def _parse_event_proto_str(proto_str):
"""
Converts the serialized Event protobufs from `tf.data.TFRecordDataset` into a `Message` object.
"""
event_msg = Event()
event_msg.ParseFromString(proto_str)
return event_msg
def _add_event(self, step, summary):
t = time.time()
e = Event(wall_time=t, step=step, summary=summary)
self.queue.append(e)
if t - self.last_append >= self.tick or len(self.queue) >= self.max_queue_len:
self.flush()
self.last_append = t
def rename_events(input_path, output_path, old_tags, new_tag):
# Make a record writer
with tf.io.TFRecordWriter(str(output_path)) as writer:
# Iterate event records
for rec in tf.data.TFRecordDataset([str(input_path)]):
# Read event
ev = Event()
ev.MergeFromString(rec.numpy())
# Check if it is a summary
if ev.summary:
# Iterate summary values
for v in ev.summary.value:
# Check if the tag should be renamed
if v.tag in old_tags:
# Rename with new tag name
v.tag = new_tag
writer.write(ev.SerializeToString())
def write_summary(dpath, aggregations_per_key):
writer = tf.summary.FileWriter(dpath)
for key, (steps, wall_times, aggregations) in aggregations_per_key.items():
for step, wall_time, aggregation in zip(steps, wall_times, aggregations):
summary = tf.Summary(value=[tf.Summary.Value(tag=key, simple_value=aggregation)])
scalar_event = Event(wall_time=wall_time, step=step, summary=summary)
writer.add_event(scalar_event)
writer.flush()
def iter_log_events(tf_event_file):
import tensorflow as tf
from tensorflow.core.util.event_pb2 import Event
for event in tf.data.TFRecordDataset([tf_event_file]):
event = Event.FromString(event.numpy())
if event.summary.value:
assert len(event.summary.value
) == 1, "Unexpected length for event summary"
value = event.summary.value[0]
yield value.tag, value.simple_value
def remove_events(input_file, groups_to_remove):
new_file = input_file + ".new"
# Make a record writer
with tf.io.TFRecordWriter(new_file) as writer:
# Iterate event records
for rec in tf.data.TFRecordDataset([input_file]):
# Read event
ev = Event()
ev.MergeFromString(rec.numpy())
# Check if it is a summary event
if ev.summary:
orig_values = [v for v in ev.summary.value]
filtered_values = [v for v in orig_values if not is_tag_matching_group(v.tag, groups_to_remove)]
#print(f"filtered_values={len(filtered_values)}, orig_values={len(orig_values)}")
if len(filtered_values) != len(orig_values):
# for v in orig_values:
# print(v)
del ev.summary.value[:]
ev.summary.value.extend(filtered_values)
writer.write(ev.SerializeToString())
os.rename(new_file, input_file)
def rename_events(input_file, old_tags, new_tag):
new_file = input_file + ".new"
# Make a record writer
with tf.io.TFRecordWriter(new_file) as writer:
# Iterate event records
for rec in tf.data.TFRecordDataset([input_file]):
# Read event
ev = Event()
ev.MergeFromString(rec.numpy())
# Check if it is a summary
#print(ev)
if ev.summary:
# Iterate summary values
for v in ev.summary.value:
#print(v)
# Check if the tag should be renamed
if v.tag in old_tags:
# Rename with new tag name
v.tag = new_tag
writer.write(ev.SerializeToString())
os.rename(new_file, input_file)
def write_summary(dpath, extracts):
dpath = dpath / FOLDER_NAME / dpath.name
writer = tf.summary.FileWriter(dpath)
for key, (steps, wall_times, values) in extracts.items():
for episode, step, wall_time, value in zip(steps, wall_times, values):
summary = tf.Summary(
value=[tf.Summary.Value(tag=key, simple_value=value)])
scalar_event = Event(episode=episode,
wall_time=wall_time,
step=step,
summary=summary)
writer.add_event(scalar_event)
writer.flush()
def write_summary(dpath, dname, fname, aggregations_per_key, steps,
wall_times):
fpath = os.path.join(dpath, dname)
fpath = os.path.abspath(fpath)
writer = tf.summary.FileWriter(fpath)
for key, aggregations in aggregations_per_key.items():
for step, wall_time, aggregation in zip(steps, wall_times,
aggregations):
summary = tf.Summary(
value=[tf.Summary.Value(tag=key, simple_value=aggregation)])
scalar_event = Event(wall_time=wall_time,
step=step,
summary=summary)
writer.add_event(scalar_event)
writer.flush()
def read_events(stream):
'''
Read and return as a generator a sequence of Event protos from
file-like object `stream`.
'''
header_size = struct.calcsize('<QI')
len_size = struct.calcsize('<Q')
footer_size = struct.calcsize('<I')
while True:
header = stream.read(header_size)
if len(header) == 0:
break
elif len(header) < header_size:
raise SummaryReaderException(
'unexpected EOF (expected a %d-byte header, '
'got %d bytes)' % (header_size, len(header)))
data_len, len_crc = struct.unpack('<QI', header)
len_crc_actual = masked_crc(header[:len_size])
if len_crc_actual != len_crc:
raise SummaryReaderException('incorrect length CRC (%d != %d)' %
(len_crc_actual, len_crc))
data = stream.read(data_len)
if len(data) < data_len:
raise SummaryReaderException(
'unexpected EOF (expected %d bytes, got %d)' %
(data_len, len(data)))
yield Event.FromString(data)
footer = stream.read(footer_size)
if len(footer) < footer_size:
raise SummaryReaderException(
'unexpected EOF (expected a %d-byte footer, '
'got %d bytes)' % (footer_size, len(footer)))
data_crc, = struct.unpack('<I', footer)
data_crc_actual = masked_crc(data)
if data_crc_actual != data_crc:
raise SummaryReaderException('incorrect data CRC (%d != %d)' %
(data_crc_actual, data_crc))
def _add_event(self, event: event_pb2.Event, step: Optional[int]) -> None:
# _add_event is vendored from TensorFlow: tensorflow/python/summary/writer/writer.py
event.wall_time = time.time()
if step is not None:
event.step = int(step)
self.writer.add_event(event)
def _write_event(self, summary_value, step):
self.writer.add_event(
Event(wall_time=round(time.time()),
step=step,
summary=Summary(value=[summary_value])))
