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 testTrainWithSummary(self):
with tf.Graph().as_default():
images = tf.placeholder(tf.float32, image_shape(None), name='images')
labels = tf.placeholder(tf.float32, [None, 1000], name='labels')
tf.train.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with tf.contrib.summary.always_record_summaries():
with tf.contrib.summary.create_summary_file_writer(
logdir, max_queue=0,
name='t0').as_default():
model = resnet50.ResNet50(data_format())
logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels)
tf.contrib.summary.scalar(name='loss', tensor=loss)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)
train_op = optimizer.minimize(loss)
init = tf.global_variables_initializer()
self.assertEqual(321, len(tf.global_variables()))
batch_size = 32
with tf.Session() as sess:
sess.run(init)
sess.run(tf.contrib.summary.summary_writer_initializer_op())
np_images, np_labels = random_batch(batch_size)
sess.run([train_op, tf.contrib.summary.all_summary_ops()],
feed_dict={images: np_images, labels: np_labels})
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'loss')
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 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 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)
# 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 testSummaryGlobalStep(self):
global_step = training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_summary_file_writer(
logdir, max_queue=0,
name='t2').as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, global_step=global_step)
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'scalar')
def testWriteSummaries(self):
e = SimpleEvaluator(IdentityModel())
e(3.0)
e([5.0, 7.0, 9.0])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
e.all_metric_results(logdir)
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 6.0)
def testSummaryName(self):
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_summary_file_writer(
logdir, max_queue=0,
name='t2').as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0)
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'scalar')
def testWriteSummaries(self):
e = SimpleEvaluator(IdentityModel())
e(3.0)
e([5.0, 7.0, 9.0])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
e.all_metric_results(logdir)
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 6.0)
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 testWriteSummaries(self):
m = metrics.Mean()
m([1, 10, 100])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_summary_file_writer(
logdir, max_queue=0,
name="t0").as_default(), summary_ops.always_record_summaries():
m.result() # As a side-effect will write summaries.
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 37.0)
def testWriteSummaries(self):
m = metrics.Mean()
m([1, 10, 100])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_summary_file_writer(
logdir, max_queue=0,
name="t0").as_default(), summary_ops.always_record_summaries():
m.result() # As a side-effect will write summaries.
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 37.0)
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_string = checkpoint_utils.load_variable(
config.logdir, name="_CHECKPOINTABLE_OBJECT_GRAPH")
object_graph = checkpointable_object_graph_pb2.CheckpointableObjectGraph(
)
object_graph.ParseFromString(object_graph_string)
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)
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 testWriteSummariesGraph(self):
with context.graph_mode(), ops.Graph().as_default(), self.test_session():
e = SimpleEvaluator(IdentityModel())
ds = dataset_ops.Dataset.from_tensor_slices([3.0, 5.0, 7.0, 9.0])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
init_op, call_op, results_op = e.evaluate_on_dataset(
ds, summary_logdir=logdir)
variables.global_variables_initializer().run()
e.run_evaluation(init_op, call_op, results_op)
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 6.0)
def testDefunSummarys(self):
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_summary_file_writer(
logdir, max_queue=0,
name='t1').as_default(), summary_ops.always_record_summaries():
@function.defun
def write():
summary_ops.scalar('scalar', 2.0)
write()
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 2.0)
def testWriteSummariesGraph(self):
with context.graph_mode(), ops.Graph().as_default(), self.test_session(
):
e = SimpleEvaluator(IdentityModel())
ds = dataset_ops.Dataset.from_tensor_slices([3.0, 5.0, 7.0, 9.0])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
init_op, call_op, results_op = e.evaluate_on_dataset(
ds, summary_logdir=logdir)
variables.global_variables_initializer().run()
e.run_evaluation(init_op, call_op, results_op)
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 6.0)
def test_train(self):
device, data_format = device_and_data_format()
model = resnet50.ResNet50(data_format)
tf.train.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with tf.contrib.summary.create_summary_file_writer(
logdir, max_queue=0,
name='t0').as_default(), tf.contrib.summary.always_record_summaries():
with tf.device(device):
optimizer = tf.train.GradientDescentOptimizer(0.1)
images, labels = random_batch(2)
train_one_step(model, images, labels, optimizer)
self.assertEqual(320, len(model.variables))
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'loss')
def test_train(self):
device, data_format = device_and_data_format()
model = resnet50.ResNet50(data_format)
tf.train.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with tf.contrib.summary.create_summary_file_writer(
logdir, max_queue=0, name='t0').as_default(
), tf.contrib.summary.always_record_summaries():
with tf.device(device):
optimizer = tf.train.GradientDescentOptimizer(0.1)
images, labels = random_batch(2)
train_one_step(model, images, labels, optimizer)
self.assertEqual(320, len(model.variables))
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'loss')
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_string = checkpoint_utils.load_variable(
config.logdir, name="_CHECKPOINTABLE_OBJECT_GRAPH")
object_graph = checkpointable_object_graph_pb2.CheckpointableObjectGraph()
object_graph.ParseFromString(object_graph_string)
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)
def testTrainWithSummary(self):
with tf.Graph().as_default():
images = tf.placeholder(tf.float32,
image_shape(None),
name='images')
labels = tf.placeholder(tf.float32, [None, 1000], name='labels')
tf.train.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with tf.contrib.summary.always_record_summaries():
with tf.contrib.summary.create_summary_file_writer(
logdir, max_queue=0, name='t0').as_default():
model = resnet50.ResNet50(data_format())
logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels)
tf.contrib.summary.scalar(name='loss', tensor=loss)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=0.01)
train_op = optimizer.minimize(loss)
init = tf.global_variables_initializer()
self.assertEqual(321, len(tf.global_variables()))
batch_size = 32
with tf.Session() as sess:
sess.run(init)
sess.run(tf.contrib.summary.summary_writer_initializer_op())
np_images, np_labels = random_batch(batch_size)
sess.run(
[train_op, tf.contrib.summary.all_summary_ops()],
feed_dict={
images: np_images,
labels: np_labels
})
events = summary_test_util.events_from_file(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'loss')
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 = os.path.join(snli_1_0_dir, "snli_1.0_train.txt")
os.makedirs(snli_1_0_dir)
# Four sentences in total.
with open(fake_train_file, "wt") as f:
f.write(
"gold_label\tsentence1_binary_parse\tsentence2_binary_parse\t"
"sentence1_parse\tsentence2_parse\tsentence1\tsentence2\t"
"captionID\tpairID\tlabel1\tlabel2\tlabel3\tlabel4\tlabel5\n")
f.write("neutral\t( ( Foo bar ) . )\t( ( foo . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
f.write("contradiction\t( ( Bar foo ) . )\t( ( baz . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
f.write("entailment\t( ( Quux quuz ) . )\t( ( grault . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
f.write("entailment\t( ( Quuz quux ) . )\t( ( garply . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
glove_dir = os.path.join(self._temp_data_dir, "glove")
os.makedirs(glove_dir)
glove_file = os.path.join(glove_dir, "glove.42B.300d.txt")
words = [".", "foo", "bar", "baz", "quux", "quuz", "grault", "garply"]
with open(glove_file, "wt") as f:
for i, word in enumerate(words):
f.write("%s " % word)
for j in range(data.WORD_VECTOR_LEN):
f.write("%.5f" % (i * 0.1))
if j < data.WORD_VECTOR_LEN - 1:
f.write(" ")
else:
f.write("\n")
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)
print(embed)
# 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.
spinn.train_spinn(embed, 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))
self.assertLess(train_losses[-1], train_losses[0])
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 = os.path.join(snli_1_0_dir, "snli_1.0_train.txt")
os.makedirs(snli_1_0_dir)
# Four sentences in total.
with open(fake_train_file, "wt") as f:
f.write("gold_label\tsentence1_binary_parse\tsentence2_binary_parse\t"
"sentence1_parse\tsentence2_parse\tsentence1\tsentence2\t"
"captionID\tpairID\tlabel1\tlabel2\tlabel3\tlabel4\tlabel5\n")
f.write("neutral\t( ( Foo bar ) . )\t( ( foo . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
f.write("contradiction\t( ( Bar foo ) . )\t( ( baz . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
f.write("entailment\t( ( Quux quuz ) . )\t( ( grault . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
f.write("entailment\t( ( Quuz quux ) . )\t( ( garply . )\t"
"DummySentence1Parse\tDummySentence2Parse\t"
"Foo bar.\tfoo baz.\t"
"4705552913.jpg#2\t4705552913.jpg#2r1n\t"
"neutral\tentailment\tneutral\tneutral\tneutral\n")
glove_dir = os.path.join(self._temp_data_dir, "glove")
os.makedirs(glove_dir)
glove_file = os.path.join(glove_dir, "glove.42B.300d.txt")
words = [".", "foo", "bar", "baz", "quux", "quuz", "grault", "garply"]
with open(glove_file, "wt") as f:
for i, word in enumerate(words):
f.write("%s " % word)
for j in range(data.WORD_VECTOR_LEN):
f.write("%.5f" % (i * 0.1))
if j < data.WORD_VECTOR_LEN - 1:
f.write(" ")
else:
f.write("\n")
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)
print(embed)
# 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.
spinn.train_spinn(embed, 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))
self.assertLess(train_losses[-1], train_losses[0])
