def _build_session_groups(self):
"""Returns a list of SessionGroups protobuffers from the summary
data."""
# Algorithm: We keep a dict 'groups_by_name' mapping a SessionGroup name
# (str) to a SessionGroup protobuffer. We traverse the runs associated with
# the plugin--each representing a single session. We form a Session
# protobuffer from each run and add it to the relevant SessionGroup object
# in the 'groups_by_name' dict. We create the SessionGroup object, if this
# is the first session of that group we encounter.
groups_by_name = {}
run_to_tag_to_content = self._context.hparams_metadata(
self._experiment_id,
run_tag_filter=provider.RunTagFilter(tags=[
metadata.SESSION_START_INFO_TAG,
metadata.SESSION_END_INFO_TAG,
]),
)
# The TensorBoard runs with session start info are the
# "sessions", which are not necessarily the runs that actually
# contain metrics (may be in subdirectories).
session_names = [
run for (run, tags) in run_to_tag_to_content.items()
if metadata.SESSION_START_INFO_TAG in tags
]
metric_runs = set()
metric_tags = set()
for session_name in session_names:
for metric in self._experiment.metric_infos:
metric_name = metric.name
(run, tag) = metrics.run_tag_from_session_and_metric(
session_name, metric_name)
metric_runs.add(run)
metric_tags.add(tag)
all_metric_evals = self._context.read_last_scalars(
self._experiment_id,
run_tag_filter=provider.RunTagFilter(runs=metric_runs,
tags=metric_tags),
)
for (session_name, tag_to_content) in run_to_tag_to_content.items():
if metadata.SESSION_START_INFO_TAG not in tag_to_content:
continue
start_info = metadata.parse_session_start_info_plugin_data(
tag_to_content[metadata.SESSION_START_INFO_TAG])
end_info = None
if metadata.SESSION_END_INFO_TAG in tag_to_content:
end_info = metadata.parse_session_end_info_plugin_data(
tag_to_content[metadata.SESSION_END_INFO_TAG])
session = self._build_session(session_name, start_info, end_info,
all_metric_evals)
if session.status in self._request.allowed_statuses:
self._add_session(session, start_info, groups_by_name)
# Compute the session group's aggregated metrics for each group.
groups = groups_by_name.values()
for group in groups:
# We sort the sessions in a group so that the order is deterministic.
group.sessions.sort(key=operator.attrgetter("name"))
self._aggregate_metrics(group)
return groups
def run(self):
"""Executes the request.
Returns:
An array of tuples representing the metric evaluations--each of the
form (<wall time in secs>, <training step>, <metric value>).
"""
run, tag = metrics.run_tag_from_session_and_metric(
self._request.session_name, self._request.metric_name)
body, _ = self._scalars_plugin_instance.scalars_impl(
tag, run, self._experiment, scalars_plugin.OutputFormat.JSON)
return body
def _build_session_metric_values(self, session_name, all_metric_evals):
"""Builds the session metric values."""
# result is a list of api_pb2.MetricValue instances.
result = []
metric_infos = self._experiment.metric_infos
for metric_info in metric_infos:
metric_name = metric_info.name
(run, tag) = metrics.run_tag_from_session_and_metric(
session_name, metric_name)
datum = all_metric_evals.get(run, {}).get(tag)
if not datum:
# It's ok if we don't find the metric in the session.
# We skip it here. For filtering and sorting purposes its value is None.
continue
result.append(
api_pb2.MetricValue(
name=metric_name,
wall_time_secs=datum.wall_time,
training_step=datum.step,
value=datum.value,
))
return result
