def __init__(self, model, state_manager=None, optimizer=None, model_dir=None,
config=None):
"""Initialize the Estimator.
Args:
model: The time series model to wrap (inheriting from TimeSeriesModel).
state_manager: The state manager to use, or (by default)
PassthroughStateManager if none is needed.
optimizer: The optimization algorithm to use when training, inheriting
from tf.train.Optimizer. Defaults to Adam with step size 0.02.
model_dir: See `Estimator`.
config: See `Estimator`.
"""
input_statistics_generator = math_utils.InputStatisticsFromMiniBatch(
dtype=model.dtype, num_features=model.num_features)
if state_manager is None:
state_manager = state_management.PassthroughStateManager()
if optimizer is None:
optimizer = train.AdamOptimizer(0.02)
self._model = model
model_fn = ts_head_lib.time_series_regression_head(
model, state_manager, optimizer,
input_statistics_generator=input_statistics_generator).create_estimator_spec
super(TimeSeriesRegressor, self).__init__(
model_fn=model_fn,
model_dir=model_dir,
config=config)
def test_metrics_consistent(self):
# Tests that the identity metrics used to report in-sample predictions match
# the behavior of standard metrics.
g = ops.Graph()
with g.as_default():
features = {
feature_keys.TrainEvalFeatures.TIMES:
array_ops.zeros((1, 1)),
feature_keys.TrainEvalFeatures.VALUES:
array_ops.zeros((1, 1, 1)),
"ticker":
array_ops.reshape(
math_ops.cast(variables.Variable(
name="ticker",
initial_value=0,
dtype=dtypes.int64,
collections=[ops.GraphKeys.LOCAL_VARIABLES
]).count_up_to(10),
dtype=dtypes.float32), (1, 1, 1))
}
model_fn = ts_head_lib.time_series_regression_head(
model=_TickerModel(),
state_manager=state_management.PassthroughStateManager(),
optimizer=train.GradientDescentOptimizer(
0.001)).create_estimator_spec
outputs = model_fn(features=features,
labels=None,
mode=estimator_lib.ModeKeys.EVAL)
metric_update_ops = [
metric[1] for metric in outputs.eval_metric_ops.values()
]
loss_mean, loss_update = metrics.mean(outputs.loss)
metric_update_ops.append(loss_update)
with self.test_session() as sess:
coordinator = coordinator_lib.Coordinator()
queue_runner_impl.start_queue_runners(sess, coord=coordinator)
variables.local_variables_initializer().run()
sess.run(metric_update_ops)
loss_evaled, metric_evaled, nested_metric_evaled = sess.run(
(loss_mean, outputs.eval_metric_ops["ticker"][0],
outputs.eval_metric_ops[
feature_keys.FilteringResults.STATE_TUPLE][0][0]))
# The custom model_utils metrics for in-sample predictions should be in
# sync with the Estimator's mean metric for model loss.
self.assertAllClose(0., loss_evaled)
self.assertAllClose((((0., ), ), ), metric_evaled)
self.assertAllClose((((0., ), ), ), nested_metric_evaled)
coordinator.request_stop()
coordinator.join()
def test_metrics_consistent(self):
# Tests that the identity metrics used to report in-sample predictions match
# the behavior of standard metrics.
g = ops.Graph()
with g.as_default():
features = {
feature_keys.TrainEvalFeatures.TIMES:
array_ops.zeros((1, 1)),
feature_keys.TrainEvalFeatures.VALUES:
array_ops.zeros((1, 1, 1)),
"ticker":
array_ops.reshape(
math_ops.cast(
variables.Variable(
name="ticker",
initial_value=0,
dtype=dtypes.int64,
collections=[ops.GraphKeys.LOCAL_VARIABLES])
.count_up_to(10),
dtype=dtypes.float32), (1, 1, 1))
}
model_fn = ts_head_lib.time_series_regression_head(
model=_TickerModel(),
state_manager=state_management.PassthroughStateManager(),
optimizer=train.GradientDescentOptimizer(0.001)).create_estimator_spec
outputs = model_fn(
features=features, labels=None, mode=estimator_lib.ModeKeys.EVAL)
metric_update_ops = [
metric[1] for metric in outputs.eval_metric_ops.values()]
loss_mean, loss_update = metrics.mean(outputs.loss)
metric_update_ops.append(loss_update)
with self.test_session() as sess:
coordinator = coordinator_lib.Coordinator()
queue_runner_impl.start_queue_runners(sess, coord=coordinator)
variables.local_variables_initializer().run()
sess.run(metric_update_ops)
loss_evaled, metric_evaled, nested_metric_evaled = sess.run(
(loss_mean, outputs.eval_metric_ops["ticker"][0],
outputs.eval_metric_ops[feature_keys.FilteringResults.STATE_TUPLE][
0][0]))
# The custom model_utils metrics for in-sample predictions should be in
# sync with the Estimator's mean metric for model loss.
self.assertAllClose(0., loss_evaled)
self.assertAllClose((((0.,),),), metric_evaled)
self.assertAllClose((((0.,),),), nested_metric_evaled)
coordinator.request_stop()
coordinator.join()
def _stub_model_fn():
return ts_head_lib.time_series_regression_head(
model=_StubModel(),
state_manager=state_management.PassthroughStateManager(),
optimizer=train.AdamOptimizer(0.001)).create_estimator_spec
def _stub_model_fn():
return ts_head_lib.time_series_regression_head(
model=_StubModel(),
state_manager=state_management.PassthroughStateManager(),
optimizer=train.AdamOptimizer(0.001)).create_estimator_spec
