• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
16 #ifndef TENSORFLOW_CORE_KERNELS_XENT_OP_H_
17 #define TENSORFLOW_CORE_KERNELS_XENT_OP_H_
18 // Functor definition for XentOp, must be compilable by nvcc.
19 
20 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
21 
22 #include "tensorflow/core/framework/tensor_types.h"
23 
24 namespace tensorflow {
25 namespace functor {
26 
27 // Functor used by XentOp to do the computations.
28 template <typename Device, typename T>
29 struct XentFunctor {
30   // Computes Cross Entropy loss and backprop.
31   //
32   // logits: batch_size, num_classes.
33   // labels: batch_size, num_classes.
34   // scratch: temporary tensor, dims: batch_size, 1
35   // loss: output tensor for the loss, dims: batch_size.
36   // backprop: output tensor for the backprop, dims: batch_size, num_classes.
37   void operator()(const Device &d,
38                   const Eigen::DSizes<Eigen::DenseIndex, 2> &shape,
39                   const Eigen::array<Eigen::DenseIndex, 2> &logits_bcast,
40                   const Eigen::array<Eigen::DenseIndex, 2> &labels_bcast,
41                   typename TTypes<T>::ConstMatrix logits,
42                   typename TTypes<T>::ConstMatrix labels,
43                   typename TTypes<T>::Matrix scratch,
44                   typename TTypes<T>::Vec loss,
45                   typename TTypes<T>::Matrix backprop);
46 };
47 
48 // Eigen code implementing XentFunctor::operator().
49 // This code works for both CPU and GPU and is used by the functor
50 // specializations for both device types.
51 template <typename Device, typename T>
52 struct XentEigenImpl {
ComputeXentEigenImpl53   static void Compute(const Device &d,
54                       const Eigen::DSizes<Eigen::DenseIndex, 2> &shape,
55                       const Eigen::array<Eigen::DenseIndex, 2> &logits_bcast,
56                       const Eigen::array<Eigen::DenseIndex, 2> &labels_bcast,
57                       typename TTypes<T>::ConstMatrix logits,
58                       typename TTypes<T>::ConstMatrix labels,
59                       typename TTypes<T>::Matrix scratch,
60                       typename TTypes<T>::Vec loss,
61                       typename TTypes<T>::Matrix backprop) {
62     // NOTE(touts): This duplicates some of the computations in softmax_op
63     // because we need the intermediate (logits -max(logits)) values to
64     // avoid a log(exp()) in the computation of the loss.
65 
66     const int kBatchDim = 0;
67     const int kClassDim = 1;
68 
69     const int batch_size = shape[kBatchDim];
70     const int num_classes = shape[kClassDim];
71 
72 // These arrays are used to reduce along the class dimension, and broadcast
73 // the resulting value to all classes.
74 #if !defined(EIGEN_HAS_INDEX_LIST)
75     Eigen::array<int, 1> along_class;
76     along_class[0] = kClassDim;
77     Eigen::array<int, 1> batch_only;
78     batch_only[0] = batch_size;
79     Eigen::array<int, 2> batch_by_one;
80     batch_by_one[0] = batch_size;
81     batch_by_one[1] = 1;
82     Eigen::array<int, 2> one_by_class;
83     one_by_class[0] = 1;
84     one_by_class[1] = num_classes;
85 #else
86     Eigen::IndexList<Eigen::type2index<kClassDim> > along_class;
87     Eigen::IndexList<int, Eigen::type2index<1> > batch_by_one;
88     batch_by_one.set(0, batch_size);
89     Eigen::IndexList<int> batch_only;
90     batch_only.set(0, batch_size);
91     Eigen::IndexList<Eigen::type2index<1>, int> one_by_class;
92     one_by_class.set(1, num_classes);
93 #endif
94 
95     // max_logits along classes.
96     scratch.reshape(batch_only).device(d) =
97         logits.broadcast(logits_bcast).maximum(along_class);
98 
99     // logits - max_logits.
100     backprop.device(d) =
101         logits.broadcast(logits_bcast) - scratch.broadcast(one_by_class);
102 
103     // sum(exp(logits - max_logits)) along classes.
104     scratch.reshape(batch_only).device(d) = backprop.exp().sum(along_class);
105 
106     // NOTE(keveman): Eigen on GPU dispatches to an optimized implementation
107     // for an expression of the form lhs = rhs.sum().
108     // lhs = -rhs.sum() doesn't match the above pattern, so folding in the
109     // negation before calling sum().
110     //  sum(-labels *
111     //     ((logits - max_logits) - log(sum(exp(logits - max_logits)))))
112     //  along classes
113     loss.device(d) = (labels.broadcast(labels_bcast) *
114                       (scratch.log().eval().broadcast(one_by_class) - backprop))
115                          .eval()
116                          .sum(along_class);
117 
118     // backprop: prob - labels, where
119     //   prob = exp(logits - max_logits) / sum(exp(logits - max_logits))
120     backprop.device(d) = (backprop.exp() / scratch.broadcast(one_by_class)) -
121                          labels.broadcast(labels_bcast);
122   }
123 };
124 
125 }  // namespace functor
126 }  // namespace tensorflow
127 
128 #endif  // TENSORFLOW_CORE_KERNELS_XENT_OP_H_
129