Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/tensorflow_addons/losses/contrastive.py: 75%

1# Copyright 2019 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"""Implements contrastive loss.""" 

16 

17import tensorflow as tf 

18from typeguard import typechecked 

19 

20from tensorflow_addons.utils.keras_utils import LossFunctionWrapper 

21from tensorflow_addons.utils.types import TensorLike, Number 

22 

23 

24@tf.keras.utils.register_keras_serializable(package="Addons") 

25@tf.function 

26def contrastive_loss( 

27 y_true: TensorLike, y_pred: TensorLike, margin: Number = 1.0 

28) -> tf.Tensor: 

29 r"""Computes the contrastive loss between `y_true` and `y_pred`. 

30 

31 This loss encourages the embedding to be close to each other for 

32 the samples of the same label and the embedding to be far apart at least 

33 by the margin constant for the samples of different labels. 

34 

35 The euclidean distances `y_pred` between two embedding matrices 

36 `a` and `b` with shape `[batch_size, hidden_size]` can be computed 

37 as follows: 

38 

39 >>> a = tf.constant([[1, 2], 

40 ... [3, 4], 

41 ... [5, 6]], dtype=tf.float16) 

42 >>> b = tf.constant([[5, 9], 

43 ... [3, 6], 

44 ... [1, 8]], dtype=tf.float16) 

45 >>> y_pred = tf.linalg.norm(a - b, axis=1) 

46 >>> y_pred 

47 <tf.Tensor: shape=(3,), dtype=float16, numpy=array([8.06 , 2. , 4.473], 

48 dtype=float16)> 

49 

50 <... Note: constants a & b have been used purely for 

51 example purposes and have no significant value ...> 

52 

53 See: http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf 

54 

55 Args: 

56 y_true: 1-D integer `Tensor` with shape `[batch_size]` of 

57 binary labels indicating positive vs negative pair. 

58 y_pred: 1-D float `Tensor` with shape `[batch_size]` of 

59 distances between two embedding matrices. 

60 margin: margin term in the loss definition. 

61 

62 Returns: 

63 contrastive_loss: 1-D float `Tensor` with shape `[batch_size]`. 

64 """ 

65 y_pred = tf.convert_to_tensor(y_pred) 

66 y_true = tf.dtypes.cast(y_true, y_pred.dtype) 

67 return y_true * tf.math.square(y_pred) + (1.0 - y_true) * tf.math.square( 

68 tf.math.maximum(margin - y_pred, 0.0) 

69 ) 

70 

71 

72@tf.keras.utils.register_keras_serializable(package="Addons") 

73class ContrastiveLoss(LossFunctionWrapper): 

74 r"""Computes the contrastive loss between `y_true` and `y_pred`. 

75 

76 This loss encourages the embedding to be close to each other for 

77 the samples of the same label and the embedding to be far apart at least 

78 by the margin constant for the samples of different labels. 

79 

80 See: http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf 

81 

82 We expect labels `y_true` to be provided as 1-D integer `Tensor` 

83 with shape `[batch_size]` of binary integer labels. And `y_pred` must be 

84 1-D float `Tensor` with shape `[batch_size]` of distances between two 

85 embedding matrices. 

86 

87 The euclidean distances `y_pred` between two embedding matrices 

88 `a` and `b` with shape `[batch_size, hidden_size]` can be computed 

89 as follows: 

90 

91 >>> a = tf.constant([[1, 2], 

92 ... [3, 4],[5, 6]], dtype=tf.float16) 

93 >>> b = tf.constant([[5, 9], 

94 ... [3, 6],[1, 8]], dtype=tf.float16) 

95 >>> y_pred = tf.linalg.norm(a - b, axis=1) 

96 >>> y_pred 

97 <tf.Tensor: shape=(3,), dtype=float16, numpy=array([8.06 , 2. , 4.473], 

98 dtype=float16)> 

99 

100 <... Note: constants a & b have been used purely for 

101 example purposes and have no significant value ...> 

102 

103 Args: 

104 margin: `Float`, margin term in the loss definition. 

105 Default value is 1.0. 

106 reduction: (Optional) Type of `tf.keras.losses.Reduction` to apply. 

107 Default value is `SUM_OVER_BATCH_SIZE`. 

108 name: (Optional) name for the loss. 

109 """ 

110 

111 @typechecked 

112 def __init__( 

113 self, 

114 margin: Number = 1.0, 

115 reduction: str = tf.keras.losses.Reduction.SUM_OVER_BATCH_SIZE, 

116 name: str = "contrastive_loss", 

117 ): 

118 super().__init__( 

119 contrastive_loss, reduction=reduction, name=name, margin=margin 

120 )