def _infer_model(
self, input_fn, feed_fn=None, outputs=None, as_iterable=False):
# Check that model has been trained.
checkpoint_path = saver.latest_checkpoint(self._model_dir)
if not checkpoint_path:
raise NotFittedError("Couldn't find trained model at %s."
% self._model_dir)
with ops.Graph().as_default() as g:
random_seed.set_random_seed(self._config.tf_random_seed)
contrib_framework.create_global_step(g)
features = self._get_features_from_input_fn(input_fn)
predictions = self._get_predict_ops(features)
# If predictions is single output - wrap it into dict, and remember to
# return not a dict.
return_dict = isinstance(predictions, dict)
if not return_dict:
predictions = {'predictions': predictions}
# Filter what to run predictions on, if outputs provided.
if outputs:
existing_keys = predictions.keys()
predictions = {
key: value for key, value in predictions.items() if key in outputs
}
if not predictions:
raise ValueError('Expected to run at least one output from %s, '
'provided %s.' % (existing_keys, outputs))
if as_iterable:
return self._infer_model_as_iterable(
checkpoint_path, predictions, feed_fn, return_dict)
else:
return self._infer_model_single(
checkpoint_path, predictions, feed_fn, return_dict)
def _decode_infer_model(input_fn,
feed_fn=None,
model_fn=None,
runCfg=None,
model_dir=None,
outputs=None,
as_iterable=True,
iterate_batches=False):
# Check that model has been trained.
checkpoint_path = saver.latest_checkpoint(model_dir)
if not checkpoint_path:
raise NotFittedError("Couldn't find trained model at %s." % model_dir)
with tf.Graph().as_default() as g:
random_seed.set_random_seed(runCfg.tf_random_seed)
contrib_framework.create_global_step(g)
features = input_fn()
model_result, _, _ = model_fn(features,
None,
mode=tf.contrib.learn.ModeKeys.INFER)
mon_sess = monitored_session.MonitoredSession(
session_creator=monitored_session.ChiefSessionCreator(
checkpoint_filename_with_path=checkpoint_path,
scaffold=None,
config=runCfg._session_config))
return _decode_predict_generator(mon_sess, model_result, feed_fn,
iterate_batches)
def _evaluate_model(self,
input_fn,
steps,
feed_fn=None,
metrics=None,
name=''):
# TODO(wicke): Remove this once Model and associated code are gone.
if (hasattr(self._config, 'execution_mode') and
self._config.execution_mode not in ('all', 'evaluate', 'eval_evalset')):
return None, None
# Check that model has been trained.
checkpoint_path = self._model_dir
latest_path = saver.latest_checkpoint(checkpoint_path)
if not latest_path:
raise NotFittedError("Couldn't find trained model at %s."
% checkpoint_path)
# Setup output directory.
eval_dir = os.path.join(self._model_dir, 'eval' if not name else
'eval_' + name)
with ops.Graph().as_default() as g:
random_seed.set_random_seed(self._config.tf_random_seed)
global_step = contrib_framework.create_global_step(g)
features, labels = input_fn()
self._check_inputs(features, labels)
# The default return type of _get_eval_ops is ModelFnOps. But there are
# some subclasses of tf.contrib.learn.Estimator which override this
# method and use the legacy signature, namely _get_eval_ops returns an
# `eval_dict` dictionary of Tensors. The following else-statement code
# covers these cases, but will soon be deleted after the subclasses are
# updated.
# TODO(b/32664904): Update subclasses and delete the else-statement.
eval_ops = self._get_eval_ops(features, labels, metrics)
if isinstance(eval_ops, model_fn_lib.ModelFnOps): # Default signature
eval_dict = eval_ops.eval_metric_ops
else: # Legacy signature
eval_dict = eval_ops
update_op, eval_dict = self._extract_metric_update_ops(eval_dict)
eval_results, current_global_step = graph_actions.evaluate(
graph=g,
output_dir=eval_dir,
checkpoint_path=checkpoint_path,
eval_dict=eval_dict,
update_op=update_op,
global_step_tensor=global_step,
supervisor_master=self._config.evaluation_master,
feed_fn=feed_fn,
max_steps=steps)
return eval_results, current_global_step
def _infer_model(
self, input_fn, feed_fn=None, outputs=None, as_iterable=True):
# Check that model has been trained.
checkpoint_path = saver.latest_checkpoint(self._model_dir)
if not checkpoint_path:
raise NotFittedError("Couldn't find trained model at %s."
% self._model_dir)
with ops.Graph().as_default() as g:
random_seed.set_random_seed(self._config.tf_random_seed)
contrib_framework.create_global_step(g)
features = self._get_features_from_input_fn(input_fn)
# The default return type of _get_predict_ops is ModelFnOps. But there are
# some subclasses of tf.contrib.learn.Estimator which override this
# method and use the legacy signature, namely _get_predict_ops returns a
# `predictions` Tensor or dict or Tensors. The following else-statement
# code covers these cases, but will soon be deleted after the subclasses
# are updated.
# TODO(b/32664904): Update subclasses and delete the else-statement.
infer_ops = self._get_predict_ops(features)
if isinstance(infer_ops, model_fn_lib.ModelFnOps): # Default signature
predictions = infer_ops.predictions
else: # Legacy signature
predictions = infer_ops
# If predictions is single output - wrap it into dict, and remember to
# return not a dict.
return_dict = isinstance(predictions, dict)
if not return_dict:
predictions = {'predictions': predictions}
# Filter what to run predictions on, if outputs provided.
if outputs:
existing_keys = predictions.keys()
predictions = {
key: value
for key, value in six.iteritems(predictions) if key in outputs
}
if not predictions:
raise ValueError('Expected to run at least one output from %s, '
'provided %s.' % (existing_keys, outputs))
if as_iterable:
return self._infer_model_as_iterable(
checkpoint_path, predictions, feed_fn, return_dict)
else:
return self._infer_model_single(
checkpoint_path, predictions, feed_fn, return_dict)
def _infer_model(self, input_fn, feed_fn=None, outputs=None):
# Check that model has been trained.
checkpoint_path = saver.latest_checkpoint(self._model_dir)
if not checkpoint_path:
raise NotFittedError("Couldn't find trained model at %s." %
self._model_dir)
with ops.Graph().as_default() as g:
random_seed.set_random_seed(self._config.tf_random_seed)
contrib_framework.create_global_step(g)
features = self._get_features_from_input_fn(input_fn)
predictions = self._get_predict_ops(features)
# If predictions is single output - wrap it into dict, and remember to
# return not a dict.
return_dict = True
if not isinstance(predictions, dict):
predictions, return_dict = {'predictions': predictions}, False
# Filter what to run predictions on, if outputs provided.
if outputs:
existing_keys = predictions.keys()
predictions = {
key: value
for key, value in predictions.items() if key in outputs
}
if not predictions:
raise ValueError(
'Expected to run at least one output from %s, '
'provided %s.' % (existing_keys, outputs))
if feed_fn is None:
preds = graph_actions.infer(checkpoint_path, predictions)
else:
preds = {}
def _feed_fn():
while True:
yield feed_fn()
outputs = graph_actions.run_feeds(
output_dict=predictions,
feed_dicts=_feed_fn(),
restore_checkpoint_path=checkpoint_path)
for key in predictions:
preds[key] = np.concatenate(
[output[key] for output in outputs], axis=0)
if return_dict:
return preds
return preds['predictions']
def _predict(self, x, axis=-1, batch_size=None):
if self._graph is None:
raise NotFittedError()
# Use the batch size for fitting if the user did not specify one.
if batch_size is None:
batch_size = self.batch_size
predict_data_feeder = setup_train_data_feeder(
x, None, n_classes=None,
batch_size=batch_size,
shuffle=False, epochs=1)
preds = np.array(list(self._infer_model(
input_fn=predict_data_feeder.input_builder,
feed_fn=predict_data_feeder.get_feed_dict_fn(),
as_iterable=True)))
if self.n_classes > 1 and axis != -1:
preds = preds.argmax(axis=axis)
return preds
def _evaluate_model(self,
input_fn,
steps,
feed_fn=None,
metrics=None,
name=''):
# TODO(wicke): Remove this once Model and associated code are gone.
if (hasattr(self._config, 'execution_mode')
and self._config.execution_mode
not in ('all', 'evaluate', 'eval_evalset')):
return None, None
# Check that model has been trained.
checkpoint_path = self._model_dir
latest_path = saver.latest_checkpoint(checkpoint_path)
if not latest_path:
raise NotFittedError("Couldn't find trained model at %s." %
checkpoint_path)
# Setup output directory.
eval_dir = os.path.join(self._model_dir,
'eval' if not name else 'eval_' + name)
with ops.Graph().as_default() as g:
random_seed.set_random_seed(self._config.tf_random_seed)
global_step = contrib_framework.create_global_step(g)
features, targets = input_fn()
self._check_inputs(features, targets)
eval_dict = self._get_eval_ops(features, targets, metrics)
update_op, eval_dict = self._extract_metric_update_ops(eval_dict)
eval_results, current_global_step = graph_actions.evaluate(
graph=g,
output_dir=eval_dir,
checkpoint_path=checkpoint_path,
eval_dict=eval_dict,
update_op=update_op,
global_step_tensor=global_step,
supervisor_master=self._config.master,
feed_fn=feed_fn,
max_steps=steps)
return eval_results, current_global_step
