def _dummy_export_outputs(export_output_key, logits, predictions):
"""Returns a dummy export output dictionary for the given key."""
export_outputs = None
if export_output_key == ExportOutputKeys.CLASSIFICATION_CLASSES:
export_outputs = {
export_output_key:
tf.estimator.export.ClassificationOutput(
classes=tf.as_string(logits))
}
elif export_output_key == ExportOutputKeys.CLASSIFICATION_SCORES:
export_outputs = {
export_output_key:
tf.estimator.export.ClassificationOutput(scores=logits)
}
elif export_output_key == ExportOutputKeys.REGRESSION:
export_outputs = {
export_output_key:
tf.estimator.export.RegressionOutput(value=logits)
}
elif export_output_key == ExportOutputKeys.PREDICTION:
export_outputs = {
export_output_key:
tf.estimator.export.PredictOutput(outputs=predictions)
}
elif export_output_key == ExportOutputKeys.INVALID:
export_outputs = {export_output_key: predictions}
return export_outputs
def _hash_values_to_bins(self, values):
"""Converts a non-sparse tensor of values to bin indices."""
hash_bins = self.num_bins
mask = None
# If mask_value is set, the zeroth bin is reserved for it.
if self.mask_value is not None and hash_bins > 1:
hash_bins -= 1
mask = tf.equal(values, self.mask_value)
# Convert all values to strings before hashing.
if values.dtype.is_integer:
values = tf.as_string(values)
# Hash the strings.
if self.strong_hash:
values = tf.strings.to_hash_bucket_strong(values,
hash_bins,
name='hash',
key=self.salt)
else:
values = tf.strings.to_hash_bucket_fast(values,
hash_bins,
name='hash')
if mask is not None:
values = tf.add(values, tf.ones_like(values))
values = tf.where(mask, tf.zeros_like(values), values)
return values
def apply_randomization(features, label, randomize_prob):
"""Randomize each categorical feature with some probability."""
rnd_tok = lambda: tf.as_string(tf.random.uniform([], 0, 99999999, tf.int32))
for idx in CAT_FEATURE_INDICES:
key = feature_name(idx)
# Ignore lint since tf.cond should evaluate lambda immediately.
features[key] = tf.cond(tf.random.uniform([]) < randomize_prob,
rnd_tok,
lambda: features[key]) # pylint: disable=cell-var-from-loop
return features, label
def _hash_values_to_bins(self, values):
"""Converts a non-sparse tensor of values to bin indices."""
str_to_hash_bucket = self._get_string_to_hash_bucket_fn()
num_available_bins = self.num_bins
mask = None
# If mask_value is set, the zeroth bin is reserved for it.
if self.mask_value is not None and num_available_bins > 1:
num_available_bins -= 1
mask = tf.equal(values, self.mask_value)
# Convert all values to strings before hashing.
if values.dtype.is_integer:
values = tf.as_string(values)
values = str_to_hash_bucket(values, num_available_bins, name='hash')
if mask is not None:
values = tf.add(values, tf.compat.v1.ones_like(values))
values = tf.compat.v1.where(mask, tf.compat.v1.zeros_like(values), values)
return values
def rnd_tok():
return tf.as_string(
tf.random.uniform(tf.shape(features[key]), 0, 99999999, tf.int32)) # pylint: disable=cell-var-from-loop
def gather(self, string_values, indices):
tf.print(tf.gather(tf.as_string(string_values), indices))
def rnd_tok():
return tf.as_string(
tf.random.uniform(tf.shape(features[key]), 0, 99999999, tf.int32))
