def MetaGraph(self):
"""Return the metagraph definition, if there is one.
Raises:
ValueError: If there is no metagraph for this run.
Returns:
The `meta_graph_def` proto.
"""
if self._meta_graph is None:
raise ValueError('There is no metagraph in this EventAccumulator')
meta_graph = tf.MetaGraphDef()
meta_graph.ParseFromString(self._meta_graph)
return meta_graph
def _generate_test_data(self, run_name, experiment_name):
"""Generates the test data directory.
The test data has a single run of the given name, containing:
- a graph definition and metagraph definition
Arguments:
run_name: The directory under self.logdir into which to write
events.
"""
run_path = os.path.join(self.logdir, run_name)
writer = tf.summary.FileWriter(run_path)
# Add a simple graph event.
graph_def = tf.GraphDef()
node1 = graph_def.node.add()
node1.name = 'a'
node2 = graph_def.node.add()
node2.name = 'b'
node2.attr['very_large_attr'].s = b'a' * 2048 # 2 KB attribute
meta_graph_def = tf.MetaGraphDef(graph_def=graph_def)
if self._only_use_meta_graph:
writer.add_meta_graph(meta_graph_def)
else:
writer.add_graph(graph_def)
writer.flush()
writer.close()
# Write data for the run to the database.
# TODO(nickfelt): Figure out why reseting the graph is necessary.
tf.reset_default_graph()
db_writer = tf.contrib.summary.create_db_writer(
db_uri=self.db_path,
experiment_name=experiment_name,
run_name=run_name,
user_name='user')
with db_writer.as_default(), tf.contrib.summary.always_record_summaries():
tf.contrib.summary.scalar('mytag', 1)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.contrib.summary.summary_writer_initializer_op())
sess.run(tf.contrib.summary.all_summary_ops())
def _ProcessEvent(self, event):
"""Called whenever an event is loaded."""
if self._first_event_timestamp is None:
self._first_event_timestamp = event.wall_time
if event.HasField('file_version'):
new_file_version = _ParseFileVersion(event.file_version)
if self.file_version and self.file_version != new_file_version:
## This should not happen.
tf.logging.warn(('Found new file_version for event.proto. This will '
'affect purging logic for TensorFlow restarts. '
'Old: {0} New: {1}').format(self.file_version,
new_file_version))
self.file_version = new_file_version
self._MaybePurgeOrphanedData(event)
## Process the event.
# GraphDef and MetaGraphDef are handled in a special way:
# If no graph_def Event is available, but a meta_graph_def is, and it
# contains a graph_def, then use the meta_graph_def.graph_def as our graph.
# If a graph_def Event is available, always prefer it to the graph_def
# inside the meta_graph_def.
if event.HasField('graph_def'):
if self._graph is not None:
tf.logging.warn(
('Found more than one graph event per run, or there was '
'a metagraph containing a graph_def, as well as one or '
'more graph events. Overwriting the graph with the '
'newest event.'))
self._graph = event.graph_def
self._graph_from_metagraph = False
elif event.HasField('meta_graph_def'):
if self._meta_graph is not None:
tf.logging.warn(('Found more than one metagraph event per run. '
'Overwriting the metagraph with the newest event.'))
self._meta_graph = event.meta_graph_def
if self._graph is None or self._graph_from_metagraph:
# We may have a graph_def in the metagraph. If so, and no
# graph_def is directly available, use this one instead.
meta_graph = tf.MetaGraphDef()
meta_graph.ParseFromString(self._meta_graph)
if meta_graph.graph_def:
if self._graph is not None:
tf.logging.warn(
('Found multiple metagraphs containing graph_defs,'
'but did not find any graph events. Overwriting the '
'graph with the newest metagraph version.'))
self._graph_from_metagraph = True
self._graph = meta_graph.graph_def.SerializeToString()
elif event.HasField('tagged_run_metadata'):
tag = event.tagged_run_metadata.tag
if tag in self._tagged_metadata:
tf.logging.warn('Found more than one "run metadata" event with tag ' +
tag + '. Overwriting it with the newest event.')
self._tagged_metadata[tag] = event.tagged_run_metadata.run_metadata
elif event.HasField('summary'):
for value in event.summary.value:
value = data_compat.migrate_value(value)
if value.HasField('metadata'):
tag = value.tag
# We only store the first instance of the metadata. This check
# is important: the `FileWriter` does strip metadata from all
# values except the first one per each tag, but a new
# `FileWriter` is created every time a training job stops and
# restarts. Hence, we must also ignore non-initial metadata in
# this logic.
if tag not in self.summary_metadata:
self.summary_metadata[tag] = value.metadata
plugin_data = value.metadata.plugin_data
if plugin_data.plugin_name:
with self._plugin_tag_locks[plugin_data.plugin_name]:
self._plugin_to_tag_to_content[plugin_data.plugin_name][tag] = (
plugin_data.content)
else:
tf.logging.warn(
('This summary with tag %r is oddly not associated with a '
'plugin.'), tag)
for summary_type, summary_func in SUMMARY_TYPES.items():
if value.HasField(summary_type):
datum = getattr(value, summary_type)
tag = value.tag
if summary_type == 'tensor' and not tag:
# This tensor summary was created using the old method that used
# plugin assets. We must still continue to support it.
tag = value.node_name
getattr(self, summary_func)(tag, event.wall_time, event.step, datum)