Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/layers/activation/softmax.py: 44%
32 statements
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
1# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Softmax activation layer."""
18import tensorflow.compat.v2 as tf
20from keras.src import backend
21from keras.src.engine.base_layer import Layer
22from keras.src.utils import tf_utils
24# isort: off
25from tensorflow.python.util.tf_export import keras_export
28def _large_compatible_negative(tensor_type):
29 """Large negative number as Tensor.
31 This function is necessary because the standard value for epsilon
32 in this module (-1e9) cannot be represented using tf.float16
34 Args:
35 tensor_type: a dtype to determine the type.
37 Returns:
38 a large negative number.
39 """
40 # In case of dtype=float16 (e.g., for mixed-precision), the largest
41 # negative number (dtypes.float16.min) is divided by 2, in order to
42 # avoid overflows when summing negative inputs.
43 if tensor_type == tf.float16:
44 return tf.float16.min / 2.0
45 return -1e9
48@keras_export("keras.layers.Softmax")
49class Softmax(Layer):
50 """Softmax activation function.
52 Example without mask:
54 >>> inp = np.asarray([1., 2., 1.])
55 >>> layer = tf.keras.layers.Softmax()
56 >>> layer(inp).numpy()
57 array([0.21194157, 0.5761169 , 0.21194157], dtype=float32)
58 >>> mask = np.asarray([True, False, True], dtype=bool)
59 >>> layer(inp, mask).numpy()
60 array([0.5, 0. , 0.5], dtype=float32)
62 Input shape:
63 Arbitrary. Use the keyword argument `input_shape`
64 (tuple of integers, does not include the samples axis)
65 when using this layer as the first layer in a model.
67 Output shape:
68 Same shape as the input.
70 Args:
71 axis: Integer, or list of Integers, axis along which the softmax
72 normalization is applied.
73 Call arguments:
74 inputs: The inputs, or logits to the softmax layer.
75 mask: A boolean mask of the same shape as `inputs`. The mask
76 specifies 1 to keep and 0 to mask. Defaults to `None`.
79 Returns:
80 Softmaxed output with the same shape as `inputs`.
81 """
83 def __init__(self, axis=-1, **kwargs):
84 super().__init__(**kwargs)
85 self.supports_masking = True
86 self.axis = axis
88 def call(self, inputs, mask=None):
89 if mask is not None:
90 # Since mask is 1.0 for positions we want to keep and 0.0 for masked
91 # positions, this operation will create a tensor which is 0.0 for
92 # positions we want to attend and -1e.9 for masked positions.
93 adder = (1.0 - tf.cast(mask, inputs.dtype)) * (
94 _large_compatible_negative(inputs.dtype)
95 )
97 # Since we are adding it to the raw scores before the softmax, this
98 # is effectively the same as removing these entirely.
99 inputs += adder
100 if isinstance(self.axis, (tuple, list)):
101 if len(self.axis) > 1:
102 return tf.exp(
103 inputs
104 - tf.reduce_logsumexp(inputs, axis=self.axis, keepdims=True)
105 )
106 else:
107 return backend.softmax(inputs, axis=self.axis[0])
108 return backend.softmax(inputs, axis=self.axis)
110 def get_config(self):
111 config = {"axis": self.axis}
112 base_config = super().get_config()
113 return dict(list(base_config.items()) + list(config.items()))
115 @tf_utils.shape_type_conversion
116 def compute_output_shape(self, input_shape):
117 return input_shape