File size: 2,836 Bytes
97b6013 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 |
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Keras-based softmax layer with optional masking."""
# pylint: disable=g-classes-have-attributes
from __future__ import absolute_import
from __future__ import division
# from __future__ import google_type_annotations
from __future__ import print_function
import tensorflow as tf
@tf.keras.utils.register_keras_serializable(package='Text')
class MaskedSoftmax(tf.keras.layers.Layer):
"""Performs a softmax with optional masking on a tensor.
Arguments:
mask_expansion_axes: Any axes that should be padded on the mask tensor.
normalization_axes: On which axes the softmax should perform.
"""
def __init__(self,
mask_expansion_axes=None,
normalization_axes=None,
**kwargs):
self._mask_expansion_axes = mask_expansion_axes
if normalization_axes is None:
self._normalization_axes = (-1,)
else:
self._normalization_axes = normalization_axes
super(MaskedSoftmax, self).__init__(**kwargs)
def call(self, scores, mask=None):
if mask is not None:
for _ in range(len(scores.shape) - len(mask.shape)):
mask = tf.expand_dims(mask, axis=self._mask_expansion_axes)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
adder = (1.0 - tf.cast(mask, scores.dtype)) * -10000.0
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
scores += adder
if len(self._normalization_axes) == 1:
return tf.nn.softmax(scores, axis=self._normalization_axes[0])
else:
return tf.math.exp(scores - tf.math.reduce_logsumexp(
scores, axis=self._normalization_axes, keepdims=True))
def get_config(self):
config = {
'mask_expansion_axes': self._mask_expansion_axes,
'normalization_axes': self._normalization_axes
}
base_config = super(MaskedSoftmax, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
|