param_shapes.keys()
],
kernel=kernel,
num_steps_between_results=1)
tr_states = transform_mcmc_states(states, transform)
ess_op = tfp.mcmc.effective_sample_size(tr_states)
init = tf.global_variables_initializer()
with tf.Session() as sess:
init.run()
start_time = time.time()
(orig_samples, samples,
is_accepted, ess,
final_step_size,
log_accept_ratio) = sess.run([states, tr_states,
kernel_results.is_accepted,
ess_op,
kernel_results.extra.step_size_assign,
kernel_results.log_accept_ratio])
sampling_time = time.time() - start_time
results = collections.OrderedDict()
results['samples'] = collections.OrderedDict()
i = 0
for param in param_shapes.keys():
results['samples'][param] = samples[i]
i = i + 1
results['orig_samples'] = orig_samples
results['is_accepted'] = is_accepted
results['acceptance_rate'] = np.sum(is_accepted) * 100. / float(num_samples)
results['ess'] = ess
results['step_size'] = [s[0] for s in final_step_size]
results['sampling_time'] = sampling_time
results['log_accept_ratio'] = log_accept_ratio
return results
def fit_with_hmc(model,
observed_time_series,
num_results=100,
num_warmup_steps=50,
num_leapfrog_steps=15,
initial_state=None,
initial_step_size=None,
chain_batch_shape=(),
num_variational_steps=150,
variational_optimizer=None,
seed=None,
name=None):
"""Draw posterior samples using Hamiltonian Monte Carlo (HMC).
Markov chain Monte Carlo (MCMC) methods are considered the gold standard of
Bayesian inference; under suitable conditions and in the limit of infinitely
many draws they generate samples from the true posterior distribution. HMC [1]
uses gradients of the model's log-density function to propose samples,
allowing it to exploit posterior geometry. However, it is computationally more
expensive than variational inference and relatively sensitive to tuning.
This method attempts to provide a sensible default approach for fitting
StructuralTimeSeries models using HMC. It first runs variational inference as
a fast posterior approximation, and initializes the HMC sampler from the
variational posterior, using the posterior standard deviations to set
per-variable step sizes (equivalently, a diagonal mass matrix). During the
warmup phase, it adapts the step size to target an acceptance rate of 0.75,
which is thought to be in the desirable range for optimal mixing [2].
Args:
model: An instance of `StructuralTimeSeries` representing a
time-series model. This represents a joint distribution over
time-series and their parameters with batch shape `[b1, ..., bN]`.
observed_time_series: `float` `Tensor` of shape
`concat([sample_shape, model.batch_shape, [num_timesteps, 1]]) where
`sample_shape` corresponds to i.i.d. observations, and the trailing `[1]`
dimension may (optionally) be omitted if `num_timesteps > 1`.
num_results: Integer number of Markov chain draws.
Default value: `100`.
num_warmup_steps: Integer number of steps to take before starting to
collect results. The warmup steps are also used to adapt the step size
towards a target acceptance rate of 0.75.
Default value: `50`.
num_leapfrog_steps: Integer number of steps to run the leapfrog integrator
for. Total progress per HMC step is roughly proportional to
`step_size * num_leapfrog_steps`.
Default value: `15`.
initial_state: Optional Python `list` of `Tensor`s, one for each model
parameter, representing the initial state(s) of the Markov chain(s). These
should have shape `concat([chain_batch_shape, param.prior.batch_shape,
param.prior.event_shape])`. If `None`, the initial state is set
automatically using a sample from a variational posterior.
Default value: `None`.
initial_step_size: Python `list` of `Tensor`s, one for each model parameter,
representing the step size for the leapfrog integrator. Must
broadcast with the shape of `initial_state`. Larger step sizes lead to
faster progress, but too-large step sizes make rejection exponentially
more likely. If `None`, the step size is set automatically using the
standard deviation of a variational posterior.
Default value: `None`.
chain_batch_shape: Batch shape (Python `tuple`, `list`, or `int`) of chains
to run in parallel.
Default value: `[]` (i.e., a single chain).
num_variational_steps: Python `int` number of steps to run the variational
optimization to determine the initial state and step sizes.
Default value: `200`.
variational_optimizer: Optional `tf.train.Optimizer` instance to use in
the variational optimization. If `None`, defaults to
`tf.train.AdamOptimizer(0.1)`.
Default value: `None`.
seed: Python integer to seed the random number generator.
name: Python `str` name prefixed to ops created by this function.
Default value: `None` (i.e., 'fit_with_hmc').
Returns:
samples: Python `list` of `Tensors` representing posterior samples of model
parameters, with shapes `[concat([[num_results], chain_batch_shape,
param.prior.batch_shape, param.prior.event_shape]) for param in
model.parameters]`.
kernel_results: A (possibly nested) `tuple`, `namedtuple` or `list` of
`Tensor`s representing internal calculations made within the HMC sampler.
#### Examples
Assume we've built a structural time-series model:
```python
day_of_week = tfp.sts.Seasonal(
num_seasons=7,
observed_time_series=observed_time_series,
name='day_of_week')
local_linear_trend = tfp.sts.LocalLinearTrend(
observed_time_series=observed_time_series,
name='local_linear_trend')
model = tfp.sts.Sum(components=[day_of_week, local_linear_trend],
observed_time_series=observed_time_series)
```
To draw posterior samples using HMC under default settings:
```python
samples, kernel_results = tfp.sts.fit_with_hmc(model, observed_time_series)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
samples_, kernel_results_ = sess.run((samples, kernel_results))
print("acceptance rate: {}".format(
np.mean(kernel_results_.inner_results.is_accepted, axis=0)))
print("posterior means: {}".format(
{param.name: np.mean(param_draws, axis=0)
for (param, param_draws) in zip(model.parameters, samples_)}))
```
We can also run multiple chains. This may help diagnose convergence issues
and allows us to exploit vectorization to draw samples more quickly, although
warmup still requires the same number of sequential steps.
```python
from matplotlib import pylab as plt
samples, kernel_results = tfp.sts.fit_with_hmc(
model, observed_time_series, chain_batch_shape=[10])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
samples_, kernel_results_ = sess.run((samples, kernel_results))
print("acceptance rate: {}".format(
np.mean(kernel_results_.inner_results.is_accepted, axis=0)))
# Plot the sampled traces for each parameter. If the chains have mixed, their
# traces should all cover the same region of state space, frequently crossing
# over each other.
for (param, param_draws) in zip(model.parameters, samples_):
if param.prior.event_shape.ndims > 0:
print("Only plotting traces for scalar parameters, skipping {}".format(
param.name))
continue
plt.figure(figsize=[10, 4])
plt.title(param.name)
plt.plot(param_draws)
plt.ylabel(param.name)
plt.xlabel("HMC step")
# Combining the samples from multiple chains into a single dimension allows
# us to easily pass sampled parameters to downstream forecasting methods.
combined_samples_ = [np.reshape(param_draws,
[-1] + list(param_draws.shape[2:]))
for param_draws in samples_]
```
For greater flexibility, you may prefer to implement your own sampler using
the TensorFlow Probability primitives in `tfp.mcmc`. The following recipe
constructs a basic HMC sampler, using a `TransformedTransitionKernel` to
incorporate constraints on the parameter space.
```python
transformed_hmc_kernel = mcmc.TransformedTransitionKernel(
inner_kernel=mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=model.joint_log_prob(observed_time_series),
step_size=step_size,
num_leapfrog_steps=num_leapfrog_steps,
step_size_update_fn=tfp.mcmc.make_simple_step_size_update_policy(
num_adaptation_steps=num_adaptation_steps),
state_gradients_are_stopped=True,
seed=seed),
bijector=[param.bijector for param in model.parameters])
# Initialize from a Uniform[-2, 2] distribution in unconstrained space.
initial_state = [tfp.sts.sample_uniform_initial_state(
param, return_constrained=True) for param in model.parameters]
samples, kernel_results = tfp.mcmc.sample_chain(
kernel=transformed_hmc_kernel,
num_results=num_results,
current_state=initial_state,
num_burnin_steps=num_warmup_steps)
```
#### References
[1]: Radford Neal. MCMC Using Hamiltonian Dynamics. _Handbook of Markov Chain
Monte Carlo_, 2011. https://arxiv.org/abs/1206.1901
[2] M.J. Betancourt, Simon Byrne, and Mark Girolami. Optimizing The
Integrator Step Size for Hamiltonian Monte Carlo.
https://arxiv.org/abs/1411.6669
"""
with tf.name_scope(name, 'fit_with_hmc',
values=[observed_time_series]) as name:
observed_time_series = tf.convert_to_tensor(observed_time_series,
name='observed_time_series')
seed = tfd.SeedStream(seed, salt='StructuralTimeSeries_fit_with_hmc')
# Initialize state and step sizes from a variational posterior if not
# specified.
if initial_step_size is None or initial_state is None:
# To avoid threading variational distributions through the training
# while loop, we build our own copy here. `make_template` ensures
# that our variational distributions share the optimized parameters.
def make_variational():
return build_factored_variational_loss(
model, observed_time_series,
init_batch_shape=chain_batch_shape, seed=seed())
make_variational = tf.make_template('make_variational', make_variational)
_, variational_distributions = make_variational()
minimize_op = _minimize_in_graph(
build_loss_fn=lambda: make_variational()[0], # return just the loss.
num_steps=num_variational_steps,
optimizer=variational_optimizer)
with tf.control_dependencies([minimize_op]):
if initial_state is None:
initial_state = [tf.stop_gradient(d.sample())
for d in variational_distributions.values()]
# Set step sizes using the unconstrained variational distribution.
if initial_step_size is None:
initial_step_size = [
transformed_q.distribution.stddev()
for transformed_q in variational_distributions.values()]
# Multiple chains manifest as an extra param batch dimension, so we need to
# add a corresponding batch dimension to `observed_time_series`.
observed_time_series = pad_batch_dimension_for_multiple_chains(
observed_time_series, model, chain_batch_shape=chain_batch_shape)
# When the initial step size depends on a variational optimization, we
# can't initialize step size variables before the optimization runs.
# Instead we initialize with a dummy value of the appropriate
# shape, then wrap the HMC chain with `control_dependencies` to ensure the
# variational step sizes are assigned before HMC actually runs.
step_size = [tf.get_variable(
initializer=tf.zeros_like(sample_uniform_initial_state(
param, init_sample_shape=chain_batch_shape,
return_constrained=False)),
name='{}_step_size'.format(param.name),
trainable=False,
use_resource=True)
for (param, ss) in zip(model.parameters, initial_step_size)]
step_size_init_op = tf.group(
[tf.assign(ss, initial_ss)
for (ss, initial_ss) in zip(step_size, initial_step_size)])
# Run HMC to sample from the posterior on parameters.
with tf.control_dependencies([step_size_init_op]):
samples, kernel_results = mcmc.sample_chain(
num_results=num_results,
current_state=initial_state,
num_burnin_steps=num_warmup_steps,
kernel=mcmc.TransformedTransitionKernel(
inner_kernel=mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=model.joint_log_prob(observed_time_series),
step_size=step_size,
num_leapfrog_steps=num_leapfrog_steps,
step_size_update_fn=mcmc.make_simple_step_size_update_policy(
num_adaptation_steps=int(num_warmup_steps * 0.8),
decrement_multiplier=0.1,
increment_multiplier=0.1),
state_gradients_are_stopped=True,
seed=seed()),
bijector=[param.bijector for param in model.parameters]),
parallel_iterations=1 if seed is not None else 10)
return samples, kernel_results
