def testWriterInitAndClose(self):
logdir = self.get_temp_dir()
get_total = lambda: len(summary_test_util.events_from_logdir(logdir))
with summary_ops.always_record_summaries():
writer = summary_ops.create_file_writer(
logdir, max_queue=100, flush_millis=1000000)
self.assertEqual(1, get_total()) # file_version Event
# Calling init() again while writer is open has no effect
writer.init()
self.assertEqual(1, get_total())
try:
# Not using .as_default() to avoid implicit flush when exiting
writer.set_as_default()
summary_ops.scalar('one', 1.0, step=1)
self.assertEqual(1, get_total())
# Calling .close() should do an implicit flush
writer.close()
self.assertEqual(2, get_total())
# Calling init() on a closed writer should start a new file
time.sleep(1.1) # Ensure filename has a different timestamp
writer.init()
files = sorted(gfile.Glob(os.path.join(logdir, '*tfevents*')))
self.assertEqual(2, len(files))
get_total = lambda: len(summary_test_util.events_from_file(files[1]))
self.assertEqual(1, get_total()) # file_version Event
summary_ops.scalar('two', 2.0, step=2)
writer.close()
self.assertEqual(2, get_total())
finally:
# Clean up by resetting default writer
summary_ops.create_file_writer(None).set_as_default()
def testWriterInitAndClose(self):
logdir = self.get_temp_dir()
with summary_ops.always_record_summaries():
writer = summary_ops.create_file_writer(
logdir, max_queue=100, flush_millis=1000000)
with writer.as_default():
summary_ops.scalar('one', 1.0, step=1)
with self.cached_session() as sess:
sess.run(summary_ops.summary_writer_initializer_op())
get_total = lambda: len(summary_test_util.events_from_logdir(logdir))
self.assertEqual(1, get_total()) # file_version Event
# Running init() again while writer is open has no effect
sess.run(writer.init())
self.assertEqual(1, get_total())
sess.run(summary_ops.all_summary_ops())
self.assertEqual(1, get_total())
# Running close() should do an implicit flush
sess.run(writer.close())
self.assertEqual(2, get_total())
# Running init() on a closed writer should start a new file
time.sleep(1.1) # Ensure filename has a different timestamp
sess.run(writer.init())
sess.run(summary_ops.all_summary_ops())
sess.run(writer.close())
files = sorted(gfile.Glob(os.path.join(logdir, '*tfevents*')))
self.assertEqual(2, len(files))
self.assertEqual(2, len(summary_test_util.events_from_file(files[1])))
def testSharedName(self):
logdir = self.get_temp_dir()
with summary_ops.always_record_summaries():
# Create with default shared name (should match logdir)
writer1 = summary_ops.create_file_writer(logdir)
with writer1.as_default():
summary_ops.scalar('one', 1.0, step=1)
# Create with explicit logdir shared name (should be same resource/file)
shared_name = 'logdir:' + logdir
writer2 = summary_ops.create_file_writer(logdir, name=shared_name)
with writer2.as_default():
summary_ops.scalar('two', 2.0, step=2)
# Create with different shared name (should be separate resource/file)
writer3 = summary_ops.create_file_writer(logdir, name='other')
with writer3.as_default():
summary_ops.scalar('three', 3.0, step=3)
with self.cached_session() as sess:
# Run init ops across writers sequentially to avoid race condition.
# TODO(nickfelt): fix race condition in resource manager lookup or create
sess.run(writer1.init())
sess.run(writer2.init())
time.sleep(1.1) # Ensure filename has a different timestamp
sess.run(writer3.init())
sess.run(summary_ops.all_summary_ops())
sess.run([writer1.flush(), writer2.flush(), writer3.flush()])
event_files = iter(sorted(gfile.Glob(os.path.join(logdir, '*tfevents*'))))
# First file has tags "one" and "two"
events = summary_test_util.events_from_file(next(event_files))
self.assertEqual('brain.Event:2', events[0].file_version)
tags = [e.summary.value[0].tag for e in events[1:]]
self.assertItemsEqual(['one', 'two'], tags)
# Second file has tag "three"
events = summary_test_util.events_from_file(next(event_files))
self.assertEqual('brain.Event:2', events[0].file_version)
tags = [e.summary.value[0].tag for e in events[1:]]
self.assertItemsEqual(['three'], tags)
# No more files
self.assertRaises(StopIteration, lambda: next(event_files))
def testSharedName(self):
logdir = self.get_temp_dir()
with summary_ops.always_record_summaries():
# Create with default shared name (should match logdir)
writer1 = summary_ops.create_file_writer(logdir)
with writer1.as_default():
summary_ops.scalar('one', 1.0, step=1)
summary_ops.flush()
# Create with explicit logdir shared name (should be same resource/file)
shared_name = 'logdir:' + logdir
writer2 = summary_ops.create_file_writer(logdir, name=shared_name)
with writer2.as_default():
summary_ops.scalar('two', 2.0, step=2)
summary_ops.flush()
# Create with different shared name (should be separate resource/file)
time.sleep(1.1) # Ensure filename has a different timestamp
writer3 = summary_ops.create_file_writer(logdir, name='other')
with writer3.as_default():
summary_ops.scalar('three', 3.0, step=3)
summary_ops.flush()
event_files = iter(sorted(gfile.Glob(os.path.join(logdir, '*tfevents*'))))
# First file has tags "one" and "two"
events = iter(summary_test_util.events_from_file(next(event_files)))
self.assertEqual('brain.Event:2', next(events).file_version)
self.assertEqual('one', next(events).summary.value[0].tag)
self.assertEqual('two', next(events).summary.value[0].tag)
self.assertRaises(StopIteration, lambda: next(events))
# Second file has tag "three"
events = iter(summary_test_util.events_from_file(next(event_files)))
self.assertEqual('brain.Event:2', next(events).file_version)
self.assertEqual('three', next(events).summary.value[0].tag)
self.assertRaises(StopIteration, lambda: next(events))
# No more files
self.assertRaises(StopIteration, lambda: next(event_files))
def testTrainSpinn(self):
"""Test with fake toy SNLI data and GloVe vectors."""
# 1. Create and load a fake SNLI data file and a fake GloVe embedding file.
snli_1_0_dir = os.path.join(self._temp_data_dir, "snli/snli_1.0")
fake_train_file = self._create_test_data(snli_1_0_dir)
vocab = data.load_vocabulary(self._temp_data_dir)
word2index, embed = data.load_word_vectors(self._temp_data_dir, vocab)
train_data = data.SnliData(fake_train_file, word2index)
dev_data = data.SnliData(fake_train_file, word2index)
test_data = data.SnliData(fake_train_file, word2index)
# 2. Create a fake config.
config = _test_spinn_config(
data.WORD_VECTOR_LEN, 4,
logdir=os.path.join(self._temp_data_dir, "logdir"))
# 3. Test training of a SPINN model.
trainer = spinn.train_or_infer_spinn(
embed, word2index, train_data, dev_data, test_data, config)
# 4. Load train loss values from the summary files and verify that they
# decrease with training.
summary_file = glob.glob(os.path.join(config.logdir, "events.out.*"))[0]
events = summary_test_util.events_from_file(summary_file)
train_losses = [event.summary.value[0].simple_value for event in events
if event.summary.value
and event.summary.value[0].tag == "train/loss"]
self.assertEqual(config.epochs, len(train_losses))
# 5. Verify that checkpoints exist and contains all the expected variables.
self.assertTrue(glob.glob(os.path.join(config.logdir, "ckpt*")))
object_graph = trackable_utils.object_metadata(
checkpoint_management.latest_checkpoint(config.logdir))
ckpt_variable_names = set()
for node in object_graph.nodes:
for attribute in node.attributes:
ckpt_variable_names.add(attribute.full_name)
self.assertIn("global_step", ckpt_variable_names)
for v in trainer.variables:
variable_name = v.name[:v.name.index(":")] if ":" in v.name else v.name
self.assertIn(variable_name, ckpt_variable_names)