1 # Copyright 2016 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 """Deprecated Wrappers for Neural Net (NN) Cross Entropy Operations.""" 16 17 from __future__ import absolute_import 18 from __future__ import division 19 from __future__ import print_function 20 21 from tensorflow.python.ops import nn 22 23 24 # TODO(b/33392402): Formally deprecate this API. 25 # After LSC (see b/33392402#comment1), this API will be deprecated and callers 26 # will be suggested to use the (updated version of) 27 # tf.nn.softmax_cross_entropy_with_logits. 28 def deprecated_flipped_softmax_cross_entropy_with_logits(logits, 29 labels, 30 dim=-1, 31 name=None): 32 """Computes softmax cross entropy between `logits` and `labels`. 33 34 This function diffs from tf.nn.softmax_cross_entropy_with_logits only in the 35 argument order. 36 37 Measures the probability error in discrete classification tasks in which the 38 classes are mutually exclusive (each entry is in exactly one class). For 39 example, each CIFAR-10 image is labeled with one and only one label: an image 40 can be a dog or a truck, but not both. 41 42 **NOTE:** While the classes are mutually exclusive, their probabilities 43 need not be. All that is required is that each row of `labels` is 44 a valid probability distribution. If they are not, the computation of the 45 gradient will be incorrect. 46 47 If using exclusive `labels` (wherein one and only 48 one class is true at a time), see `sparse_softmax_cross_entropy_with_logits`. 49 50 **WARNING:** This op expects unscaled logits, since it performs a `softmax` 51 on `logits` internally for efficiency. Do not call this op with the 52 output of `softmax`, as it will produce incorrect results. 53 54 `logits` and `labels` must have the same shape `[batch_size, num_classes]` 55 and the same dtype (either `float16`, `float32`, or `float64`). 56 57 Args: 58 logits: Unscaled log probabilities. 59 labels: Each row `labels[i]` must be a valid probability distribution. 60 dim: The class dimension. Defaulted to -1 which is the last dimension. 61 name: A name for the operation (optional). 62 63 Returns: 64 A 1-D `Tensor` of length `batch_size` of the same type as `logits` with the 65 softmax cross entropy loss. 66 """ 67 return nn.softmax_cross_entropy_with_logits( 68 labels=labels, logits=logits, dim=dim, name=name) 69 70 71 # TODO(b/33392402): Formally deprecate this API. 72 # After LSC (see b/33392402#comment1), this API will be deprecated and callers 73 # will be suggested to use the (updated version of) 74 # tf.nn.sparse_softmax_cross_entropy_with_logits. 75 def deprecated_flipped_sparse_softmax_cross_entropy_with_logits(logits, 76 labels, 77 name=None): 78 """Computes sparse softmax cross entropy between `logits` and `labels`. 79 80 This function diffs from tf.nn.sparse_softmax_cross_entropy_with_logits only 81 in the argument order. 82 83 Measures the probability error in discrete classification tasks in which the 84 classes are mutually exclusive (each entry is in exactly one class). For 85 example, each CIFAR-10 image is labeled with one and only one label: an image 86 can be a dog or a truck, but not both. 87 88 **NOTE:** For this operation, the probability of a given label is considered 89 exclusive. That is, soft classes are not allowed, and the `labels` vector 90 must provide a single specific index for the true class for each row of 91 `logits` (each minibatch entry). For soft softmax classification with 92 a probability distribution for each entry, see 93 `softmax_cross_entropy_with_logits`. 94 95 **WARNING:** This op expects unscaled logits, since it performs a softmax 96 on `logits` internally for efficiency. Do not call this op with the 97 output of `softmax`, as it will produce incorrect results. 98 99 A common use case is to have logits of shape `[batch_size, num_classes]` and 100 labels of shape `[batch_size]`. But higher dimensions are supported. 101 102 Args: 103 104 logits: Unscaled log probabilities of rank `r` and shape 105 `[d_0, d_1, ..., d_{r-2}, num_classes]` and dtype `float32` or `float64`. 106 labels: `Tensor` of shape `[d_0, d_1, ..., d_{r-2}]` and dtype `int32` or 107 `int64`. Each entry in `labels` must be an index in `[0, num_classes)`. 108 Other values will raise an exception when this op is run on CPU, and 109 return `NaN` for corresponding corresponding loss and gradient rows 110 on GPU. 111 name: A name for the operation (optional). 112 113 Returns: 114 A `Tensor` of the same shape as `labels` and of the same type as `logits` 115 with the softmax cross entropy loss. 116 117 Raises: 118 ValueError: If logits are scalars (need to have rank >= 1) or if the rank 119 of the labels is not equal to the rank of the logits minus one. 120 """ 121 return nn.sparse_softmax_cross_entropy_with_logits( 122 labels=labels, logits=logits, name=name) 123 124 125 # TODO(b/33392402): Formally deprecate this API. 126 # After LSC (see b/33392402#comment1), this API will be deprecated and callers 127 # will be suggested to use the (updated version of) 128 # tf.nn.sigmoid_cross_entropy_with_logits. 129 def deprecated_flipped_sigmoid_cross_entropy_with_logits(logits, 130 targets, 131 name=None): 132 """Computes sigmoid cross entropy given `logits`. 133 134 This function diffs from tf.nn.sigmoid_cross_entropy_with_logits only in the 135 argument order. 136 137 Measures the probability error in discrete classification tasks in which each 138 class is independent and not mutually exclusive. For instance, one could 139 perform multilabel classification where a picture can contain both an elephant 140 and a dog at the same time. 141 142 For brevity, let `x = logits`, `z = targets`. The logistic loss is 143 144 z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x)) 145 = z * -log(1 / (1 + exp(-x))) + (1 - z) * -log(exp(-x) / (1 + exp(-x))) 146 = z * log(1 + exp(-x)) + (1 - z) * (-log(exp(-x)) + log(1 + exp(-x))) 147 = z * log(1 + exp(-x)) + (1 - z) * (x + log(1 + exp(-x)) 148 = (1 - z) * x + log(1 + exp(-x)) 149 = x - x * z + log(1 + exp(-x)) 150 151 For x < 0, to avoid overflow in exp(-x), we reformulate the above 152 153 x - x * z + log(1 + exp(-x)) 154 = log(exp(x)) - x * z + log(1 + exp(-x)) 155 = - x * z + log(1 + exp(x)) 156 157 Hence, to ensure stability and avoid overflow, the implementation uses this 158 equivalent formulation 159 160 max(x, 0) - x * z + log(1 + exp(-abs(x))) 161 162 `logits` and `targets` must have the same type and shape. 163 164 Args: 165 logits: A `Tensor` of type `float32` or `float64`. 166 targets: A `Tensor` of the same type and shape as `logits`. 167 name: A name for the operation (optional). 168 169 Returns: 170 A `Tensor` of the same shape as `logits` with the componentwise 171 logistic losses. 172 173 Raises: 174 ValueError: If `logits` and `targets` do not have the same shape. 175 """ 176 return nn.sigmoid_cross_entropy_with_logits( 177 labels=targets, logits=logits, name=name) 178
