def loss(self, model, guide, *args, **kwargs):
"""
:returns: returns an estimate of the ELBO
:rtype: float
Evaluates the ELBO with an estimator that uses num_particles many samples/particles.
"""
elbo = 0.0
for weight, model_trace, guide_trace, log_r in self._get_traces(
model, guide, *args, **kwargs):
elbo_particle = weight * 0
log_pdf = "batch_log_pdf" if (
self.enum_discrete and weight.size(0) > 1) else "log_pdf"
for name in model_trace.nodes.keys():
if model_trace.nodes[name]["type"] == "sample":
if model_trace.nodes[name]["is_observed"]:
elbo_particle += model_trace.nodes[name][log_pdf]
else:
elbo_particle += model_trace.nodes[name][log_pdf]
elbo_particle -= guide_trace.nodes[name][log_pdf]
# drop terms of weight zero to avoid nans
if isinstance(weight, numbers.Number):
if weight == 0.0:
elbo_particle = torch_zeros_like(elbo_particle)
else:
elbo_particle[weight == 0] = 0.0
elbo += torch_data_sum(weight * elbo_particle)
loss = -elbo
return loss
def _compute_elbo_reparam(model_trace, guide_trace, non_reparam_nodes):
# prepare a list of all the cost nodes, each of which is +- log_pdf
cost_nodes = []
for name, model_site in model_trace.nodes.items():
if model_site["type"] == "sample":
if model_site["is_observed"]:
cost_nodes.append(CostNode(model_site["log_pdf"], True))
else:
# cost node from model sample
cost_nodes.append(CostNode(model_site["log_pdf"], True))
# cost node from guide sample
guide_site = guide_trace.nodes[name]
zero_expectation = name in non_reparam_nodes
cost_nodes.append(CostNode(-guide_site["log_pdf"], not zero_expectation))
# compute the elbo; if all stochastic nodes are reparameterizable, we're done
# this bit is never differentiated: it's here for getting an estimate of the elbo itself
elbo = torch_data_sum(sum(c.cost for c in cost_nodes))
# compute the surrogate elbo, removing terms whose gradient is zero
# this is the bit that's actually differentiated
# XXX should the user be able to control if these terms are included?
surrogate_elbo = sum(c.cost for c in cost_nodes if c.nonzero_expectation)
return elbo, surrogate_elbo
def loss(self, model, guide, *args, **kwargs):
"""
:returns: returns an estimate of the ELBO
:rtype: float
Evaluates the ELBO with an estimator that uses num_particles many samples/particles.
"""
elbo = 0.0
for weight, model_trace, guide_trace in self._get_traces(model, guide, *args, **kwargs):
guide_trace.log_pdf(), model_trace.log_pdf()
elbo_particle = 0.0
for name in model_trace.nodes.keys():
if model_trace.nodes[name]["type"] == "sample":
if model_trace.nodes[name]["is_observed"]:
elbo_particle += model_trace.nodes[name]["log_pdf"]
else:
elbo_particle += model_trace.nodes[name]["log_pdf"]
elbo_particle -= guide_trace.nodes[name]["log_pdf"]
elbo += torch_data_sum(weight * elbo_particle)
loss = -elbo
if np.isnan(loss):
warnings.warn('Encountered NAN loss')
return loss
def _loss_and_grads_particle(self, weight, model_trace, guide_trace):
# get info regarding rao-blackwellization of vectorized map_data
guide_vec_md_info = guide_trace.graph["vectorized_map_data_info"]
model_vec_md_info = model_trace.graph["vectorized_map_data_info"]
guide_vec_md_condition = guide_vec_md_info['rao-blackwellization-condition']
model_vec_md_condition = model_vec_md_info['rao-blackwellization-condition']
do_vec_rb = guide_vec_md_condition and model_vec_md_condition
if not do_vec_rb:
warnings.warn(
"Unable to do fully-vectorized Rao-Blackwellization in TraceGraph_ELBO. "
"Falling back to higher-variance gradient estimator. "
"Try to avoid these issues in your model and guide:\n{}".format("\n".join(
guide_vec_md_info["warnings"] | model_vec_md_info["warnings"])))
guide_vec_md_nodes = guide_vec_md_info['nodes'] if do_vec_rb else set()
model_vec_md_nodes = model_vec_md_info['nodes'] if do_vec_rb else set()
# have the trace compute all the individual (batch) log pdf terms
# so that they are available below
guide_trace.compute_batch_log_pdf(site_filter=lambda name, site: name in guide_vec_md_nodes)
guide_trace.log_pdf()
model_trace.compute_batch_log_pdf(site_filter=lambda name, site: name in model_vec_md_nodes)
model_trace.log_pdf()
# prepare a list of all the cost nodes, each of which is +- log_pdf
cost_nodes = []
non_reparam_nodes = set(guide_trace.nonreparam_stochastic_nodes)
for name, model_site in model_trace.nodes.items():
if model_site["type"] == "sample":
if model_site["is_observed"]:
cost_nodes.append(CostNode(model_site["log_pdf"], True))
else:
# cost node from model sample
cost_nodes.append(CostNode(model_site["log_pdf"], True))
# cost node from guide sample
guide_site = guide_trace.nodes[name]
zero_expectation = name in non_reparam_nodes
cost_nodes.append(CostNode(-guide_site["log_pdf"], not zero_expectation))
# compute the elbo; if all stochastic nodes are reparameterizable, we're done
# this bit is never differentiated: it's here for getting an estimate of the elbo itself
elbo = torch_data_sum(sum(c.cost for c in cost_nodes))
# compute the surrogate elbo, removing terms whose gradient is zero
# this is the bit that's actually differentiated
# XXX should the user be able to control if these terms are included?
surrogate_elbo = sum(c.cost for c in cost_nodes if c.nonzero_expectation)
# the following computations are only necessary if we have non-reparameterizable nodes
baseline_loss = 0.0
if non_reparam_nodes:
# recursively compute downstream cost nodes for all sample sites in model and guide
# (even though ultimately just need for non-reparameterizable sample sites)
# 1. downstream costs used for rao-blackwellization
# 2. model observe sites (as well as terms that arise from the model and guide having different
# dependency structures) are taken care of via 'children_in_model' below
topo_sort_guide_nodes = list(reversed(list(networkx.topological_sort(guide_trace))))
topo_sort_guide_nodes = [x for x in topo_sort_guide_nodes
if guide_trace.nodes[x]["type"] == "sample"]
downstream_guide_cost_nodes = {}
downstream_costs = {}
for node in topo_sort_guide_nodes:
node_log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
downstream_costs[node] = model_trace.nodes[node][node_log_pdf_key] - \
guide_trace.nodes[node][node_log_pdf_key]
nodes_included_in_sum = set([node])
downstream_guide_cost_nodes[node] = set([node])
for child in guide_trace.successors(node):
child_cost_nodes = downstream_guide_cost_nodes[child]
downstream_guide_cost_nodes[node].update(child_cost_nodes)
if nodes_included_in_sum.isdisjoint(child_cost_nodes): # avoid duplicates
if node_log_pdf_key == 'log_pdf':
downstream_costs[node] += downstream_costs[child].sum()
else:
downstream_costs[node] += downstream_costs[child]
nodes_included_in_sum.update(child_cost_nodes)
missing_downstream_costs = downstream_guide_cost_nodes[node] - nodes_included_in_sum
# include terms we missed because we had to avoid duplicates
for missing_node in missing_downstream_costs:
mn_log_pdf_key = 'batch_log_pdf' if missing_node in guide_vec_md_nodes else 'log_pdf'
if node_log_pdf_key == 'log_pdf':
downstream_costs[node] += (model_trace.nodes[missing_node][mn_log_pdf_key] -
guide_trace.nodes[missing_node][mn_log_pdf_key]).sum()
else:
downstream_costs[node] += model_trace.nodes[missing_node][mn_log_pdf_key] - \
guide_trace.nodes[missing_node][mn_log_pdf_key]
# finish assembling complete downstream costs
# (the above computation may be missing terms from model)
# XXX can we cache some of the sums over children_in_model to make things more efficient?
for site in non_reparam_nodes:
children_in_model = set()
for node in downstream_guide_cost_nodes[site]:
children_in_model.update(model_trace.successors(node))
# remove terms accounted for above
children_in_model.difference_update(downstream_guide_cost_nodes[site])
for child in children_in_model:
child_log_pdf_key = 'batch_log_pdf' if child in model_vec_md_nodes else 'log_pdf'
site_log_pdf_key = 'batch_log_pdf' if site in guide_vec_md_nodes else 'log_pdf'
assert (model_trace.nodes[child]["type"] == "sample")
if site_log_pdf_key == 'log_pdf':
downstream_costs[site] += model_trace.nodes[child][child_log_pdf_key].sum()
else:
downstream_costs[site] += model_trace.nodes[child][child_log_pdf_key]
# construct all the reinforce-like terms.
# we include only downstream costs to reduce variance
# optionally include baselines to further reduce variance
# XXX should the average baseline be in the param store as below?
elbo_reinforce_terms = 0.0
for node in non_reparam_nodes:
guide_site = guide_trace.nodes[node]
log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
downstream_cost = downstream_costs[node]
baseline = 0.0
(nn_baseline, nn_baseline_input, use_decaying_avg_baseline, baseline_beta,
baseline_value) = _get_baseline_options(guide_site)
use_nn_baseline = nn_baseline is not None
use_baseline_value = baseline_value is not None
assert(not (use_nn_baseline and use_baseline_value)), \
"cannot use baseline_value and nn_baseline simultaneously"
if use_decaying_avg_baseline:
avg_downstream_cost_old = pyro.param("__baseline_avg_downstream_cost_" + node,
ng_zeros(1), tags="__tracegraph_elbo_internal_tag")
avg_downstream_cost_new = (1 - baseline_beta) * downstream_cost + \
baseline_beta * avg_downstream_cost_old
avg_downstream_cost_old.data = avg_downstream_cost_new.data # XXX copy_() ?
baseline += avg_downstream_cost_old
if use_nn_baseline:
# block nn_baseline_input gradients except in baseline loss
baseline += nn_baseline(detach_iterable(nn_baseline_input))
elif use_baseline_value:
# it's on the user to make sure baseline_value tape only points to baseline params
baseline += baseline_value
if use_nn_baseline or use_baseline_value:
# accumulate baseline loss
baseline_loss += torch.pow(downstream_cost.detach() - baseline, 2.0).sum()
guide_log_pdf = guide_site[log_pdf_key] / guide_site["scale"] # not scaled by subsampling
if use_nn_baseline or use_decaying_avg_baseline or use_baseline_value:
if downstream_cost.size() != baseline.size():
raise ValueError("Expected baseline at site {} to be {} instead got {}".format(
node, downstream_cost.size(), baseline.size()))
downstream_cost = downstream_cost - baseline
elbo_reinforce_terms += (guide_log_pdf * downstream_cost.detach()).sum()
surrogate_elbo += elbo_reinforce_terms
# collect parameters to train from model and guide
trainable_params = set(site["value"]
for trace in (model_trace, guide_trace)
for site in trace.nodes.values()
if site["type"] == "param")
if trainable_params:
surrogate_loss = -surrogate_elbo
torch_backward(weight * (surrogate_loss + baseline_loss))
pyro.get_param_store().mark_params_active(trainable_params)
loss = -elbo
return weight * loss
def loss_and_grads(self, model, guide, *args, **kwargs):
"""
:returns: returns an estimate of the ELBO
:rtype: float
Computes the ELBO as well as the surrogate ELBO that is used to form the gradient estimator.
Performs backward on the latter. Num_particle many samples are used to form the estimators.
"""
elbo = 0.0
# grab a trace from the generator
for weight, model_trace, guide_trace, log_r in self._get_traces(
model, guide, *args, **kwargs):
elbo_particle = weight * 0
surrogate_elbo_particle = weight * 0
# compute elbo and surrogate elbo
log_pdf = "batch_log_pdf" if (
self.enum_discrete and weight.size(0) > 1) else "log_pdf"
for name in model_trace.nodes.keys():
if model_trace.nodes[name]["type"] == "sample":
if model_trace.nodes[name]["is_observed"]:
elbo_particle += model_trace.nodes[name][log_pdf]
surrogate_elbo_particle += model_trace.nodes[name][
log_pdf]
else:
lp_lq = model_trace.nodes[name][
log_pdf] - guide_trace.nodes[name][log_pdf]
elbo_particle += lp_lq
if guide_trace.nodes[name]["fn"].reparameterized:
surrogate_elbo_particle += lp_lq
else:
# XXX should the user be able to control inclusion of the -logq term below?
surrogate_elbo_particle += model_trace.nodes[name][log_pdf] + \
log_r.detach() * guide_trace.nodes[name][log_pdf]
# drop terms of weight zero to avoid nans
if isinstance(weight, numbers.Number):
if weight == 0.0:
elbo_particle = torch_zeros_like(elbo_particle)
surrogate_elbo_particle = torch_zeros_like(
surrogate_elbo_particle)
else:
weight_eq_zero = (weight == 0)
elbo_particle[weight_eq_zero] = 0.0
surrogate_elbo_particle[weight_eq_zero] = 0.0
elbo += torch_data_sum(weight * elbo_particle)
surrogate_elbo_particle = torch_sum(weight *
surrogate_elbo_particle)
# collect parameters to train from model and guide
trainable_params = set(site["value"]
for trace in (model_trace, guide_trace)
for site in trace.nodes.values()
if site["type"] == "param")
if trainable_params:
surrogate_loss_particle = -surrogate_elbo_particle
torch_backward(surrogate_loss_particle)
pyro.get_param_store().mark_params_active(trainable_params)
loss = -elbo
return loss
def loss_and_grads(self, model, guide, *args, **kwargs):
"""
:returns: returns an estimate of the ELBO
:rtype: float
Computes the ELBO as well as the surrogate ELBO that is used to form the gradient estimator.
Performs backward on the latter. Num_particle many samples are used to form the estimators.
If baselines are present, a baseline loss is also constructed and differentiated.
"""
elbo = 0.0
for weight, model_trace, guide_trace in self._get_traces(
model, guide, *args, **kwargs):
# get info regarding rao-blackwellization of vectorized map_data
guide_vec_md_info = guide_trace.graph["vectorized_map_data_info"]
model_vec_md_info = model_trace.graph["vectorized_map_data_info"]
guide_vec_md_condition = guide_vec_md_info[
'rao-blackwellization-condition']
model_vec_md_condition = model_vec_md_info[
'rao-blackwellization-condition']
do_vec_rb = guide_vec_md_condition and model_vec_md_condition
if not do_vec_rb:
warnings.warn(
"Unable to do fully-vectorized Rao-Blackwellization in TraceGraph_ELBO. "
"Falling back to higher-variance gradient estimator. "
"Try to avoid these issues in your model and guide:\n{}".
format("\n".join(guide_vec_md_info["warnings"]
| model_vec_md_info["warnings"])))
guide_vec_md_nodes = guide_vec_md_info[
'nodes'] if do_vec_rb else set()
model_vec_md_nodes = model_vec_md_info[
'nodes'] if do_vec_rb else set()
# have the trace compute all the individual (batch) log pdf terms
# so that they are available below
guide_trace.compute_batch_log_pdf(
site_filter=lambda name, site: name in guide_vec_md_nodes)
guide_trace.log_pdf()
model_trace.compute_batch_log_pdf(
site_filter=lambda name, site: name in model_vec_md_nodes)
model_trace.log_pdf()
# prepare a list of all the cost nodes, each of which is +- log_pdf
cost_nodes = []
non_reparam_nodes = set(guide_trace.nonreparam_stochastic_nodes)
for site in model_trace.nodes.keys():
model_trace_site = model_trace.nodes[site]
log_pdf_key = 'batch_log_pdf' if site in model_vec_md_nodes else 'log_pdf'
if model_trace_site["type"] == "sample":
if model_trace_site["is_observed"]:
cost_node = (model_trace_site[log_pdf_key], True)
cost_nodes.append(cost_node)
else:
# cost node from model sample
cost_node1 = (model_trace_site[log_pdf_key], True)
# cost node from guide sample
zero_expectation = site in non_reparam_nodes
cost_node2 = (-guide_trace.nodes[site][log_pdf_key],
not zero_expectation)
cost_nodes.extend([cost_node1, cost_node2])
elbo_particle = 0.0
surrogate_elbo_particle = 0.0
baseline_loss_particle = 0.0
elbo_reinforce_terms_particle = 0.0
elbo_no_zero_expectation_terms_particle = 0.0
# compute the elbo; if all stochastic nodes are reparameterizable, we're done
# this bit is never differentiated: it's here for getting an estimate of the elbo itself
for cost_node in cost_nodes:
elbo_particle += cost_node[0].sum()
elbo += weight * torch_data_sum(elbo_particle)
# compute the elbo, removing terms whose gradient is zero
# this is the bit that's actually differentiated
# XXX should the user be able to control if these terms are included?
for cost_node in cost_nodes:
if cost_node[1]:
elbo_no_zero_expectation_terms_particle += cost_node[
0].sum()
surrogate_elbo_particle += weight * elbo_no_zero_expectation_terms_particle
# the following computations are only necessary if we have non-reparameterizable nodes
if len(non_reparam_nodes) > 0:
# recursively compute downstream cost nodes for all sample sites in model and guide
# (even though ultimately just need for non-reparameterizable sample sites)
# 1. downstream costs used for rao-blackwellization
# 2. model observe sites (as well as terms that arise from the model and guide having different
# dependency structures) are taken care of via 'children_in_model' below
topo_sort_guide_nodes = list(
reversed(list(networkx.topological_sort(guide_trace))))
topo_sort_guide_nodes = [
x for x in topo_sort_guide_nodes
if guide_trace.nodes[x]["type"] == "sample"
]
downstream_guide_cost_nodes = {}
downstream_costs = {}
for node in topo_sort_guide_nodes:
node_log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
downstream_costs[node] = model_trace.nodes[node][node_log_pdf_key] - \
guide_trace.nodes[node][node_log_pdf_key]
nodes_included_in_sum = set([node])
downstream_guide_cost_nodes[node] = set([node])
for child in guide_trace.successors(node):
child_cost_nodes = downstream_guide_cost_nodes[child]
downstream_guide_cost_nodes[node].update(
child_cost_nodes)
if nodes_included_in_sum.isdisjoint(
child_cost_nodes): # avoid duplicates
if node_log_pdf_key == 'log_pdf':
downstream_costs[node] += downstream_costs[
child].sum()
else:
downstream_costs[node] += downstream_costs[
child]
nodes_included_in_sum.update(child_cost_nodes)
missing_downstream_costs = downstream_guide_cost_nodes[
node] - nodes_included_in_sum
# include terms we missed because we had to avoid duplicates
for missing_node in missing_downstream_costs:
mn_log_pdf_key = 'batch_log_pdf' if missing_node in guide_vec_md_nodes else 'log_pdf'
if node_log_pdf_key == 'log_pdf':
downstream_costs[node] += (
model_trace.nodes[missing_node][mn_log_pdf_key]
-
guide_trace.nodes[missing_node][mn_log_pdf_key]
).sum()
else:
downstream_costs[node] += model_trace.nodes[missing_node][mn_log_pdf_key] - \
guide_trace.nodes[missing_node][mn_log_pdf_key]
# finish assembling complete downstream costs
# (the above computation may be missing terms from model)
# XXX can we cache some of the sums over children_in_model to make things more efficient?
for site in non_reparam_nodes:
children_in_model = set()
for node in downstream_guide_cost_nodes[site]:
children_in_model.update(model_trace.successors(node))
# remove terms accounted for above
children_in_model.difference_update(
downstream_guide_cost_nodes[site])
for child in children_in_model:
child_log_pdf_key = 'batch_log_pdf' if child in model_vec_md_nodes else 'log_pdf'
site_log_pdf_key = 'batch_log_pdf' if site in guide_vec_md_nodes else 'log_pdf'
assert (model_trace.nodes[child]["type"] == "sample")
if site_log_pdf_key == 'log_pdf':
downstream_costs[site] += model_trace.nodes[child][
child_log_pdf_key].sum()
else:
downstream_costs[site] += model_trace.nodes[child][
child_log_pdf_key]
# construct all the reinforce-like terms.
# we include only downstream costs to reduce variance
# optionally include baselines to further reduce variance
# XXX should the average baseline be in the param store as below?
# for extracting baseline options from site["baseline"]
# XXX default for baseline_beta currently set here
def get_baseline_options(site_baseline):
options_dict = site_baseline.copy()
options_tuple = (options_dict.pop('nn_baseline', None),
options_dict.pop('nn_baseline_input',
None),
options_dict.pop(
'use_decaying_avg_baseline', False),
options_dict.pop('baseline_beta', 0.90),
options_dict.pop('baseline_value', None))
if options_dict:
raise ValueError(
"Unrecognized baseline options: {}".format(
options_dict.keys()))
return options_tuple
baseline_loss_particle = 0.0
for node in non_reparam_nodes:
log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
downstream_cost = downstream_costs[node]
baseline = 0.0
(nn_baseline, nn_baseline_input, use_decaying_avg_baseline,
baseline_beta, baseline_value) = get_baseline_options(
guide_trace.nodes[node]["baseline"])
use_nn_baseline = nn_baseline is not None
use_baseline_value = baseline_value is not None
assert(not (use_nn_baseline and use_baseline_value)), \
"cannot use baseline_value and nn_baseline simultaneously"
if use_decaying_avg_baseline:
avg_downstream_cost_old = pyro.param(
"__baseline_avg_downstream_cost_" + node,
ng_zeros(1),
tags="__tracegraph_elbo_internal_tag")
avg_downstream_cost_new = (1 - baseline_beta) * downstream_cost + \
baseline_beta * avg_downstream_cost_old
avg_downstream_cost_old.data = avg_downstream_cost_new.data # XXX copy_() ?
baseline += avg_downstream_cost_old
if use_nn_baseline:
# block nn_baseline_input gradients except in baseline loss
baseline += nn_baseline(
detach_iterable(nn_baseline_input))
elif use_baseline_value:
# it's on the user to make sure baseline_value tape only points to baseline params
baseline += baseline_value
if use_nn_baseline or use_baseline_value:
# construct baseline loss
baseline_loss = torch.pow(
downstream_cost.detach() - baseline, 2.0).sum()
baseline_loss_particle += weight * baseline_loss
if use_nn_baseline or use_decaying_avg_baseline or use_baseline_value:
if downstream_cost.size() != baseline.size():
raise ValueError(
"Expected baseline at site {} to be {} instead got {}"
.format(node, downstream_cost.size(),
baseline.size()))
elbo_reinforce_terms_particle += (
guide_trace.nodes[node][log_pdf_key] *
(downstream_cost - baseline).detach()).sum()
else:
elbo_reinforce_terms_particle += (
guide_trace.nodes[node][log_pdf_key] *
downstream_cost.detach()).sum()
surrogate_elbo_particle += weight * elbo_reinforce_terms_particle
torch_backward(baseline_loss_particle)
# collect parameters to train from model and guide
trainable_params = set(site["value"]
for trace in (model_trace, guide_trace)
for site in trace.nodes.values()
if site["type"] == "param")
surrogate_loss_particle = -surrogate_elbo_particle
if trainable_params:
torch_backward(surrogate_loss_particle)
# mark all params seen in trace as active so that gradient steps are taken downstream
pyro.get_param_store().mark_params_active(trainable_params)
loss = -elbo
return loss
def _loss_and_grads_particle(self, weight, model_trace, guide_trace):
# get info regarding rao-blackwellization of vectorized map_data
guide_vec_md_info = guide_trace.graph["vectorized_map_data_info"]
model_vec_md_info = model_trace.graph["vectorized_map_data_info"]
guide_vec_md_condition = guide_vec_md_info[
'rao-blackwellization-condition']
model_vec_md_condition = model_vec_md_info[
'rao-blackwellization-condition']
do_vec_rb = guide_vec_md_condition and model_vec_md_condition
if not do_vec_rb:
warnings.warn(
"Unable to do fully-vectorized Rao-Blackwellization in TraceGraph_ELBO. "
"Falling back to higher-variance gradient estimator. "
"Try to avoid these issues in your model and guide:\n{}".
format("\n".join(guide_vec_md_info["warnings"]
| model_vec_md_info["warnings"])))
guide_vec_md_nodes = guide_vec_md_info['nodes'] if do_vec_rb else set()
model_vec_md_nodes = model_vec_md_info['nodes'] if do_vec_rb else set()
# have the trace compute all the individual (batch) log pdf terms
# so that they are available below
guide_trace.compute_batch_log_pdf(
site_filter=lambda name, site: name in guide_vec_md_nodes)
guide_trace.log_pdf()
model_trace.compute_batch_log_pdf(
site_filter=lambda name, site: name in model_vec_md_nodes)
model_trace.log_pdf()
# prepare a list of all the cost nodes, each of which is +- log_pdf
cost_nodes = []
non_reparam_nodes = set(guide_trace.nonreparam_stochastic_nodes)
for name, model_site in model_trace.nodes.items():
if model_site["type"] == "sample":
if model_site["is_observed"]:
cost_nodes.append(CostNode(model_site["log_pdf"], True))
else:
# cost node from model sample
cost_nodes.append(CostNode(model_site["log_pdf"], True))
# cost node from guide sample
guide_site = guide_trace.nodes[name]
zero_expectation = name in non_reparam_nodes
cost_nodes.append(
CostNode(-guide_site["log_pdf"], not zero_expectation))
# compute the elbo; if all stochastic nodes are reparameterizable, we're done
# this bit is never differentiated: it's here for getting an estimate of the elbo itself
elbo = torch_data_sum(sum(c.cost for c in cost_nodes))
# compute the surrogate elbo, removing terms whose gradient is zero
# this is the bit that's actually differentiated
# XXX should the user be able to control if these terms are included?
surrogate_elbo = sum(c.cost for c in cost_nodes
if c.nonzero_expectation)
# the following computations are only necessary if we have non-reparameterizable nodes
baseline_loss = 0.0
if non_reparam_nodes:
# recursively compute downstream cost nodes for all sample sites in model and guide
# (even though ultimately just need for non-reparameterizable sample sites)
# 1. downstream costs used for rao-blackwellization
# 2. model observe sites (as well as terms that arise from the model and guide having different
# dependency structures) are taken care of via 'children_in_model' below
topo_sort_guide_nodes = list(
reversed(list(networkx.topological_sort(guide_trace))))
topo_sort_guide_nodes = [
x for x in topo_sort_guide_nodes
if guide_trace.nodes[x]["type"] == "sample"
]
downstream_guide_cost_nodes = {}
downstream_costs = {}
for node in topo_sort_guide_nodes:
node_log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
downstream_costs[node] = model_trace.nodes[node][node_log_pdf_key] - \
guide_trace.nodes[node][node_log_pdf_key]
nodes_included_in_sum = set([node])
downstream_guide_cost_nodes[node] = set([node])
for child in guide_trace.successors(node):
child_cost_nodes = downstream_guide_cost_nodes[child]
downstream_guide_cost_nodes[node].update(child_cost_nodes)
if nodes_included_in_sum.isdisjoint(
child_cost_nodes): # avoid duplicates
if node_log_pdf_key == 'log_pdf':
downstream_costs[node] += downstream_costs[
child].sum()
else:
downstream_costs[node] += downstream_costs[child]
nodes_included_in_sum.update(child_cost_nodes)
missing_downstream_costs = downstream_guide_cost_nodes[
node] - nodes_included_in_sum
# include terms we missed because we had to avoid duplicates
for missing_node in missing_downstream_costs:
mn_log_pdf_key = 'batch_log_pdf' if missing_node in guide_vec_md_nodes else 'log_pdf'
if node_log_pdf_key == 'log_pdf':
downstream_costs[node] += (
model_trace.nodes[missing_node][mn_log_pdf_key] -
guide_trace.nodes[missing_node][mn_log_pdf_key]
).sum()
else:
downstream_costs[node] += model_trace.nodes[missing_node][mn_log_pdf_key] - \
guide_trace.nodes[missing_node][mn_log_pdf_key]
# finish assembling complete downstream costs
# (the above computation may be missing terms from model)
# XXX can we cache some of the sums over children_in_model to make things more efficient?
for site in non_reparam_nodes:
children_in_model = set()
for node in downstream_guide_cost_nodes[site]:
children_in_model.update(model_trace.successors(node))
# remove terms accounted for above
children_in_model.difference_update(
downstream_guide_cost_nodes[site])
for child in children_in_model:
child_log_pdf_key = 'batch_log_pdf' if child in model_vec_md_nodes else 'log_pdf'
site_log_pdf_key = 'batch_log_pdf' if site in guide_vec_md_nodes else 'log_pdf'
assert (model_trace.nodes[child]["type"] == "sample")
if site_log_pdf_key == 'log_pdf':
downstream_costs[site] += model_trace.nodes[child][
child_log_pdf_key].sum()
else:
downstream_costs[site] += model_trace.nodes[child][
child_log_pdf_key]
# construct all the reinforce-like terms.
# we include only downstream costs to reduce variance
# optionally include baselines to further reduce variance
# XXX should the average baseline be in the param store as below?
elbo_reinforce_terms = 0.0
for node in non_reparam_nodes:
guide_site = guide_trace.nodes[node]
log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
downstream_cost = downstream_costs[node]
baseline = 0.0
(nn_baseline, nn_baseline_input, use_decaying_avg_baseline,
baseline_beta,
baseline_value) = _get_baseline_options(guide_site)
use_nn_baseline = nn_baseline is not None
use_baseline_value = baseline_value is not None
assert(not (use_nn_baseline and use_baseline_value)), \
"cannot use baseline_value and nn_baseline simultaneously"
if use_decaying_avg_baseline:
avg_downstream_cost_old = pyro.param(
"__baseline_avg_downstream_cost_" + node,
ng_zeros(1),
tags="__tracegraph_elbo_internal_tag")
avg_downstream_cost_new = (1 - baseline_beta) * downstream_cost + \
baseline_beta * avg_downstream_cost_old
avg_downstream_cost_old.data = avg_downstream_cost_new.data # XXX copy_() ?
baseline += avg_downstream_cost_old
if use_nn_baseline:
# block nn_baseline_input gradients except in baseline loss
baseline += nn_baseline(detach_iterable(nn_baseline_input))
elif use_baseline_value:
# it's on the user to make sure baseline_value tape only points to baseline params
baseline += baseline_value
if use_nn_baseline or use_baseline_value:
# accumulate baseline loss
baseline_loss += torch.pow(
downstream_cost.detach() - baseline, 2.0).sum()
guide_log_pdf = guide_site[log_pdf_key] / guide_site[
"scale"] # not scaled by subsampling
if use_nn_baseline or use_decaying_avg_baseline or use_baseline_value:
if downstream_cost.size() != baseline.size():
raise ValueError(
"Expected baseline at site {} to be {} instead got {}"
.format(node, downstream_cost.size(),
baseline.size()))
downstream_cost = downstream_cost - baseline
elbo_reinforce_terms += (guide_log_pdf *
downstream_cost.detach()).sum()
surrogate_elbo += elbo_reinforce_terms
# collect parameters to train from model and guide
trainable_params = set(site["value"]
for trace in (model_trace, guide_trace)
for site in trace.nodes.values()
if site["type"] == "param")
if trainable_params:
surrogate_loss = -surrogate_elbo
torch_backward(weight * (surrogate_loss + baseline_loss))
pyro.get_param_store().mark_params_active(trainable_params)
loss = -elbo
return weight * loss