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 _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 _write_custom_summaries(self, step, logs=None):
"""Writes metrics out as custom scalar summaries.
Arguments:
step: the global step to use for TensorBoard.
logs: dict. Keys are scalar summary names, values are
NumPy scalars.
"""
logs = logs or {}
if context.executing_eagerly():
# use v2 summary ops
with self.writer.as_default(), summary_ops_v2.always_record_summaries():
for name, value in logs.items():
if isinstance(value, np.ndarray):
value = value.item()
summary_ops_v2.scalar(name, value, step=step)
else:
# use FileWriter from v1 summary
for name, value in logs.items():
if isinstance(value, np.ndarray):
value = value.item()
summary = tf_summary.Summary()
summary_value = summary.value.add()
summary_value.simple_value = value
summary_value.tag = name
self.writer.add_summary(summary, step)
self.writer.flush()
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 _log_metrics(self, logs, prefix, step):
"""Writes metrics out as custom scalar summaries.
Arguments:
logs: Dict. Keys are scalar summary names, values are NumPy scalars.
prefix: String. The prefix to apply to the scalar summary names.
step: Int. The global step to use for TensorBoard.
"""
if logs is None:
logs = {}
with context.eager_mode():
with summary_ops_v2.always_record_summaries():
train_writer = self._get_writer(self._train_run_name)
with train_writer.as_default():
self._train_summaries(train_writer,
logs,
prefix=prefix,
step=step)
eval_writer = self._get_writer(self._validation_run_name)
with eval_writer.as_default():
self._eval_summaries(eval_writer,
logs,
prefix=prefix,
step=step)
def write_weights(self, mode: str, models: Iterable[Model], step: int,
visualize: bool) -> None:
# Similar to TF implementation, but multiple models
with self.tf_summary_writers[mode].as_default(
), summary_ops_v2.always_record_summaries():
for model in models:
for layer in model.layers:
for weight in layer.weights:
weight_name = weight.name.replace(':', '_')
weight_name = "{}_{}".format(model.model_name,
weight_name)
with tfops.init_scope():
weight = backend.get_value(weight)
summary_ops_v2.histogram(weight_name,
weight,
step=step)
if visualize:
weight = self._weight_to_image(
weight=weight, kernel_channels_last=True)
if weight is not None:
summary_ops_v2.image(
weight_name,
weight,
step=step,
max_images=weight.shape[0])
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 _write_custom_summaries(self, step, logs=None):
"""Writes metrics out as custom scalar summaries.
Arguments:
step: the global step to use for TensorBoard.
logs: dict. Keys are scalar summary names, values are
NumPy scalars.
"""
logs = logs or {}
if context.executing_eagerly():
# use v2 summary ops
with self.writer.as_default(
), summary_ops_v2.always_record_summaries():
for name, value in logs.items():
if isinstance(value, np.ndarray):
value = value.item()
summary_ops_v2.scalar(name, value, step=step)
else:
# use FileWriter from v1 summary
for name, value in logs.items():
if isinstance(value, np.ndarray):
value = value.item()
summary = tf_summary.Summary()
summary_value = summary.value.add()
summary_value.simple_value = value
summary_value.tag = name
self.writer.add_summary(summary, step)
self.writer.flush()
def start(self):
"""Starts the evaluation loop."""
optimizer_checkpoint = tracking_util.Checkpoint(iter=self._iterations)
checkpoint = tracking_util.Checkpoint(model=self.model,
optimizer=optimizer_checkpoint)
for latest_checkpoint in checkpoint_utils.checkpoints_iterator(
self.checkpoint_dir):
try:
# `expect_partial` because the checkpoint can have other `Trackable`s
# such as `optimizer`.
checkpoint.restore(latest_checkpoint).expect_partial()
except (errors_impl.OpError, ) as e:
# A couple errors can happen here with the coordinator racing to write
# checkpoint:
# 1) OpError: open failed for <file path>: No such file or directory
# 2) NotFoundError (subclass of OpError): Unsuccessful
# TensorSliceReader constructor.
# TODO(rchao): Remove this except block once b/150954027 is resolved.
logging.info(
'SidecarEvaluator has an error loading '
'checkpoint: %s. Retrying. Error: %s: %s',
latest_checkpoint, e.__class__.__name__, e)
continue
if self._iterations.numpy() == _ITERATIONS_UNINITIALIZED:
raise RuntimeError(
'`iterations` cannot be loaded from the '
'checkpoint file. Please ensure `iterations` is '
'tracked in the `checkpoint` saved by the coordinator.')
logging.info(
'Evaluation starts: Model weights loaded from latest '
'checkpoint file: %s.', latest_checkpoint)
# TODO(rchao): Support arbitrary callback for extensibility.
self.model.evaluate(self.data, steps=self.steps)
logging.info('End of evaluation. Accuracy: %r', [
metric.result().numpy()
for metric in self.model.compiled_metrics.metrics
])
if self._summary_writer:
with summary_ops_v2.always_record_summaries(
), self._summary_writer.as_default():
for metric in self.model.compiled_metrics.metrics:
summary_ops_v2.scalar(
metric.name,
metric.result(),
step=self._iterations.read_value())
# TODO(rchao): Make the max evaluation robust in case users save the
# checkpoints with epoch format {epoch:03d}.
if (self.max_evaluations and latest_checkpoint.endswith(
'-{}'.format(self.max_evaluations))):
# Exit the loop because we have evaluated the final checkpoint file.
logging.info(
'Last checkpoint evaluated. SidecarEvaluator stops.')
return
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.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())
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 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 _scatter(self, epoch, logs={}):
x = self.scatter_x_model.predict_generator(self.eval_seq)
y = self.scatter_y_model.predict_generator(self.eval_seq)
# =====================================================================
writer = self._get_writer(self._train_run_name)
with context.eager_mode(), writer.as_default(), \
summary_ops_v2.always_record_summaries():
# ------------------------------------------------------------------
for i, name in self.scatter:
fig, ax = plt.subplots()
ax.hist2d(x[i],
y[i],
bins=50,
density=True,
norm=colors.SymLogNorm(linthresh=0.01,
linscale=2,
vmin=-1.0,
vmax=2.0))
ax.set_xlabel(name[0])
ax.set_ylabel(name[1])
fig.canvas.draw()
# ------------------------------------------------------------------
plot = np.array(fig.canvas.renderer.buffer_rgba())
plot = np.expand_dims(plot, 0)
summary_ops_v2.image("2DHistogram", plot, step=step)
# ------------------------------------------------------------------
plt.close()
writer.flush()
def _log_metrics(self, logs, prefix, step):
"""Writes metrics out as custom scalar summaries.
Arguments:
logs: Dict. Keys are scalar summary names, values are NumPy scalars.
prefix: String. The prefix to apply to the scalar summary names.
step: Int. The global step to use for TensorBoard.
"""
if logs is None:
logs = {}
# Group metrics by the name of their associated file writer. Values
# are lists of metrics, as (name, scalar_value) pairs.
validation_prefix = 'val_'
logs_by_writer = []
for (name, value) in logs.items():
if name in ('batch', 'size', 'num_steps'):
# Scrub non-metric items.
continue
name = prefix + name # assign batch or epoch prefix
logs_by_writer.append((name, value))
with context.eager_mode():
with summary_ops_v2.always_record_summaries():
if not logs_by_writer:
# Don't create a "validation" events file if we don't
# actually have any validation data.
pass
with self.writer.as_default():
for (name, value) in logs_by_writer:
summary_ops_v2.scalar(name, value, step=step)
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 __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.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 test_summaries_in_tf_function(self):
if not context.executing_eagerly():
return
class MyLayer(keras.layers.Layer):
def call(self, inputs):
summary_ops_v2.scalar('mean', math_ops.reduce_mean(inputs))
return inputs
tmp_dir = self.get_temp_dir()
writer = summary_ops_v2.create_file_writer_v2(tmp_dir)
with writer.as_default(), summary_ops_v2.always_record_summaries():
my_layer = MyLayer()
x = array_ops.ones((10, 10))
def my_fn(x):
return my_layer(x)
_ = my_fn(x)
event_file = gfile.Glob(os.path.join(tmp_dir, 'events*'))
self.assertLen(event_file, 1)
event_file = event_file[0]
tags = set()
for e in summary_iterator.summary_iterator(event_file):
for val in e.summary.value:
tags.add(val.tag)
self.assertEqual(set(['my_layer/mean']), tags)
def testSummaryWithAutoOutsideCompilation(self):
strategy = get_tpu_strategy()
def host_computation(x):
summary.scalar("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")
self.assertEqual(events[1].summary.value[0].simple_value, 3.0)
def testSummaryInWhile(self):
strategy = get_tpu_strategy()
def host_computation(x):
summary.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 testSummaryInCond(self, take_true_branch):
strategy = get_tpu_strategy()
def host_computation(x):
summary.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 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 on_epoch_end(self, epoch, logs={}):
if epoch % self._period == 0 and self._period != 0:
_map, average_precisions = self.evaluate_map()
print('\n')
for label, average_precision in average_precisions.items():
print(self._yolo.labels[label],
'{:.4f}'.format(average_precision))
print('mAP: {:.4f}'.format(_map))
if self._save_best and self._save_name is not None and _map > self.bestMap:
print("mAP improved from {} to {}, saving model to {}.".format(
self.bestMap, _map, self._save_name))
self.bestMap = _map
self.model.save(self._save_name)
else:
print("mAP did not improve from {}.".format(self.bestMap))
if self._tensorboard is not None:
with summary_ops_v2.always_record_summaries():
with self._tensorboard._val_writer.as_default():
name = "mAP" # Remove 'val_' prefix.
summary_ops_v2.scalar('epoch_' + name,
_map,
step=epoch)
def _log_epoch_metrics(self, epoch, logs):
"""Writes epoch metrics out as scalar summaries.
Arguments:
epoch: Int. The global step to use for TensorBoard.
logs: Dict. Keys are scalar summary names, values are scalars.
"""
if not logs:
return
train_logs = {
k: v
for k, v in logs.items() if not k.startswith('val_')
}
val_logs = {k: v for k, v in logs.items() if k.startswith('val_')}
train_logs = self._collect_learning_rate(train_logs)
with summary_ops_v2.always_record_summaries():
if train_logs:
with self._train_writer.as_default():
for name, value in train_logs.items():
summary_ops_v2.scalar(name, value, step=epoch)
if val_logs:
with self._val_writer.as_default():
for name, value in val_logs.items():
name = name[4:] # Remove 'val_' prefix.
summary_ops_v2.scalar(name, value, step=epoch)
def save_trace(self):
"""Saves tensorboard profile to event file."""
step = self.current_step()
with self.writer.as_default(), summary_ops_v2.always_record_summaries():
tf.summary.trace_export(
name='profile_batch', step=step, profiler_outdir=self.full_write_dir)
self.is_tracing = False
def testIntegerSummaries(self):
step = training_util.create_global_step()
writer = self.create_db_writer()
def adder(x, y):
state_ops.assign_add(step, 1)
summary_ops.generic('x', x)
summary_ops.generic('y', y)
sum_ = x + y
summary_ops.generic('sum', sum_)
return sum_
with summary_ops.always_record_summaries():
with writer.as_default():
self.assertEqual(5, adder(int64(2), int64(3)).numpy())
six.assertCountEqual(
self, [1, 1, 1],
get_all(self.db, 'SELECT step FROM Tensors WHERE dtype IS NOT NULL'))
six.assertCountEqual(self, ['x', 'y', 'sum'],
get_all(self.db, 'SELECT tag_name FROM Tags'))
x_id = get_one(self.db, 'SELECT tag_id FROM Tags WHERE tag_name = "x"')
y_id = get_one(self.db, 'SELECT tag_id FROM Tags WHERE tag_name = "y"')
sum_id = get_one(self.db, 'SELECT tag_id FROM Tags WHERE tag_name = "sum"')
with summary_ops.always_record_summaries():
with writer.as_default():
self.assertEqual(9, adder(int64(4), int64(5)).numpy())
six.assertCountEqual(
self, [1, 1, 1, 2, 2, 2],
get_all(self.db, 'SELECT step FROM Tensors WHERE dtype IS NOT NULL'))
six.assertCountEqual(self, [x_id, y_id, sum_id],
get_all(self.db, 'SELECT tag_id FROM Tags'))
self.assertEqual(2, get_tensor(self.db, x_id, 1))
self.assertEqual(3, get_tensor(self.db, y_id, 1))
self.assertEqual(5, get_tensor(self.db, sum_id, 1))
self.assertEqual(4, get_tensor(self.db, x_id, 2))
self.assertEqual(5, get_tensor(self.db, y_id, 2))
self.assertEqual(9, get_tensor(self.db, sum_id, 2))
six.assertCountEqual(
self, ['experiment'],
get_all(self.db, 'SELECT experiment_name FROM Experiments'))
six.assertCountEqual(self, ['run'],
get_all(self.db, 'SELECT run_name FROM Runs'))
six.assertCountEqual(self, ['user'],
get_all(self.db, 'SELECT user_name FROM Users'))
def testIntegerSummaries(self):
step = training_util.create_global_step()
writer = self.create_db_writer()
def adder(x, y):
state_ops.assign_add(step, 1)
summary_ops.generic('x', x)
summary_ops.generic('y', y)
sum_ = x + y
summary_ops.generic('sum', sum_)
return sum_
with summary_ops.always_record_summaries():
with writer.as_default():
self.assertEqual(5, adder(int64(2), int64(3)).numpy())
six.assertCountEqual(
self, [1, 1, 1],
get_all(self.db, 'SELECT step FROM Tensors WHERE dtype IS NOT NULL'))
six.assertCountEqual(self, ['x', 'y', 'sum'],
get_all(self.db, 'SELECT tag_name FROM Tags'))
x_id = get_one(self.db, 'SELECT tag_id FROM Tags WHERE tag_name = "x"')
y_id = get_one(self.db, 'SELECT tag_id FROM Tags WHERE tag_name = "y"')
sum_id = get_one(self.db, 'SELECT tag_id FROM Tags WHERE tag_name = "sum"')
with summary_ops.always_record_summaries():
with writer.as_default():
self.assertEqual(9, adder(int64(4), int64(5)).numpy())
six.assertCountEqual(
self, [1, 1, 1, 2, 2, 2],
get_all(self.db, 'SELECT step FROM Tensors WHERE dtype IS NOT NULL'))
six.assertCountEqual(self, [x_id, y_id, sum_id],
get_all(self.db, 'SELECT tag_id FROM Tags'))
self.assertEqual(2, get_tensor(self.db, x_id, 1))
self.assertEqual(3, get_tensor(self.db, y_id, 1))
self.assertEqual(5, get_tensor(self.db, sum_id, 1))
self.assertEqual(4, get_tensor(self.db, x_id, 2))
self.assertEqual(5, get_tensor(self.db, y_id, 2))
self.assertEqual(9, get_tensor(self.db, sum_id, 2))
six.assertCountEqual(
self, ['experiment'],
get_all(self.db, 'SELECT experiment_name FROM Experiments'))
six.assertCountEqual(self, ['run'],
get_all(self.db, 'SELECT run_name FROM Runs'))
six.assertCountEqual(self, ['user'],
get_all(self.db, 'SELECT user_name FROM Users'))
def _log_raw_confusion_matrix(self, logs, prefix, step):
"""
Logs the confusion matrix values as raw text.
Arguments:
logs: Training log dictionary with metric nams as keys, <dict>.
prefix: The prefix to apply to the summary names, <str>.
step: The global step to use for TensorBoard, <int>.
"""
if logs is None:
logs = {}
# Group metrics by the name of their associated file writer. Values
# are lists of metrics, as (name, scalar_value) pairs.
logs_by_writer = {
self._train_run_name: [],
self._validation_run_name: [],
}
# Get confusion matrix values
for (name, value) in logs.items():
if name.endswith('confusion_matrix'):
# Assign writer
if name.startswith(self._validation_prefix):
name = name[len(self._validation_prefix):]
writer_name = self._validation_run_name
else:
writer_name = self._train_run_name
# Add prefix and suffix
name = prefix + name + '_values'
# Convert to text
value = tf.identity(value)
value = tf.expand_dims(tf.keras.backend.flatten(value), axis=0)
value = tf.strings.as_string(
value,
precision=tf.keras.backend.epsilon(),
scientific=False)
# Add to writer list
logs_by_writer[writer_name].append((name, value))
# Iterate over writers (train, val)
with context.eager_mode():
with summary_ops_v2.always_record_summaries():
for writer_name in logs_by_writer:
these_logs = logs_by_writer[writer_name]
if not these_logs:
# Skip if empts (no validation metric)
continue
# Write logs
writer = self._get_writer(writer_name)
with writer.as_default():
for (name, value) in these_logs:
tf.summary.text(name, value, step=step)
writer.flush()
def _log_trace(self, global_batch_idx):
with self.summary_writers['train'].as_default(
), summary_ops_v2.always_record_summaries():
summary_ops_v2.trace_export(
name='batch_{}'.format(global_batch_idx),
step=global_batch_idx,
profiler_outdir=self.train_log_dir)
self.is_tracing = False
def graph(self, model):
from tensorflow.python.ops import summary_ops_v2
from tensorflow.python.keras import backend as K
with self.train_writer.as_default():
with summary_ops_v2.always_record_summaries():
if not model.run_eagerly:
summary_ops_v2.graph(K.get_graph(), step=0)
def write_epoch_models(self, mode: str) -> None:
with self.tf_summary_writers[mode].as_default(), summary_ops_v2.always_record_summaries():
summary_ops_v2.graph(backend.get_graph(), step=0)
for model in self.network.epoch_models:
summary_writable = (model.__class__.__name__ == 'Sequential'
or (hasattr(model, '_is_graph_network') and model._is_graph_network))
if summary_writable:
summary_ops_v2.keras_model(model.model_name, model, step=0)
def make_summary(self, step, tensor, type):
with context.eager_mode(), self.writer.as_default(
), summary_ops_v2.always_record_summaries():
for i in range(self.n_channels):
summary_ops_v2.image("%s image dim %d" % (type, i),
tensor[:, :, :, i, tf.newaxis],
max_images=3,
step=step)
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 _custom_step(features, labels):
del labels
logits = model(features)
with summary_ops_v2.always_record_summaries(
), writer.as_default():
summary_ops_v2.scalar('logits',
logits,
step=model.optimizer.iterations)
return logits
def testGraphSummary(self):
training_util.get_or_create_global_step()
name = 'hi'
graph = graph_pb2.GraphDef(node=(node_def_pb2.NodeDef(name=name),))
with summary_ops.always_record_summaries():
with self.create_db_writer().as_default():
summary_ops.graph(graph)
six.assertCountEqual(self, [name],
get_all(self.db, 'SELECT node_name FROM Nodes'))
def testGraphSummary(self):
training_util.get_or_create_global_step()
name = 'hi'
graph = graph_pb2.GraphDef(node=(node_def_pb2.NodeDef(name=name),))
with summary_ops.always_record_summaries():
with self.create_db_writer().as_default():
summary_ops.graph(graph)
six.assertCountEqual(self, [name],
get_all(self.db, 'SELECT node_name FROM Nodes'))
def _log_confusion_matrix(self, logs, prefix, step):
"""
Logs the confusion matrix as image.
Arguments:
logs: Training log dictionary with metric nams as keys, <dict>.
prefix: The prefix to apply to the summary names, <str>.
step: The global step to use for TensorBoard, <int>.
"""
if logs is None:
logs = {}
# Group metrics by the name of their associated file writer. Values
# are lists of metrics, as (name, scalar_value) pairs.
logs_by_writer = {
self._train_run_name: [],
self._validation_run_name: [],
}
# Get confusion matrix values
for (name, value) in logs.items():
if name.endswith('confusion_matrix'):
# Assign writer
if name.startswith(self._validation_prefix):
name = name[len(self._validation_prefix):]
writer_name = self._validation_run_name
else:
writer_name = self._train_run_name
# Add prefix
name = prefix + name
# Plot confusion matrix and decode figure
value = tf.identity(value)
value = plot_confusion_matrix(value,
class_names=self.class_names,
norm=True)
value = decode_figure(value)
# Add to writer list
logs_by_writer[writer_name].append((name, value))
# Iterate over writers (train, val)
with context.eager_mode():
with summary_ops_v2.always_record_summaries():
for writer_name in logs_by_writer:
these_logs = logs_by_writer[writer_name]
if not these_logs:
# Skip if empts (no validation metric)
continue
# Write logs
writer = self._get_writer(writer_name)
with writer.as_default():
for (name, value) in these_logs:
summary_ops_v2.image(name, value, step=step)
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 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 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 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 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 testDbURIOpen(self):
tmpdb_path = os.path.join(self.get_temp_dir(), 'tmpDbURITest.sqlite')
tmpdb_uri = six.moves.urllib_parse.urljoin("file:", tmpdb_path)
tmpdb_writer = summary_ops.create_db_writer(
tmpdb_uri,
"experimentA",
"run1",
"user1")
with summary_ops.always_record_summaries():
with tmpdb_writer.as_default():
summary_ops.scalar('t1', 2.0)
tmpdb = sqlite3.connect(tmpdb_path)
num = get_one(tmpdb, 'SELECT count(*) FROM Tags WHERE tag_name = "t1"')
self.assertEqual(num, 1)
tmpdb.close()
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 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 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 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 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 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 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 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.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())
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 f():
with summary_ops.create_file_writer(
summary_logdir).as_default(), summary_ops.always_record_summaries():
return self._all_metric_results()
def testSharing_withExplicitSummaryFileWriters(self):
logdir = self.get_temp_dir()
with session.Session() as sess:
# Initial file writer via FileWriter(session=?)
writer1 = writer.FileWriter(session=sess, logdir=logdir)
writer1.add_summary(self._createTaggedSummary("one"), 1)
writer1.flush()
# Next one via create_file_writer(), should use same file
writer2 = summary_ops_v2.create_file_writer(logdir=logdir)
with summary_ops_v2.always_record_summaries(), writer2.as_default():
summary2 = summary_ops_v2.scalar("two", 2.0, step=2)
sess.run(writer2.init())
sess.run(summary2)
sess.run(writer2.flush())
# Next has different shared name, should be in separate file
time.sleep(1.1) # Ensure filename has a different timestamp
writer3 = summary_ops_v2.create_file_writer(logdir=logdir, name="other")
with summary_ops_v2.always_record_summaries(), writer3.as_default():
summary3 = summary_ops_v2.scalar("three", 3.0, step=3)
sess.run(writer3.init())
sess.run(summary3)
sess.run(writer3.flush())
# Next uses a second session, should be in separate file
time.sleep(1.1) # Ensure filename has a different timestamp
with session.Session() as other_sess:
writer4 = summary_ops_v2.create_file_writer(logdir=logdir)
with summary_ops_v2.always_record_summaries(), writer4.as_default():
summary4 = summary_ops_v2.scalar("four", 4.0, step=4)
other_sess.run(writer4.init())
other_sess.run(summary4)
other_sess.run(writer4.flush())
# Next via FileWriter(session=?) uses same second session, should be in
# same separate file. (This checks sharing in the other direction)
writer5 = writer.FileWriter(session=other_sess, logdir=logdir)
writer5.add_summary(self._createTaggedSummary("five"), 5)
writer5.flush()
# One more via create_file_writer(), should use same file
writer6 = summary_ops_v2.create_file_writer(logdir=logdir)
with summary_ops_v2.always_record_summaries(), writer6.as_default():
summary6 = summary_ops_v2.scalar("six", 6.0, step=6)
sess.run(writer6.init())
sess.run(summary6)
sess.run(writer6.flush())
event_paths = iter(sorted(glob.glob(os.path.join(logdir, "event*"))))
# First file should have tags "one", "two", and "six"
events = summary_iterator.summary_iterator(next(event_paths))
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.assertEqual("six", next(events).summary.value[0].tag)
self.assertRaises(StopIteration, lambda: next(events))
# Second file should have just "three"
events = summary_iterator.summary_iterator(next(event_paths))
self.assertEqual("brain.Event:2", next(events).file_version)
self.assertEqual("three", next(events).summary.value[0].tag)
self.assertRaises(StopIteration, lambda: next(events))
# Third file should have "four" and "five"
events = summary_iterator.summary_iterator(next(event_paths))
self.assertEqual("brain.Event:2", next(events).file_version)
self.assertEqual("four", next(events).summary.value[0].tag)
self.assertEqual("five", next(events).summary.value[0].tag)
self.assertRaises(StopIteration, lambda: next(events))
# No more files
self.assertRaises(StopIteration, lambda: next(event_paths))
def testShouldRecordSummary(self):
self.assertFalse(summary_ops.should_record_summaries())
with summary_ops.always_record_summaries():
self.assertTrue(summary_ops.should_record_summaries())
