def test_graph_def_experimental_filter_graph(self):
# Create a `GraphDef`
graph_def = graph_pb2.GraphDef()
graph_def.node.add(name="alice", op="Person")
graph_def.node.add(name="bob", op="Person")
graph_def.node[1].attr["small"].s = b"small_attr_value"
graph_def.node[1].attr["large"].s = (
b"large_attr_value" * 100 # 1600 bytes > 1024 limit
)
graph_def.node.add(name="friendship",
op="Friendship",
input=["alice", "bob"])
# Simulate legacy graph event
old_event = event_pb2.Event()
old_event.step = 0
old_event.wall_time = 456.75
old_event.graph_def = graph_def.SerializeToString()
new_events = self._migrate_event(old_event,
experimental_filter_graph=True)
new_event = new_events[1]
tensor = tensor_util.make_ndarray(new_event.summary.value[0].tensor)
new_graph_def_bytes = tensor[0]
new_graph_def = graph_pb2.GraphDef.FromString(new_graph_def_bytes)
expected_graph_def = graph_pb2.GraphDef()
expected_graph_def.CopyFrom(graph_def)
del expected_graph_def.node[1].attr["large"]
expected_graph_def.node[1].attr["_too_large_attrs"].list.s.append(
b"large")
self.assertProtoEquals(expected_graph_def, new_graph_def)
def merge_graph_defs(graph_defs):
"""Merges GraphDefs by adding unique prefix, `graph_{ind}`, to names.
All GraphDefs are expected to be of TensorBoard's.
When collecting graphs using the `tf.summary.trace` API, node names are not
guranteed to be unique. When non-unique names are not considered, it can
lead to graph visualization showing them as one which creates inaccurate
depiction of the flow of the graph (e.g., if there are A -> B -> C and D ->
B -> E, you may see {A, D} -> B -> E). To prevent such graph, we checked
for uniquenss while merging but it resulted in
https://github.com/tensorflow/tensorboard/issues/1929.
To remedy these issues, we simply "apply name scope" on each graph by
prefixing it with unique name (with a chance of collision) to create
unconnected group of graphs.
In case there is only one graph def passed, it returns the original
graph_def. In case no graph defs are passed, it returns an empty GraphDef.
Args:
graph_defs: TensorBoard GraphDefs to merge.
Returns:
TensorBoard GraphDef that merges all graph_defs with unique prefixes.
Raises:
ValueError in case GraphDef versions mismatch.
"""
if len(graph_defs) == 1:
return graph_defs[0]
elif len(graph_defs) == 0:
return graph_pb2.GraphDef()
dst_graph_def = graph_pb2.GraphDef()
if graph_defs[0].versions.producer:
dst_graph_def.versions.CopyFrom(graph_defs[0].versions)
for index, graph_def in enumerate(graph_defs):
if dst_graph_def.versions.producer != graph_def.versions.producer:
raise ValueError("Cannot combine GraphDefs of different versions.")
_add_with_prepended_names(
"graph_%d" % (index + 1),
graph_def,
dst_graph_def,
)
return dst_graph_def
def _migrate_graph_event(old_event, experimental_filter_graph=False):
result = event_pb2.Event()
result.wall_time = old_event.wall_time
result.step = old_event.step
value = result.summary.value.add(tag=graphs_metadata.RUN_GRAPH_NAME)
graph_bytes = old_event.graph_def
# TODO(@davidsoergel): Move this stopgap to a more appropriate place.
if experimental_filter_graph:
try:
graph_def = graph_pb2.GraphDef().FromString(graph_bytes)
# The reason for the RuntimeWarning catch here is b/27494216, whereby
# some proto parsers incorrectly raise that instead of DecodeError
# on certain kinds of malformed input. Triggering this seems to require
# a combination of mysterious circumstances.
except (message.DecodeError, RuntimeWarning):
logger.warning(
"Could not parse GraphDef of size %d. Skipping.",
len(graph_bytes),
)
return (old_event, )
# Use the default filter parameters:
# limit_attr_size=1024, large_attrs_key="_too_large_attrs"
process_graph.prepare_graph_for_ui(graph_def)
graph_bytes = graph_def.SerializeToString()
value.tensor.CopyFrom(tensor_util.make_tensor_proto([graph_bytes]))
value.metadata.plugin_data.plugin_name = graphs_metadata.PLUGIN_NAME
# `value.metadata.plugin_data.content` left as the empty proto
value.metadata.data_class = summary_pb2.DATA_CLASS_BLOB_SEQUENCE
# In the short term, keep both the old event and the new event to
# maintain compatibility.
return (old_event, result)
def graph_impl(
self,
run,
tag,
is_conceptual,
experiment=None,
limit_attr_size=None,
large_attrs_key=None,
):
"""Result of the form `(body, mime_type)`, or `None` if no graph
exists."""
if self._data_provider:
graph_blob_sequences = self._data_provider.read_blob_sequences(
experiment_id=experiment,
plugin_name=metadata.PLUGIN_NAME,
run_tag_filter=provider.RunTagFilter(runs=[run], tags=[tag]),
)
blob_datum_list = graph_blob_sequences.get(run, {}).get(tag, ())
try:
blob_ref = blob_datum_list[0].values[0]
except IndexError:
return None
# Always use the blob_key approach for now, even if there is a direct url.
graph_raw = self._data_provider.read_blob(blob_ref.blob_key)
# This method ultimately returns pbtxt, but we have to deserialize and
# later reserialize this anyway, because a) this way we accept binary
# protobufs too, and b) below we run `prepare_graph_for_ui` on the graph.
graph = graph_pb2.GraphDef.FromString(graph_raw)
elif is_conceptual:
tensor_events = self._multiplexer.Tensors(run, tag)
# Take the first event if there are multiple events written from different
# steps.
keras_model_config = json.loads(
tensor_events[0].tensor_proto.string_val[0]
)
graph = keras_util.keras_model_to_graph_def(keras_model_config)
elif tag:
tensor_events = self._multiplexer.Tensors(run, tag)
# Take the first event if there are multiple events written from different
# steps.
run_metadata = config_pb2.RunMetadata.FromString(
tensor_events[0].tensor_proto.string_val[0]
)
graph = graph_pb2.GraphDef()
for func_graph in run_metadata.function_graphs:
graph_util.combine_graph_defs(
graph, func_graph.pre_optimization_graph
)
else:
graph = self._multiplexer.Graph(run)
# This next line might raise a ValueError if the limit parameters
# are invalid (size is negative, size present but key absent, etc.).
process_graph.prepare_graph_for_ui(
graph, limit_attr_size, large_attrs_key
)
return (str(graph), "text/x-protobuf") # pbtxt
def graph_impl(self,
run,
tag,
is_conceptual,
limit_attr_size=None,
large_attrs_key=None):
"""Result of the form `(body, mime_type)`, or `None` if no graph exists."""
if is_conceptual:
tensor_events = self._multiplexer.Tensors(run, tag)
# Take the first event if there are multiple events written from different
# steps.
keras_model_config = json.loads(
tensor_events[0].tensor_proto.string_val[0])
graph = keras_util.keras_model_to_graph_def(keras_model_config)
elif tag:
tensor_events = self._multiplexer.Tensors(run, tag)
# Take the first event if there are multiple events written from different
# steps.
run_metadata = config_pb2.RunMetadata.FromString(
tensor_events[0].tensor_proto.string_val[0])
graph = graph_pb2.GraphDef()
for func_graph in run_metadata.function_graphs:
graph_util.combine_graph_defs(
graph, func_graph.pre_optimization_graph)
else:
graph = self._multiplexer.Graph(run)
self.graph = graph
# This next line might raise a ValueError if the limit parameters
# are invalid (size is negative, size present but key absent, etc.).
process_graph.prepare_graph_for_ui(graph, limit_attr_size,
large_attrs_key)
return (str(graph), 'text/x-protobuf') # pbtxt
def _migrate_graph_event(old_event, experimental_filter_graph=False):
result = event_pb2.Event()
result.wall_time = old_event.wall_time
result.step = old_event.step
value = result.summary.value.add(tag=graphs_metadata.RUN_GRAPH_NAME)
graph_bytes = old_event.graph_def
# TODO(@davidsoergel): Move this stopgap to a more appropriate place.
if experimental_filter_graph:
try:
graph_def = graph_pb2.GraphDef().FromString(graph_bytes)
except message.DecodeError:
logger.warning(
"Could not parse GraphDef of size %d. Skipping.",
len(graph_bytes),
)
return (old_event, )
# Use the default filter parameters:
# limit_attr_size=1024, large_attrs_key="_too_large_attrs"
process_graph.prepare_graph_for_ui(graph_def)
graph_bytes = graph_def.SerializeToString()
value.tensor.CopyFrom(tensor_util.make_tensor_proto([graph_bytes]))
value.metadata.plugin_data.plugin_name = graphs_metadata.PLUGIN_NAME
# `value.metadata.plugin_data.content` left as the empty proto
value.metadata.data_class = summary_pb2.DATA_CLASS_BLOB_SEQUENCE
# In the short term, keep both the old event and the new event to
# maintain compatibility.
return (old_event, result)
def _create_example_graph_bytes(large_attr_size):
graph_def = graph_pb2.GraphDef()
graph_def.node.add(name="alice", op="Person")
graph_def.node.add(name="bob", op="Person")
graph_def.node[1].attr["small"].s = b"small_attr_value"
graph_def.node[1].attr["large"].s = b"l" * large_attr_size
graph_def.node.add(name="friendship",
op="Friendship",
input=["alice", "bob"])
return graph_def.SerializeToString()
def test_graph_def(self):
# Create a `GraphDef` and write it to disk as an event.
logdir = self.get_temp_dir()
writer = test_util.FileWriter(logdir)
graph_def = graph_pb2.GraphDef()
graph_def.node.add(name="alice", op="Person")
graph_def.node.add(name="bob", op="Person")
graph_def.node.add(name="friendship",
op="Friendship",
input=["alice", "bob"])
writer.add_graph(graph=None, graph_def=graph_def, global_step=123)
writer.flush()
# Read in the `Event` containing the written `graph_def`.
files = os.listdir(logdir)
self.assertLen(files, 1)
event_file = os.path.join(logdir, files[0])
self.assertIn("tfevents", event_file)
loader = event_file_loader.EventFileLoader(event_file)
events = list(loader.Load())
self.assertLen(events, 2)
self.assertEqual(events[0].WhichOneof("what"), "file_version")
self.assertEqual(events[1].WhichOneof("what"), "graph_def")
old_event = events[1]
new_events = self._migrate_event(old_event)
self.assertLen(new_events, 2)
self.assertIs(new_events[0], old_event)
new_event = new_events[1]
self.assertEqual(new_event.WhichOneof("what"), "summary")
self.assertLen(new_event.summary.value, 1)
tensor = tensor_util.make_ndarray(new_event.summary.value[0].tensor)
self.assertEqual(
new_event.summary.value[0].metadata.data_class,
summary_pb2.DATA_CLASS_BLOB_SEQUENCE,
)
self.assertEqual(
new_event.summary.value[0].metadata.plugin_data.plugin_name,
graphs_metadata.PLUGIN_NAME,
)
self.assertEqual(tensor.shape, (1, ))
new_graph_def_bytes = tensor[0]
self.assertIsInstance(new_graph_def_bytes, bytes)
self.assertGreaterEqual(len(new_graph_def_bytes), 16)
new_graph_def = graph_pb2.GraphDef.FromString(new_graph_def_bytes)
self.assertProtoEquals(graph_def, new_graph_def)
def _filtered_graph_bytes(graph_bytes):
try:
graph_def = graph_pb2.GraphDef().FromString(graph_bytes)
# The reason for the RuntimeWarning catch here is b/27494216, whereby
# some proto parsers incorrectly raise that instead of DecodeError
# on certain kinds of malformed input. Triggering this seems to require
# a combination of mysterious circumstances.
except (message.DecodeError, RuntimeWarning):
logger.warning(
"Could not parse GraphDef of size %d. Skipping.", len(graph_bytes),
)
return None
# Use the default filter parameters:
# limit_attr_size=1024, large_attrs_key="_too_large_attrs"
process_graph.prepare_graph_for_ui(graph_def)
return graph_def.SerializeToString()
def Graph(self):
"""Return the graph definition, if there is one.
If the graph is stored directly, return that. If no graph is stored
directly but a metagraph is stored containing a graph, return that.
Raises:
ValueError: If there is no graph for this run.
Returns:
The `graph_def` proto.
"""
graph = graph_pb2.GraphDef()
if self._graph is not None:
graph.ParseFromString(self._graph)
return graph
raise ValueError("There is no graph in this EventAccumulator")
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)
with test_util.FileWriterCache.get(run_path) as writer:
# Add a simple graph event.
graph_def = graph_pb2.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 = meta_graph_pb2.MetaGraphDef(graph_def=graph_def)
if self._only_use_meta_graph:
writer.add_meta_graph(meta_graph_def)
else:
writer.add_graph(graph=None, graph_def=graph_def)
# Write data for the run to the database.
# TODO(nickfelt): Figure out why reseting the graph is necessary.
tf.compat.v1.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.compat.v1.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
sess.run(tf.contrib.summary.summary_writer_initializer_op())
sess.run(tf.contrib.summary.all_summary_ops())
def _get_graph(self, plugin, *args, **kwargs):
"""Fetch and return the graph as a proto."""
(graph_pbtxt, mime_type) = plugin.graph_impl(*args, **kwargs)
self.assertEqual(mime_type, "text/x-protobuf")
return text_format.Parse(graph_pbtxt, graph_pb2.GraphDef())
def init_graph(self, request):
self.graph = graph_pb2.GraphDef()
return http_util.Respond(request, 'init', 'text/plain')
def __init__(self):
self._tb_graph = graph_pb2.GraphDef()
