def testBasicTraining(self):
"""Test that we can learn a constant label of 0.0 for a fixed example."""
hparams = model.create_hparams(
'sequence_features=[Observation.code],'
'time_crossed_features=[Observation.code:'
'Observation.value.quantity.value:Observation.value.quantity.unit:'
'Observation.value.string],'
'time_concat_bucket_sizes=[12],'
'learning_rate=0.5')
time_crossed_features = [
features.split(':') for features in hparams.time_crossed_features
]
estimator = model.make_estimator(hparams, LABEL_VALUES,
FLAGS.test_tmpdir)
estimator.train(input_fn=model.get_input_fn(
mode=tf.estimator.ModeKeys.TRAIN,
input_files=[self.input_data_dir],
label_name='label.length_of_stay_range.class',
dedup=hparams.dedup,
time_windows=hparams.time_windows,
include_age=hparams.include_age,
categorical_context_features=hparams.categorical_context_features,
sequence_features=hparams.sequence_features,
time_crossed_features=time_crossed_features,
batch_size=10),
steps=100)
estimator.evaluate(
input_fn=model.get_input_fn(
mode=tf.estimator.ModeKeys.EVAL,
input_files=[self.input_data_dir],
label_name='label.length_of_stay_range.class',
dedup=hparams.dedup,
time_windows=hparams.time_windows,
include_age=hparams.include_age,
categorical_context_features=hparams.
categorical_context_features,
sequence_features=hparams.sequence_features,
time_crossed_features=time_crossed_features,
# Use a batch_size larger than the dataset to ensure we don't rely
# on the static batch_size anywhere.
batch_size=3),
steps=1)
results = list(
estimator.predict(input_fn=model.get_input_fn(
mode=tf.estimator.ModeKeys.EVAL,
input_files=[self.input_data_dir],
label_name='label.length_of_stay_range.class',
dedup=hparams.dedup,
time_windows=hparams.time_windows,
include_age=hparams.include_age,
categorical_context_features=hparams.
categorical_context_features,
sequence_features=hparams.sequence_features,
time_crossed_features=time_crossed_features,
batch_size=1,
shuffle=False)))
self.assertAllClose([0.0, 0.0, 0.0, 1.0],
results[0]['probabilities'],
atol=0.1)
def main(unused_argv: List[str]):
hparams = model.create_hparams(FLAGS.hparams_override)
tf.logging.info('Using hyperparameters %s', hparams)
time_crossed_features = [
cross.split(':')
for cross in hparams.time_crossed_features
if cross and cross != 'n/a'
]
train_input_fn = model.get_input_fn(
mode=tf.estimator.ModeKeys.TRAIN,
input_files=glob.glob(os.path.join(FLAGS.input_dir, 'train*')),
label_name=FLAGS.label_name,
dedup=hparams.dedup,
time_windows=hparams.time_windows,
include_age=hparams.include_age,
categorical_context_features=hparams.categorical_context_features,
sequence_features=hparams.sequence_features,
time_crossed_features=time_crossed_features,
batch_size=hparams.batch_size)
eval_input_fn = model.get_input_fn(
mode=tf.estimator.ModeKeys.EVAL,
input_files=glob.glob(os.path.join(FLAGS.input_dir, 'validation*')),
label_name=FLAGS.label_name,
dedup=hparams.dedup,
time_windows=hparams.time_windows,
include_age=hparams.include_age,
categorical_context_features=hparams.categorical_context_features,
sequence_features=hparams.sequence_features,
time_crossed_features=time_crossed_features,
# Fixing the batch size to get comparable evaluations.
batch_size=32)
train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn,
max_steps=FLAGS.num_train_steps)
eval_spec = tf.estimator.EvalSpec(
input_fn=eval_input_fn, steps=FLAGS.num_eval_steps, throttle_secs=60)
estimator = model.make_estimator(
hparams, FLAGS.label_values, FLAGS.output_dir)
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
def testInputFn(self, num_too_old_events):
event_ids = [1528917644, 1528917645, 1528917646, 1528917647]
timestamp = 1528917657
oldest_event_cutoff = event_ids[num_too_old_events]
feature_map, label = model.get_input_fn(
tf.estimator.ModeKeys.TRAIN, [self.input_data_dir],
'label.length_of_stay_range.class',
dedup=True,
time_windows=[timestamp - oldest_event_cutoff, 0],
include_age=True,
categorical_context_features=[],
sequence_features=['Observation.code'],
time_crossed_features=[[
'Observation.code', 'Observation.value.quantity.value',
'Observation.value.quantity.unit'
]],
batch_size=1,
shuffle=False)()
with self.test_session() as sess:
sess.run(tf.tables_initializer())
coord = tf.train.Coordinator()
tf.train.start_queue_runners(sess=sess, coord=coord)
feature_map['label'] = label
results = sess.run(feature_map)
self.assertAllEqual([b'above_14'], results['label'])
key = (model.SEQUENCE_KEY_PREFIX + 'Observation.code-til-%d' % 0)
self.assertAllEqual(
# Second "loinc:6" will be de-duped.
[b'loinc:2', b'loinc:4', b'loinc:6'][num_too_old_events:],
results[key].values)
self.assertAllEqual([[0, 0], [0, 1], [0,
2]][:3 - num_too_old_events],
results[key].indices)
self.assertAllEqual([1, 3 - num_too_old_events],
results[key].dense_shape)
cross_key = (model.SEQUENCE_KEY_PREFIX +
'Observation.code_Observation.value.quantity.value_'
'Observation.value.quantity.unit-til-0')
all_loincs = [
b'loinc:2-1.000000-mg/L', b'loinc:4-2.000000-n/a',
b'loinc:6-n/a-n/a'
]
self.assertAllEqual(all_loincs[num_too_old_events:],
results[cross_key].values)
self.assertAllEqual([[0, 0], [0, 1], [0,
2]][:3 - num_too_old_events],
results[cross_key].indices)
self.assertAllEqual([1, 3 - num_too_old_events],
results[cross_key].dense_shape)
self.assertAllClose([85.505402], results[model.CONTEXT_KEY_PREFIX +
model.AGE_KEY])
def testInputFnBatchDedup(self):
feature_map, label = model.get_input_fn(
tf.estimator.ModeKeys.TRAIN, [self.input_data_dir],
'label.length_of_stay_range.class',
dedup=True,
time_windows=[12, 0],
include_age=True,
categorical_context_features=[],
sequence_features=[
'Observation.code', 'Observation.value.quantity.value'
],
time_crossed_features=[[
'Observation.code', 'Observation.value.quantity.value',
'Observation.value.quantity.unit'
]],
batch_size=2,
shuffle=False)()
with self.test_session() as sess:
sess.run(tf.tables_initializer())
coord = tf.train.Coordinator()
tf.train.start_queue_runners(sess=sess, coord=coord)
feature_map['label'] = label
results = sess.run(feature_map)
self.assertAllEqual([b'above_14', b'above_14'], results['label'])
code_key = (model.SEQUENCE_KEY_PREFIX +
'Observation.code-til-%d' % 0)
self.assertAllEqual(
# First loinc:2 from example1 is out of range.
# Second "loinc:6" will be deduped.
[b'loinc:4', b'loinc:6', b'loinc:1', b'loinc:4'],
results[code_key].values)
# Indices are reordered on axis 1 due to deduping.
self.assertAllEqual([[0, 0], [0, 1], [1, 0], [1, 1]],
results[code_key].indices)
self.assertAllEqual([2, 2], results[code_key].dense_shape)
value_key = (model.SEQUENCE_KEY_PREFIX +
'Observation.value.quantity.value-til-%d' % 0)
self.assertAllEqual(
# First value from example1 is out of range.
[2.0, 1.0, 2.0],
results[value_key].values)
# Indices are reordered on axis 1 due to deduping.
self.assertAllEqual([[0, 0], [1, 0], [1, 1]],
results[value_key].indices)
self.assertAllEqual([2, 2], results[value_key].dense_shape)
self.assertAllClose([85.505402, 85.505402],
results[model.CONTEXT_KEY_PREFIX +
model.AGE_KEY])
