def testAscendingValTol0(self):
g = tf.Graph()
with g.as_default():
output = ops.best_step(self._BleuFile(), 0.0, False)
with self.session(graph=g) as sess:
best_step, last_step = sess.run(output)
self.assertEqual(best_step, 41500)
self.assertEqual(last_step, 46800)
def testTfEventAscendingValTol0(self):
g = tf.Graph()
with g.as_default():
output = ops.best_step(self._TfEventFile(), 0.0, False, 'bleu/dev')
with self.session(graph=g) as sess:
best_step, last_step = sess.run(output)
self.assertEqual(best_step, 102600)
self.assertEqual(last_step, 185200)
def testNoFile(self):
g = tf.Graph()
with g.as_default():
output = ops.best_step('')
with self.session(graph=g) as sess:
best_step, last_step = sess.run(output)
self.assertEqual(best_step, 0)
self.assertEqual(last_step, 0)
def testTolNon0(self):
g = tf.Graph()
with g.as_default():
output = ops.best_step(self._HistFile(), 0.1)
with self.session(graph=g) as sess:
best_step, last_step = sess.run(output)
self.assertEqual(best_step, 37553)
self.assertEqual(last_step, 42792)
def testTol0(self):
g = tf.Graph()
with g.as_default():
output = ops.best_step(self._HistFile())
with self.session(graph=g):
best_step, last_step = self.evaluate(output)
self.assertEqual(best_step, 42122)
self.assertEqual(last_step, 42792)
def _CreateLayerVariables(self):
p = self.params
wp = py_utils.WeightParams(shape=[],
init=py_utils.WeightInit.Constant(1.0),
collections=['DevBasedSchedule_vars'],
dtype=tf.float32)
self.CreateVariable('cur_factor', wp, trainable=False)
wp = py_utils.WeightParams(shape=[],
init=py_utils.WeightInit.Constant(0),
collections=['DevBasedSchedule_vars'],
dtype=tf.int64)
self.CreateVariable('ref_step', wp, trainable=False)
self._best_step = ops.best_step(self._metric_history.hist_file,
p.tolerance)
def __init__(self, params):
super().__init__(params)
p = self.params
with tf.variable_scope(p.name):
wp = py_utils.WeightParams(
shape=[],
init=py_utils.WeightInit.Constant(1.0),
collections=['DevBasedSchedule_vars'],
dtype=tf.float32)
self._cur_factor = py_utils.CreateVariable(
'cur_factor', wp, trainable=False)
wp = py_utils.WeightParams(
shape=[],
init=py_utils.WeightInit.Constant(0),
collections=['DevBasedSchedule_vars'],
dtype=tf.int64)
self._ref_step = py_utils.CreateVariable('ref_step', wp, trainable=False)
self._metric_history = early_stop.MetricHistory(p.metric_history)
self._best_step = ops.best_step(self._metric_history.hist_file,
p.tolerance)
def FProp(self, theta):
"""Creates an op to determine the best step from the metric history file.
Args:
theta: Not currently used.
Returns:
The created op.
This uses BestStepOp rather than reading the file directly from python in
order to ensure compatibility with DevBasedSchedule for learning-rate decay.
It is natural to use dev-based decay and early stopping together, for
example decaying when dev-set perplexity hasn't improved for n steps, and
stopping when it hasn't improved for 3n steps.
"""
del theta # not used
if self.params.window:
self._node = ops.best_step(self.metric_history.hist_file,
self.params.tolerance,
self.metric_history.minimize,
self.metric_history.metric)
else:
self._node = None
return self._node
def BestStep():
return ops.best_step(self.metric_history.hist_file,
self.params.tolerance,
self.metric_history.minimize,
self.metric_history.metric)
