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 testSummaryInCond(self, take_true_branch):
strategy = get_tpu_strategy()
def host_computation(x):
scalar_summary_v2.scalar("x", x, step=0)
return x * 2.0
@def_function.function
def step(take_true_branch):
def computation(x):
x = x + 1.0
if x < 5.0:
y = tpu.outside_compilation(host_computation, x)
y = tpu.outside_compilation(host_computation, x)
x = y
return x + 1.0
if take_true_branch:
return strategy.run(computation, args=(2.0, ))
else:
return strategy.run(computation, args=(10.0, ))
summary_writer = summary.create_file_writer(os.path.join(
os.getenv("TEST_TMPDIR", "/tmp")),
flush_millis=10000)
output_value = 12.
if take_true_branch:
output_value = 7.
with summary_writer.as_default(), summary.always_record_summaries():
self.assertAllEqual(
strategy.experimental_local_results(step(take_true_branch)),
constant_op.constant(output_value,
shape=(strategy.num_replicas_in_sync)))
def testWrite_usingDefaultStepVariable_fromLegacyGraph(self):
logdir = self.get_temp_dir()
try:
with context.graph_mode():
writer = summary_ops.create_file_writer(logdir)
mystep = variables.Variable(0, dtype=dtypes.int64)
summary_ops.set_step(mystep)
with writer.as_default():
write_op = summary_ops.write('tag', 1.0)
first_assign_op = mystep.assign_add(1)
second_assign_op = mystep.assign(10)
with self.cached_session() as sess:
sess.run(writer.init())
sess.run(mystep.initializer)
sess.run(write_op)
sess.run(first_assign_op)
sess.run(write_op)
sess.run(second_assign_op)
sess.run(write_op)
sess.run(writer.flush())
events = events_from_logdir(logdir)
self.assertEqual(4, len(events))
self.assertEqual(0, events[1].step)
self.assertEqual(1, events[2].step)
self.assertEqual(10, events[3].step)
finally:
# Reset to default state for other tests.
summary_ops.set_step(None)
def testWrite_usingDefaultStepVariable_fromFunction(self):
logdir = self.get_temp_dir()
try:
with context.eager_mode():
writer = summary_ops.create_file_writer(logdir)
@def_function.function
def f():
with writer.as_default():
summary_ops.write('tag', 1.0)
mystep = variables.Variable(0, dtype=dtypes.int64)
summary_ops.set_step(mystep)
f()
mystep.assign_add(1)
f()
mystep.assign(10)
f()
events = events_from_logdir(logdir)
self.assertEqual(4, len(events))
self.assertEqual(0, events[1].step)
self.assertEqual(1, events[2].step)
self.assertEqual(10, events[3].step)
finally:
# Reset to default state for other tests.
summary_ops.set_step(None)
def testScalarSummaryNameScope(self):
"""Test record_summaries_every_n_global_steps and all_summaries()."""
with ops.Graph().as_default(), self.test_session() as sess:
global_step = training_util.get_or_create_global_step()
global_step.initializer.run()
with ops.device('/cpu:0'):
step_increment = state_ops.assign_add(global_step, 1)
sess.run(step_increment) # Increment global step from 0 to 1
logdir = tempfile.mkdtemp()
with summary_ops.create_file_writer(logdir, max_queue=0,
name='t2').as_default():
with summary_ops.record_summaries_every_n_global_steps(2):
summary_ops.initialize()
with ops.name_scope('scope'):
summary_op = summary_ops.scalar('my_scalar', 2.0)
# Neither of these should produce a summary because
# global_step is 1 and "1 % 2 != 0"
sess.run(summary_ops.all_summary_ops())
sess.run(summary_op)
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 1)
# Increment global step from 1 to 2 and check that the summary
# is now written
sess.run(step_increment)
sess.run(summary_ops.all_summary_ops())
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag,
'scope/my_scalar')
def testSummaryInWhile(self):
strategy = get_tpu_strategy()
def host_computation(x):
scalar_summary_v2.scalar("x", x, step=0)
return x * 2.0
@def_function.function
def step():
def computation(x):
n = 0
while n < 3:
x = x + 1.0
y = tpu.outside_compilation(host_computation, x)
y = tpu.outside_compilation(host_computation, x)
x = y
n = n + 1
return y + 1.0
return strategy.run(computation, args=(2.0, ))
summary_writer = summary.create_file_writer(os.path.join(
os.getenv("TEST_TMPDIR", "/tmp")),
flush_millis=10000)
with summary_writer.as_default(), summary.always_record_summaries():
self.assertAllEqual(
strategy.experimental_local_results(step()),
constant_op.constant(31.,
shape=(strategy.num_replicas_in_sync)))
def construct(self, args):
with self.session.graph.as_default():
# Inputs
self.images = tf.placeholder(tf.float32,
[None, MNIST.H, MNIST.W, MNIST.C],
name="images")
self.labels = tf.placeholder(tf.int64, [None], name="labels")
# Computation
hidden = tf.keras.layers.Flatten()(self.images)
# TODO: Add `args.layers` number of hidden layers with size `args.hidden_layer`,
# using activation from `args.activation`, allowing "none", "relu", "tanh", "sigmoid".
# Store the results back to `hidden` variable.
output_layer = tf.keras.layers.Dense(MNIST.LABELS)(hidden)
self.predictions = tf.argmax(output_layer, axis=1)
# Training
loss = tf.keras.losses.sparse_categorical_crossentropy(
self.labels, output_layer, from_logits=True)
global_step = tf.train.create_global_step()
self.training = tf.train.AdamOptimizer().minimize(
loss, global_step=global_step, name="training")
# Summaries
accuracy = tf.math.reduce_mean(
tf.cast(tf.equal(self.labels, self.predictions), tf.float32))
confusion_matrix = tf.reshape(
tf.confusion_matrix(self.labels,
self.predictions,
weights=tf.not_equal(
self.labels, self.predictions),
dtype=tf.float32),
[1, MNIST.LABELS, MNIST.LABELS, 1])
summary_writer = tf_summary.create_file_writer(args.logdir,
flush_millis=10 *
1000)
self.summaries = {}
with summary_writer.as_default(
), tf_summary.record_summaries_every_n_global_steps(100):
self.summaries["train"] = [
tf_summary.scalar("train/loss", loss),
tf_summary.scalar("train/accuracy", accuracy)
]
with summary_writer.as_default(
), tf_summary.always_record_summaries():
for dataset in ["dev", "test"]:
self.summaries[dataset] = [
tf_summary.scalar(dataset + "/accuracy", accuracy),
tf_summary.image(dataset + "/confusion_matrix",
confusion_matrix)
]
with tf.control_dependencies(self.summaries[dataset]):
self.summaries[dataset].append(summary_writer.flush())
# Initialize variables
self.session.run(tf.global_variables_initializer())
with summary_writer.as_default():
tf_summary.initialize(session=self.session,
graph=self.session.graph)
def testFlushFunction(self):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir,
max_queue=999999,
flush_millis=999999)
with writer.as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, step=1)
flush_op = summary_ops.flush()
with self.test_session() as sess:
sess.run(summary_ops.summary_writer_initializer_op())
get_total = lambda: len(
summary_test_util.events_from_logdir(logdir))
# Note: First tf.Event is always file_version.
self.assertEqual(1, get_total())
sess.run(summary_ops.all_summary_ops())
self.assertEqual(1, get_total())
sess.run(flush_op)
self.assertEqual(2, get_total())
# Test "writer" parameter
sess.run(summary_ops.all_summary_ops())
sess.run(summary_ops.flush(writer=writer))
self.assertEqual(3, get_total())
sess.run(summary_ops.all_summary_ops())
sess.run(summary_ops.flush(writer=writer._resource)) # pylint:disable=protected-access
self.assertEqual(4, get_total())
def testWrite_usingDefaultStepVariable_fromLegacyGraph(self):
logdir = self.get_temp_dir()
try:
with context.graph_mode():
writer = summary_ops.create_file_writer(logdir)
mystep = variables.Variable(0, dtype=dtypes.int64)
summary_ops.set_step(mystep)
with writer.as_default():
write_op = summary_ops.write('tag', 1.0)
first_assign_op = mystep.assign_add(1)
second_assign_op = mystep.assign(10)
with self.cached_session() as sess:
sess.run(writer.init())
sess.run(mystep.initializer)
sess.run(write_op)
sess.run(first_assign_op)
sess.run(write_op)
sess.run(second_assign_op)
sess.run(write_op)
sess.run(writer.flush())
events = events_from_logdir(logdir)
self.assertEqual(4, len(events))
self.assertEqual(0, events[1].step)
self.assertEqual(1, events[2].step)
self.assertEqual(10, events[3].step)
finally:
# Reset to default state for other tests.
summary_ops.set_step(None)
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 _test_summary_for_replica_zero_only(self, d):
logdir = tempfile.mkdtemp()
def run_fn():
"""Function executed for each replica."""
with summary_writer.as_default():
replica_id = ds_context.get_replica_context().replica_id_in_sync_group
return summary_ops.write("a", replica_id)
with self.cached_session() as sess, d.scope(), \
summary_ops.always_record_summaries():
# We need global_step because summary writing op *always* has global_step
# as input, even when we always record summary or never record summary.
global_step = training_util.get_or_create_global_step()
if not context.executing_eagerly():
# When executing eagerly, variables are initialized immediately after
# creation, and its initializer will be None.
global_step.initializer.run()
summary_ops.set_step(0)
summary_writer = summary_ops.create_file_writer(logdir)
output = d.extended.call_for_each_replica(run_fn)
unwrapped = d.unwrap(output)
if not context.executing_eagerly():
sess.run(summary_writer.init())
sess.run(unwrapped)
sess.run(summary_writer.close())
events = _events_from_logdir(self, logdir)
# There will be 2 entries: 1 summary file header entry, and 1 entry
# written by replica 0.
self.assertLen(events, 2)
self.assertEqual(events[1].summary.value[0].tag, "a")
self.assertEqual(events[1].summary.value[0].simple_value, 0.0)
def _test_summary_for_replica_zero_only(self, d):
logdir = tempfile.mkdtemp()
def run_fn():
"""Function executed for each replica."""
with summary_writer.as_default():
replica_id = ds_context.get_replica_context().replica_id_in_sync_group
return summary_ops.write("a", replica_id)
with self.cached_session() as sess, d.scope(), \
summary_ops.always_record_summaries():
# We need global_step because summary writing op *always* has global_step
# as input, even when we always record summary or never record summary.
global_step = training_util.get_or_create_global_step()
if not context.executing_eagerly():
# When executing eagerly, variables are initialized immediately after
# creation, and its initializer will be None.
global_step.initializer.run()
summary_ops.set_step(0)
summary_writer = summary_ops.create_file_writer(logdir)
output = d.extended.call_for_each_replica(run_fn)
unwrapped = d.unwrap(output)
if not context.executing_eagerly():
sess.run(summary_writer.init())
sess.run(unwrapped)
sess.run(summary_writer.close())
events = _events_from_logdir(self, logdir)
# There will be 2 entries: 1 summary file header entry, and 1 entry
# written by replica 0.
self.assertLen(events, 2)
self.assertEqual(events[1].summary.value[0].tag, "a")
self.assertEqual(events[1].summary.value[0].simple_value, 0.0)
def testScalarSummaryNameScope(self):
"""Test record_summaries_every_n_global_steps and all_summaries()."""
with ops.Graph().as_default(), self.cached_session() as sess:
global_step = training_util.get_or_create_global_step()
global_step.initializer.run()
with ops.device('/cpu:0'):
step_increment = state_ops.assign_add(global_step, 1)
sess.run(step_increment) # Increment global step from 0 to 1
logdir = tempfile.mkdtemp()
with summary_ops.create_file_writer(logdir, max_queue=0,
name='t2').as_default():
with summary_ops.record_summaries_every_n_global_steps(2):
summary_ops.initialize()
with ops.name_scope('scope'):
summary_op = summary_ops.scalar('my_scalar', 2.0)
# Neither of these should produce a summary because
# global_step is 1 and "1 % 2 != 0"
sess.run(summary_ops.all_summary_ops())
sess.run(summary_op)
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 1)
# Increment global step from 1 to 2 and check that the summary
# is now written
sess.run(step_increment)
sess.run(summary_ops.all_summary_ops())
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'scope/my_scalar')
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.test_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 _host_call_fn(**kwargs):
"""Training host call.
Creates summaries for training metrics.
Args:
**kwargs: Dict of {str: Tensor} , with `Tensor` of shape `[batch]`. Must
contain key "global_step" with value of current global_step Tensor.
Returns:
List of summary ops to run on the CPU host.
"""
from tensorflow.python.ops import summary_ops_v2 # pylint: disable=g-direct-tensorflow-import,g-import-not-at-top
gs = tf.cast(kwargs.pop("global_step")[0], dtype=tf.int64)
for i, summary in enumerate(current_iteration.summaries):
with summary_ops_v2.create_file_writer(summary.logdir).as_default():
with summary_ops_v2.record_summaries_every_n_global_steps(
n=self.config.save_summary_steps, global_step=gs):
for j, summary_fn in enumerate(summary_fns[i]):
tensor = kwargs["summary_{}_{}".format(i, j)]
summary_fn(tensor, step=gs)
summary.clear_summary_tuples()
return tf.compat.v1.summary.all_v2_summary_ops()
def testSummaryControlFlowIfWithAutoOutsideCompilation(
self, take_true_branch):
strategy = get_tpu_strategy()
@def_function.function
def step():
def computation(x):
x = x + 1.0
if x < 5:
scalar_summary_v2.scalar("x", x, step=0)
x = x * 2.0
return x + 1.0
if take_true_branch:
return strategy.run(computation, args=(2.0, ))
else:
return strategy.run(computation, args=(10.0, ))
logdir = tempfile.mkdtemp()
summary_writer = summary.create_file_writer(logdir, flush_millis=10000)
output_value = 12.
if take_true_branch:
output_value = 7.
with summary_writer.as_default(), summary.always_record_summaries():
self.assertAllEqual(
strategy.experimental_local_results(step()),
constant_op.constant(output_value,
shape=(strategy.num_replicas_in_sync)))
if take_true_branch:
events = _events_from_logdir(self, logdir)
# There will be 2 entries: 1 summary file header entry, and 1 entry
# written by host.
#
self.assertLen(events, 2)
self.assertEqual(events[1].summary.value[0].tag, "cond/x")
def testHistogramSummaryWithAutoOutsideCompilation(self):
strategy = get_tpu_strategy()
def host_computation(x):
histogram_summary_v2.histogram("x", x, step=0)
return x * 2.0
@def_function.function
def step():
def computation(x):
x = x + 1.0
y = host_computation(x)
return y + 1.0
return strategy.run(computation, args=(2.0, ))
logdir = tempfile.mkdtemp()
summary_writer = summary.create_file_writer(logdir, flush_millis=10000)
with summary_writer.as_default(), summary.always_record_summaries():
self.assertAllEqual(
strategy.experimental_local_results(step()),
constant_op.constant(7.,
shape=(strategy.num_replicas_in_sync)))
events = _events_from_logdir(self, logdir)
# There will be 2 entries: 1 summary file header entry, and 1 entry
# written by host.
self.assertLen(events, 2)
self.assertEqual(events[1].summary.value[0].tag, "x")
def f():
with summary_ops.create_file_writer(logdir).as_default():
# Use assertAllEqual instead of assertTrue since it works in a defun.
self.assertAllEqual(summary_ops.write('default', 1, step=0), True)
with summary_ops.record_if(True):
self.assertAllEqual(summary_ops.write('set_on', 1, step=0), True)
with summary_ops.record_if(False):
self.assertAllEqual(summary_ops.write('set_off', 1, step=0), False)
def testWrite_ndarray(self):
logdir = self.get_temp_dir()
with context.eager_mode():
with summary_ops.create_file_writer(logdir).as_default():
summary_ops.write('tag', [[1, 2], [3, 4]], step=12)
events = events_from_logdir(logdir)
value = events[1].summary.value[0]
self.assertAllEqual([[1, 2], [3, 4]], to_numpy(value))
def f():
with summary_ops.create_file_writer(logdir).as_default():
with summary_ops.record_if(record_fn):
return [
summary_ops.write('tag', 1, step=step),
summary_ops.write('tag', 1, step=step),
summary_ops.write('tag', 1, step=step)
]
def _init_writer(self):
"""Sets file writer."""
if context.executing_eagerly():
self.writer = summary_ops_v2.create_file_writer(self.log_dir)
elif self.write_graph:
self.writer = tf_summary.FileWriter(self.log_dir, K.get_session().graph)
else:
self.writer = tf_summary.FileWriter(self.log_dir)
def testWrite_stringTensor(self):
logdir = self.get_temp_dir()
with context.eager_mode():
with summary_ops.create_file_writer(logdir).as_default():
summary_ops.write('tag', [b'foo', b'bar'], step=12)
events = events_from_logdir(logdir)
value = events[1].summary.value[0]
self.assertAllEqual([b'foo', b'bar'], to_numpy(value))
def f(tag_prefix):
with summary_ops.create_file_writer(logdir).as_default():
default_output = summary_ops.write(tag_prefix + '_default', 1, step=0)
with summary_ops.always_record_summaries():
on_output = summary_ops.write(tag_prefix + '_on', 1, step=0)
with summary_ops.never_record_summaries():
off_output = summary_ops.write(tag_prefix + '_off', 1, step=0)
return [default_output, on_output, off_output]
def run_metadata_graphs(self, *args, **kwargs):
assert context.executing_eagerly()
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
with writer.as_default():
summary_ops.run_metadata_graphs(*args, **kwargs)
writer.close()
events = events_from_logdir(logdir)
return events[1]
def exec_summary_op(self, summary_op_fn):
assert context.executing_eagerly()
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
with writer.as_default():
summary_op_fn()
writer.close()
events = events_from_logdir(logdir)
return events[1]
def testSummaryFileWritersInvalidInput(self):
# Test case for GitHub issue 46909
logdir = self.get_temp_dir()
with session.Session() as sess:
with self.assertRaises(errors_impl.InvalidArgumentError):
writer = summary_ops_v2.create_file_writer(logdir=logdir,
flush_millis=[1, 2])
sess.run(writer.init())
sess.run(writer.flush())
def run_metadata_graphs(self, *args, **kwargs):
assert context.executing_eagerly()
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
with writer.as_default():
summary_ops.run_metadata_graphs(*args, **kwargs)
writer.close()
events = events_from_logdir(logdir)
return events[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 __init__(self, session, logdir, max_queue=10, flush_secs=120,
filename_suffix=''):
"""Creates an `EventFileWriterV2` and an event file to write to.
On construction, this calls `tf.contrib.summary.create_file_writer` within
the graph from `session.graph` to look up a shared summary writer resource
for `logdir` if one exists, and create one if not. Creating the summary
writer resource in turn creates a new event file in `logdir` to be filled
with `Event` protocol buffers passed to `add_event`. Graph ops to control
this writer resource are added to `session.graph` during this init call;
stateful methods on this class will call `session.run()` on these ops.
Note that because the underlying resource is shared, it is possible that
other parts of the code using the same session may interact independently
with the resource, e.g. by flushing or even closing it. It is the caller's
responsibility to avoid any undesirable sharing in this regard.
The remaining arguments to the constructor (`flush_secs`, `max_queue`, and
`filename_suffix`) control the construction of the shared writer resource
if one is created. If an existing resource is reused, these arguments have
no effect. See `tf.contrib.summary.create_file_writer` for details.
Args:
session: A `tf.compat.v1.Session`. Session that will hold shared writer
resource. The writer ops will be added to session.graph during this
init call.
logdir: A string. Directory where event file will be written.
max_queue: Integer. Size of the queue for pending events and summaries.
flush_secs: Number. How often, in seconds, to flush the
pending events and summaries to disk.
filename_suffix: A string. Every event file's name is suffixed with
`filename_suffix`.
"""
self._session = session
self._logdir = logdir
self._closed = False
if not gfile.IsDirectory(self._logdir):
gfile.MakeDirs(self._logdir)
with self._session.graph.as_default():
with ops.name_scope('filewriter'):
file_writer = summary_ops_v2.create_file_writer(
logdir=self._logdir,
max_queue=max_queue,
flush_millis=flush_secs * 1000,
filename_suffix=filename_suffix)
with summary_ops_v2.always_record_summaries(), file_writer.as_default():
self._event_placeholder = array_ops.placeholder_with_default(
constant_op.constant('unused', dtypes.string),
shape=[])
self._add_event_op = summary_ops_v2.import_event(
self._event_placeholder)
self._init_op = file_writer.init()
self._flush_op = file_writer.flush()
self._close_op = file_writer.close()
self._session.run(self._init_op)
def keras_model(self, *args, **kwargs):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
with writer.as_default():
summary_ops.keras_model(*args, **kwargs)
writer.close()
events = events_from_logdir(logdir)
# The first event contains no summary values. The written content goes to
# the second event.
return events[1]
def keras_model(self, *args, **kwargs):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
with writer.as_default():
summary_ops.keras_model(*args, **kwargs)
writer.close()
events = events_from_logdir(logdir)
# The first event contains no summary values. The written content goes to
# the second event.
return events[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 testWrite_tensor(self):
logdir = self.get_temp_dir()
with context.eager_mode():
t = constant_op.constant([[1, 2], [3, 4]])
with summary_ops.create_file_writer(logdir).as_default():
summary_ops.write('tag', t, step=12)
expected = t.numpy()
events = events_from_logdir(logdir)
value = events[1].summary.value[0]
self.assertAllEqual(expected, to_numpy(value))
def testImageSummary(self):
strategy = get_tpu_strategy()
def run():
@def_function.function
def sample_sequence():
bsz = 3
max_length = 32 * 32
def f():
def body(step, tokens):
next_token = random_ops.random_uniform([bsz])
tokens = tokens.write(step, next_token)
return (step + 1, tokens)
def cond(step, tokens):
del tokens
return math_ops.less(step, max_length)
tokens_var = tensor_array_ops.TensorArray(
dtype=dtypes.float32,
size=max_length,
dynamic_size=False,
clear_after_read=False,
element_shape=(bsz, ),
name="tokens_accumulator",
)
step = constant_op.constant(0)
step, tokens_var = control_flow_ops.while_loop(
cond, body, [step, tokens_var])
image_flat = array_ops.transpose(tokens_var.stack(),
[1, 0])
image = array_ops.tile(
array_ops.reshape(image_flat, [bsz, 32, 32, 1]),
[1, 1, 1, 3])
image_summary_v2.image(
"image_sample", image,
constant_op.constant(5, dtype=dtypes.int64))
return strategy.run(f)
sample_sequence()
logdir = tempfile.mkdtemp()
summary_writer = summary.create_file_writer(logdir, flush_millis=10000)
with summary_writer.as_default(), summary.always_record_summaries():
run()
events = _events_from_logdir(self, logdir)
decoded_image = image_ops.decode_png(
events[1].summary.value[0].tensor.string_val[2]).numpy()
# Ensure that non-zero values were written to the image summary.
self.assertNotAllEqual(array_ops.zeros((3072, ), dtype=dtypes.float32),
list(decoded_image.flat))
def __init__(self, session, logdir, max_queue=10, flush_secs=120,
filename_suffix=''):
"""Creates an `EventFileWriterV2` and an event file to write to.
On construction, this calls `tf.contrib.summary.create_file_writer` within
the graph from `session.graph` to look up a shared summary writer resource
for `logdir` if one exists, and create one if not. Creating the summary
writer resource in turn creates a new event file in `logdir` to be filled
with `Event` protocol buffers passed to `add_event`. Graph ops to control
this writer resource are added to `session.graph` during this init call;
stateful methods on this class will call `session.run()` on these ops.
Note that because the underlying resource is shared, it is possible that
other parts of the code using the same session may interact independently
with the resource, e.g. by flushing or even closing it. It is the caller's
responsibility to avoid any undesirable sharing in this regard.
The remaining arguments to the constructor (`flush_secs`, `max_queue`, and
`filename_suffix`) control the construction of the shared writer resource
if one is created. If an existing resource is reused, these arguments have
no effect. See `tf.contrib.summary.create_file_writer` for details.
Args:
session: A `tf.compat.v1.Session`. Session that will hold shared writer
resource. The writer ops will be added to session.graph during this
init call.
logdir: A string. Directory where event file will be written.
max_queue: Integer. Size of the queue for pending events and summaries.
flush_secs: Number. How often, in seconds, to flush the
pending events and summaries to disk.
filename_suffix: A string. Every event file's name is suffixed with
`filename_suffix`.
"""
self._session = session
self._logdir = logdir
self._closed = False
gfile.MakeDirs(self._logdir)
with self._session.graph.as_default():
with ops.name_scope('filewriter'):
file_writer = summary_ops_v2.create_file_writer(
logdir=self._logdir,
max_queue=max_queue,
flush_millis=flush_secs * 1000,
filename_suffix=filename_suffix)
with summary_ops_v2.always_record_summaries(), file_writer.as_default():
self._event_placeholder = array_ops.placeholder_with_default(
constant_op.constant('unused', dtypes.string),
shape=[])
self._add_event_op = summary_ops_v2.import_event(
self._event_placeholder)
self._init_op = file_writer.init() # pylint: disable=assignment-from-no-return
self._flush_op = file_writer.flush() # pylint: disable=assignment-from-no-return
self._close_op = file_writer.close() # pylint: disable=assignment-from-no-return
self._session.run(self._init_op)
def testSummaryName(self):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir, max_queue=0)
with writer.as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, step=1)
with self.cached_session() as sess:
sess.run(summary_ops.summary_writer_initializer_op())
sess.run(summary_ops.all_summary_ops())
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(2, len(events))
self.assertEqual('scalar', events[1].summary.value[0].tag)
def testEagerMemory(self):
training_util.get_or_create_global_step()
logdir = self.get_temp_dir()
with summary_ops.create_file_writer(
logdir, max_queue=0,
name='t0').as_default(), summary_ops.always_record_summaries():
summary_ops.generic('tensor', 1, '')
summary_ops.scalar('scalar', 2.0)
summary_ops.histogram('histogram', [1.0])
summary_ops.image('image', [[[[1.0]]]])
summary_ops.audio('audio', [[1.0]], 1.0, 1)
def testSummaryName(self):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir, max_queue=0)
with writer.as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, step=1)
with self.cached_session() as sess:
sess.run(summary_ops.summary_writer_initializer_op())
sess.run(summary_ops.all_summary_ops())
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(2, len(events))
self.assertEqual('scalar', events[1].summary.value[0].tag)
def run_trace(self, f, step=1):
assert context.executing_eagerly()
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
summary_ops.trace_on(graph=True, profiler=False)
with writer.as_default():
f()
summary_ops.trace_export(name='foo', step=step)
writer.close()
events = events_from_logdir(logdir)
return events[1]
def testEagerMemory(self):
training_util.get_or_create_global_step()
logdir = self.get_temp_dir()
with summary_ops.create_file_writer(
logdir, max_queue=0,
name='t0').as_default(), summary_ops.always_record_summaries():
summary_ops.generic('tensor', 1, '')
summary_ops.scalar('scalar', 2.0)
summary_ops.histogram('histogram', [1.0])
summary_ops.image('image', [[[[1.0]]]])
summary_ops.audio('audio', [[1.0]], 1.0, 1)
def _init_writer(self, model):
"""Sets file writer."""
if context.executing_eagerly():
self.writer = summary_ops_v2.create_file_writer(self.log_dir)
if not model.run_eagerly and self.write_graph:
with self.writer.as_default():
summary_ops_v2.graph(K.get_graph())
elif self.write_graph:
self.writer = tf_summary.FileWriter(self.log_dir, K.get_graph())
else:
self.writer = tf_summary.FileWriter(self.log_dir)
def run_trace(self, f, step=1):
assert context.executing_eagerly()
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir)
summary_ops.trace_on(graph=True, profiler=False)
with writer.as_default():
f()
summary_ops.trace_export(name='foo', step=step)
writer.close()
events = events_from_logdir(logdir)
return events[1]
def _init_writer(self, model):
"""Sets file writer."""
if context.executing_eagerly():
self.writer = summary_ops_v2.create_file_writer(self.log_dir)
if not model.run_eagerly and self.write_graph:
with self.writer.as_default():
summary_ops_v2.graph(K.get_graph(), step=0)
elif self.write_graph:
self.writer = tf_summary.FileWriter(self.log_dir, K.get_graph())
else:
self.writer = tf_summary.FileWriter(self.log_dir)
def testSummaryGlobalStep(self):
step = training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_file_writer(
logdir, max_queue=0,
name='t2').as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, step=step)
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'scalar')
def testSummaryGlobalStep(self):
step = training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_file_writer(
logdir, max_queue=0,
name='t2').as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, step=step)
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'scalar')
def testWriteSummaries(self):
m = metrics.Mean()
m([1, 10, 100])
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_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_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 37.0)
def testMaxQueue(self):
logs = tempfile.mkdtemp()
with summary_ops.create_file_writer(
logs, max_queue=1, flush_millis=999999,
name='lol').as_default(), summary_ops.always_record_summaries():
get_total = lambda: len(summary_test_util.events_from_logdir(logs))
# Note: First tf.Event is always file_version.
self.assertEqual(1, get_total())
summary_ops.scalar('scalar', 2.0, step=1)
self.assertEqual(1, get_total())
# Should flush after second summary since max_queue = 1
summary_ops.scalar('scalar', 2.0, step=2)
self.assertEqual(3, get_total())
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 testSummaryGlobalStep(self):
training_util.get_or_create_global_step()
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir, max_queue=0)
with writer.as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
sess.run(summary_ops.summary_writer_initializer_op())
step, _ = sess.run(
[training_util.get_global_step(), summary_ops.all_summary_ops()])
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(2, len(events))
self.assertEqual(step, events[1].step)
def testWrite_gpuDeviceContext(self):
logdir = self.get_temp_dir()
with context.eager_mode():
with summary_ops.create_file_writer(logdir).as_default():
with ops.device('/GPU:0'):
value = constant_op.constant(42.0)
step = constant_op.constant(12, dtype=dtypes.int64)
summary_ops.write('tag', value, step=step).numpy()
empty_metadata = summary_pb2.SummaryMetadata()
events = events_from_logdir(logdir)
self.assertEqual(2, len(events))
self.assertEqual(12, events[1].step)
self.assertEqual(42, to_numpy(events[1].summary.value[0]))
self.assertEqual(empty_metadata, events[1].summary.value[0].metadata)
def testSummaryOps(self):
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_file_writer(
logdir, max_queue=0,
name='t0').as_default(), summary_ops.always_record_summaries():
summary_ops.generic('tensor', 1, '')
summary_ops.scalar('scalar', 2.0)
summary_ops.histogram('histogram', [1.0])
summary_ops.image('image', [[[[1.0]]]])
summary_ops.audio('audio', [[1.0]], 1.0, 1)
# The working condition of the ops is tested in the C++ test so we just
# test here that we're calling them correctly.
self.assertTrue(gfile.Exists(logdir))
def testWriterFlush(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
with writer.as_default():
summary_ops.scalar('one', 1.0, step=1)
self.assertEqual(1, get_total())
writer.flush()
self.assertEqual(2, get_total())
summary_ops.scalar('two', 2.0, step=2)
# Exiting the "as_default()" should do an implicit flush of the "two" tag
self.assertEqual(3, get_total())
def testSummaryOps(self):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(logdir, max_queue=0)
with writer.as_default(), summary_ops.always_record_summaries():
summary_ops.generic('tensor', 1, step=1)
summary_ops.scalar('scalar', 2.0, step=1)
summary_ops.histogram('histogram', [1.0], step=1)
summary_ops.image('image', [[[[1.0]]]], step=1)
summary_ops.audio('audio', [[1.0]], 1.0, 1, step=1)
with self.cached_session() as sess:
sess.run(summary_ops.summary_writer_initializer_op())
sess.run(summary_ops.all_summary_ops())
# The working condition of the ops is tested in the C++ test so we just
# test here that we're calling them correctly.
self.assertTrue(gfile.Exists(logdir))
def testWriterFlush(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
sess.run(summary_ops.all_summary_ops())
self.assertEqual(1, get_total())
sess.run(writer.flush())
self.assertEqual(2, get_total())
def testDefunSummarys(self):
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_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_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].simple_value, 2.0)
def testMaxQueue(self):
logdir = self.get_temp_dir()
writer = summary_ops.create_file_writer(
logdir, max_queue=1, flush_millis=999999)
with writer.as_default(), summary_ops.always_record_summaries():
summary_ops.scalar('scalar', 2.0, 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))
# Note: First tf.Event is always file_version.
self.assertEqual(1, get_total())
sess.run(summary_ops.all_summary_ops())
self.assertEqual(1, get_total())
# Should flush after second summary since max_queue = 1
sess.run(summary_ops.all_summary_ops())
self.assertEqual(3, get_total())
def testFlushFunction(self):
logs = tempfile.mkdtemp()
writer = summary_ops.create_file_writer(
logs, max_queue=999999, flush_millis=999999, name='lol')
with writer.as_default(), summary_ops.always_record_summaries():
get_total = lambda: len(summary_test_util.events_from_logdir(logs))
# Note: First tf.Event is always file_version.
self.assertEqual(1, get_total())
summary_ops.scalar('scalar', 2.0, step=1)
summary_ops.scalar('scalar', 2.0, step=2)
self.assertEqual(1, get_total())
summary_ops.flush()
self.assertEqual(3, get_total())
# Test "writer" parameter
summary_ops.scalar('scalar', 2.0, step=3)
summary_ops.flush(writer=writer)
self.assertEqual(4, get_total())
summary_ops.scalar('scalar', 2.0, step=4)
summary_ops.flush(writer=writer._resource) # pylint:disable=protected-access
self.assertEqual(5, get_total())
def testSummaryGraphModeCond(self):
with ops.Graph().as_default(), self.cached_session():
training_util.get_or_create_global_step()
logdir = tempfile.mkdtemp()
with summary_ops.create_file_writer(
logdir, max_queue=0,
name='t2').as_default(), summary_ops.always_record_summaries():
summary_ops.initialize()
training_util.get_or_create_global_step().initializer.run()
def f():
summary_ops.scalar('scalar', 2.0)
return constant_op.constant(True)
pred = array_ops.placeholder(dtypes.bool)
x = control_flow_ops.cond(pred, f,
lambda: constant_op.constant(False))
x.eval(feed_dict={pred: True})
events = summary_test_util.events_from_logdir(logdir)
self.assertEqual(len(events), 2)
self.assertEqual(events[1].summary.value[0].tag, 'cond/scalar')
def testWrite_usingDefaultStepConstant_fromFunction(self):
logdir = self.get_temp_dir()
try:
with context.eager_mode():
writer = summary_ops.create_file_writer(logdir)
@def_function.function
def f():
with writer.as_default():
summary_ops.write('tag', 1.0)
summary_ops.set_step(1)
f()
summary_ops.set_step(2)
f()
events = events_from_logdir(logdir)
self.assertEqual(3, len(events))
self.assertEqual(1, events[1].step)
# The step value will still be 1 because the value was captured at the
# time the function was first traced.
self.assertEqual(1, events[2].step)
finally:
# Reset to default state for other tests.
summary_ops.set_step(None)
def testWrite_usingDefaultStepConstant_fromLegacyGraph(self):
logdir = self.get_temp_dir()
try:
with context.graph_mode():
writer = summary_ops.create_file_writer(logdir)
summary_ops.set_step(1)
with writer.as_default():
write_op = summary_ops.write('tag', 1.0)
summary_ops.set_step(2)
with self.cached_session() as sess:
sess.run(writer.init())
sess.run(write_op)
sess.run(write_op)
sess.run(writer.flush())
events = events_from_logdir(logdir)
self.assertEqual(3, len(events))
self.assertEqual(1, events[1].step)
# The step value will still be 1 because the value was captured at the
# time the graph was constructed.
self.assertEqual(1, events[2].step)
finally:
# Reset to default state for other tests.
summary_ops.set_step(None)
