def test_num_steps(self):
logdir = self._test_dir('test_num_steps')
with tf.Graph().as_default():
gstep = tf.contrib.framework.get_or_create_global_step()
do_step = tf.assign_add(gstep, 1)
# Do 3 steps and save.
hooks = [basic_session_run_hooks.StopAtStepHook(num_steps=3)]
scaffold = monitored_session.Scaffold().finalize()
with monitored_session.MonitoredSession(hooks=hooks) as session:
session.run(do_step)
self.assertFalse(session.should_stop())
session.run(do_step)
self.assertFalse(session.should_stop())
session.run(do_step)
self.assertTrue(session.should_stop())
save_path = scaffold.saver.save(session._coordinated_creator.tf_sess,
os.path.join(logdir, 'step-3'))
# Restore and do 4 steps.
def load_ckpt(scaffold, sess):
scaffold.saver.restore(sess, save_path)
session_creator = monitored_session.ChiefSessionCreator(
scaffold=monitored_session.Scaffold(init_fn=load_ckpt))
hooks = [basic_session_run_hooks.StopAtStepHook(num_steps=4)]
with monitored_session.MonitoredSession(
hooks=hooks, session_creator=session_creator) as session:
self.assertEqual(4, session.run(do_step))
self.assertFalse(session.should_stop())
session.run(do_step)
self.assertFalse(session.should_stop())
session.run(do_step)
self.assertFalse(session.should_stop())
session.run(do_step)
self.assertTrue(session.should_stop())
def test_last_step(self):
logdir = self._test_dir('test_last_step')
with tf.Graph().as_default():
gstep = tf.contrib.framework.get_or_create_global_step()
do_step = tf.assign_add(gstep, 1)
scaffold = monitored_session.Scaffold()
# Run till step 3 and save.
hooks = [basic_session_run_hooks.StopAtStepHook(last_step=3)]
with monitored_session.MonitoredSession('',
scaffold=scaffold,
hooks=hooks) as session:
self.assertEqual(0, session.run(gstep))
self.assertFalse(session.should_stop())
self.assertEqual(1, session.run(do_step))
self.assertFalse(session.should_stop())
self.assertEqual(2, session.run(do_step))
self.assertFalse(session.should_stop())
self.assertEqual(3, session.run(do_step))
self.assertTrue(session.should_stop())
save_path = scaffold.saver.save(session.session,
os.path.join(logdir, 'step-3'))
# Run till step 5 and save.
def load_ckpt(scaffold, sess):
scaffold.saver.restore(sess, save_path)
scaffold = monitored_session.Scaffold(init_fn=load_ckpt)
hooks = [basic_session_run_hooks.StopAtStepHook(last_step=5)]
with monitored_session.MonitoredSession('',
scaffold=scaffold,
hooks=hooks) as session:
self.assertEqual(3, session.run(gstep))
self.assertFalse(session.should_stop())
self.assertEqual(4, session.run(do_step))
self.assertFalse(session.should_stop())
self.assertEqual(5, session.run(do_step))
self.assertTrue(session.should_stop())
def test_stop_based_on_last_step(self):
h = basic_session_run_hooks.StopAtStepHook(last_step=10)
with tf.Graph().as_default():
global_step = tf.contrib.framework.get_or_create_global_step()
no_op = tf.no_op()
h.begin()
with tf.Session() as sess:
mon_sess = monitored_session._HookedSession(sess, [h])
sess.run(tf.assign(global_step, 5))
mon_sess.run(no_op)
self.assertFalse(mon_sess.should_stop())
sess.run(tf.assign(global_step, 9))
mon_sess.run(no_op)
self.assertFalse(mon_sess.should_stop())
sess.run(tf.assign(global_step, 10))
mon_sess.run(no_op)
self.assertTrue(mon_sess.should_stop())
sess.run(tf.assign(global_step, 11))
mon_sess._should_stop = False
mon_sess.run(no_op)
self.assertTrue(mon_sess.should_stop())
def _monitored_train(graph,
output_dir,
train_op,
loss_op,
global_step_tensor=None,
init_op=None,
init_feed_dict=None,
init_fn=None,
log_every_steps=10,
supervisor_is_chief=True,
supervisor_master='',
supervisor_save_model_secs=600,
keep_checkpoint_max=5,
supervisor_save_summaries_steps=100,
feed_fn=None,
steps=None,
fail_on_nan_loss=True,
hooks=None,
max_steps=None):
"""Train a model via monitored_session.
Given `graph`, a directory to write outputs to (`output_dir`), and some ops,
run a training loop. The given `train_op` performs one step of training on the
model. The `loss_op` represents the objective function of the training. It is
expected to increment the `global_step_tensor`, a scalar integer tensor
counting training steps. This function uses `Supervisor` to initialize the
graph (from a checkpoint if one is available in `output_dir`), write summaries
defined in the graph, and write regular checkpoints as defined by
`supervisor_save_model_secs`.
Training continues until `global_step_tensor` evaluates to `max_steps`, or, if
`fail_on_nan_loss`, until `loss_op` evaluates to `NaN`. In that case the
program is terminated with exit code 1.
Args:
graph: A graph to train. It is expected that this graph is not in use
elsewhere.
output_dir: A directory to write outputs to.
train_op: An op that performs one training step when run.
loss_op: A scalar loss tensor.
global_step_tensor: A tensor representing the global step. If none is given,
one is extracted from the graph using the same logic as in `Supervisor`.
init_op: An op that initializes the graph. If `None`, use `Supervisor`'s
default.
init_feed_dict: A dictionary that maps `Tensor` objects to feed values.
This feed dictionary will be used when `init_op` is evaluated.
init_fn: Optional callable passed to Supervisor to initialize the model.
log_every_steps: Output logs regularly. The logs contain timing data and the
current loss.
supervisor_is_chief: Whether the current process is the chief supervisor in
charge of restoring the model and running standard services.
supervisor_master: The master string to use when preparing the session.
supervisor_save_model_secs: Save model every
`supervisor_save_model_secs` seconds when training.
keep_checkpoint_max: The maximum number of recent checkpoint files to
keep. As new files are created, older files are deleted. If None or 0,
all checkpoint files are kept. This is simply passed as the max_to_keep
arg to tf.Saver constructor.
supervisor_save_summaries_steps: Save summaries every
`supervisor_save_summaries_steps` seconds when training.
feed_fn: A function that is called every iteration to produce a `feed_dict`
passed to `session.run` calls. Optional.
steps: Trains for this many steps (e.g. current global step + `steps`).
fail_on_nan_loss: If true, raise `NanLossDuringTrainingError` if `loss_op`
evaluates to `NaN`. If false, continue training as if nothing happened.
hooks: List of `SessionRunHook` subclass instances. Used for callbacks
inside the training loop.
max_steps: Number of total steps for which to train model. If `None`,
train forever. Two calls fit(steps=100) means 200 training iterations.
On the other hand two calls of fit(max_steps=100) means, second call
will not do any iteration since first call did all 100 steps.
Returns:
The final loss value.
Raises:
ValueError: If `output_dir`, `train_op`, `loss_op`, or `global_step_tensor`
is not provided. See `tf.contrib.framework.get_global_step` for how we
look up the latter if not provided explicitly.
NanLossDuringTrainingError: If `fail_on_nan_loss` is `True`, and loss ever
evaluates to `NaN`.
ValueError: If both `steps` and `max_steps` are not `None`.
"""
if (steps is not None) and (max_steps is not None):
raise ValueError('Can not provide both steps and max_steps.')
if not output_dir:
raise ValueError('Output directory should be non-empty %s.' % output_dir)
if train_op is None:
raise ValueError('Missing train_op.')
if loss_op is None:
raise ValueError('Missing loss_op.')
if hooks is None:
hooks = []
if not isinstance(hooks, list):
raise ValueError('Hooks should be a list.')
with graph.as_default():
global_step_tensor = contrib_variables.assert_or_get_global_step(
graph, global_step_tensor)
if global_step_tensor is None:
raise ValueError('No "global_step" was provided or found in the graph.')
if max_steps is not None:
try:
start_step = checkpoints.load_variable(output_dir,
global_step_tensor.name)
if max_steps <= start_step:
logging.info('Skipping training since max_steps has already saved.')
return None
except: # pylint: disable=bare-except
pass
# Adapted SessionRunHooks such as ExportMonitor depend on the
# CheckpointSaverHook to be executed before they should be executed.
# The `hooks` param comprises of deprecated monitor hooks
# (such as ExportMonitor). Appending them after the basic_session_run_hooks.
all_hooks = []
with graph.as_default():
all_hooks.extend([
basic_session_run_hooks.NanTensorHook(
loss_op, fail_on_nan_loss=fail_on_nan_loss),
basic_session_run_hooks.LoggingTensorHook({
'loss': loss_op.name,
'step': global_step_tensor.name
}, every_n_iter=log_every_steps),
])
scaffold = monitored_session.Scaffold(
init_op=init_op,
init_feed_dict=init_feed_dict,
init_fn=init_fn,
saver=tf_saver.Saver(
sharded=True, max_to_keep=keep_checkpoint_max, defer_build=True))
if not supervisor_is_chief:
session_creator = monitored_session.WorkerSessionCreator(
scaffold=scaffold,
master=supervisor_master)
else:
session_creator = monitored_session.ChiefSessionCreator(
scaffold=scaffold,
checkpoint_dir=output_dir,
master=supervisor_master)
summary_writer = summary_writer_cache.SummaryWriterCache.get(output_dir)
all_hooks.append(
basic_session_run_hooks.StepCounterHook(
summary_writer=summary_writer))
all_hooks.append(
basic_session_run_hooks.SummarySaverHook(
save_steps=supervisor_save_summaries_steps,
summary_writer=summary_writer,
scaffold=scaffold))
if supervisor_save_model_secs > 0:
all_hooks.append(
basic_session_run_hooks.CheckpointSaverHook(
output_dir,
save_secs=supervisor_save_model_secs,
scaffold=scaffold))
if steps is not None or max_steps is not None:
all_hooks.append(basic_session_run_hooks.StopAtStepHook(steps, max_steps))
all_hooks.extend(hooks)
with monitored_session.MonitoredSession(
session_creator=session_creator,
hooks=all_hooks) as super_sess:
loss = None
while not super_sess.should_stop():
_, loss = super_sess.run([train_op, loss_op], feed_fn() if feed_fn else
None)
return loss
def test_raise_in_both_last_step_and_num_steps(self):
with self.assertRaises(ValueError):
basic_session_run_hooks.StopAtStepHook(num_steps=10, last_step=20)
