Home | History | Annotate | Download | only in kernels
      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_KERNELS_XENT_OP_H_
     17 #define TENSORFLOW_KERNELS_XENT_OP_H_
     18 // Functor definition for SparseXentOp, must be compilable by nvcc.
     19 
     20 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
     21 #include "tensorflow/core/framework/tensor_types.h"
     22 #include "tensorflow/core/kernels/bounds_check.h"
     23 #include "tensorflow/core/platform/macros.h"
     24 #include "tensorflow/core/platform/types.h"
     25 
     26 namespace tensorflow {
     27 
     28 namespace sparse_xent_helpers {
     29 
     30 template <typename T>
     31 typename TTypes<const T, 1>::Tensor32Bit To32BitConst(
     32     typename TTypes<T>::Vec in) {
     33   return To32Bit(typename TTypes<T>::ConstVec(in.data(), in.dimensions()));
     34 }
     35 
     36 template <typename T>
     37 typename TTypes<const T, 2>::Tensor32Bit To32BitConst(
     38     typename TTypes<T>::Matrix in) {
     39   return To32Bit(typename TTypes<T>::ConstMatrix(in.data(), in.dimensions()));
     40 }
     41 
     42 }  // namespace sparse_xent_helpers
     43 
     44 namespace generator {
     45 
     46 // Generator for calculation of the sparse Xent loss.
     47 // This generator takes the logits, the sum of the exponentiated
     48 // logits, and the label indices.  For each minibatch entry, ignoring
     49 // the batch index b, it calculates:
     50 //
     51 //   loss[j] = (log(sum_exp_logits) - logits[j]) * 1{ j == label }
     52 //
     53 // for j = 0 .. num_classes.  This value must be summed over all j for
     54 // the final loss.
     55 template <typename T, typename Index>
     56 class SparseXentLossGenerator {
     57  public:
     58   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE SparseXentLossGenerator(
     59       typename TTypes<const T, 2>::Tensor32Bit logits,
     60       typename TTypes<const T, 1>::Tensor32Bit sum_exp_logits,
     61       typename TTypes<const Index, 1>::Tensor32Bit labels,
     62       const Index max_depth)
     63       : logits_(logits),
     64         sum_exp_logits_(sum_exp_logits),
     65         labels_(labels),
     66         max_depth_(max_depth) {}
     67 
     68   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE T
     69   operator()(const Eigen::array<int, 2>& coords) const {
     70     const int batch = coords[0];
     71     const int depth = coords[1];
     72     const Index label = tensorflow::internal::SubtleMustCopy(labels_(batch));
     73     if (!FastBoundsCheck(label, max_depth_)) {
     74       return Eigen::NumTraits<T>::quiet_NaN();
     75     }
     76     return TF_PREDICT_FALSE(label == depth)
     77                ? (Eigen::numext::log(sum_exp_logits_(batch)) - logits_(coords))
     78                : T(0.0);
     79   };
     80 
     81  private:
     82   typename TTypes<const T, 2>::Tensor32Bit logits_;
     83   typename TTypes<const T, 1>::Tensor32Bit sum_exp_logits_;
     84   typename TTypes<const Index, 1>::Tensor32Bit labels_;
     85   const Index max_depth_;
     86 };
     87 
     88 // Generator for calculation of the sparse Xent gradient.
     89 // This generator takes the exponentiated logits, their sums, and the label
     90 // indices. For each minibatch entry, ignoring the batch index b, it calculates:
     91 //
     92 //   exp_logits[j] / sum_exp_logits - 1{ j == label }
     93 //
     94 // for j = 0 .. num_classes.
     95 template <typename T, typename Index>
     96 class SparseXentGradGenerator {
     97  public:
     98   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE SparseXentGradGenerator(
     99       typename TTypes<const T, 2>::Tensor32Bit exp_logits,
    100       typename TTypes<const T, 1>::Tensor32Bit sum_exp_logits,
    101       typename TTypes<const Index, 1>::Tensor32Bit labels,
    102       const Index max_depth)
    103       : exp_logits_(exp_logits),
    104         sum_exp_logits_(sum_exp_logits),
    105         labels_(labels),
    106         max_depth_(max_depth) {}
    107 
    108   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE T
    109   operator()(const Eigen::array<int, 2>& coords) const {
    110     const int batch = coords[0];
    111     const int depth = coords[1];
    112     const Index label = tensorflow::internal::SubtleMustCopy(labels_(batch));
    113     if (!FastBoundsCheck(label, max_depth_)) {
    114       return Eigen::NumTraits<T>::quiet_NaN();
    115     }
    116     T subtract = TF_PREDICT_FALSE(depth == label) ? T(1.0) : T(0.0);
    117     return exp_logits_(coords) / sum_exp_logits_(batch) - subtract;
    118   };
    119 
    120  private:
    121   typename TTypes<const T, 2>::Tensor32Bit exp_logits_;
    122   typename TTypes<const T, 1>::Tensor32Bit sum_exp_logits_;
    123   typename TTypes<const Index, 1>::Tensor32Bit labels_;
    124   const Index max_depth_;
    125 };
    126 
    127 }  // namespace generator
    128 
    129 namespace functor {
    130 
    131 // Functor used by SparseXentOp to do the computations.
    132 template <typename Device, typename T, typename Index>
    133 struct SparseXentFunctor {
    134   // Computes Cross Entropy loss and backprop.
    135   //
    136   // logits: batch_size, num_classes.
    137   // labels: num_classes.
    138   // scratch: temporary tensor, dims: batch_size, 1
    139   // loss: output tensor for the loss, dims: batch_size.
    140   // backprop: output tensor for the backprop, dims: batch_size, num_classes.
    141   void operator()(const Device& d, typename TTypes<T>::ConstMatrix logits,
    142                   typename TTypes<Index>::ConstVec labels,
    143                   typename TTypes<T>::Vec scratch, typename TTypes<T>::Vec loss,
    144                   typename TTypes<T>::Matrix backprop);
    145 };
    146 
    147 // Eigen code implementing SparseXentFunctor::operator().
    148 // This code works for both CPU and GPU and is used by the functor
    149 // specializations for both device types.
    150 template <typename Device, typename T, typename Index>
    151 struct SparseXentEigenImpl {
    152   static void Compute(const Device& d, typename TTypes<T>::ConstMatrix logits,
    153                       typename TTypes<Index>::ConstVec labels,
    154                       typename TTypes<T>::Vec scratch,
    155                       typename TTypes<T>::Vec loss,
    156                       typename TTypes<T>::Matrix backprop) {
    157     // NOTE(touts): This duplicates some of the computations in softmax_op
    158     // because we need the intermediate (logits -max(logits)) values to
    159     // avoid a log(exp()) in the computation of the loss.
    160 
    161     const int kBatchDim = 0;
    162     const int kClassDim = 1;
    163 
    164     const int batch_size = logits.dimension(kBatchDim);
    165     const int num_classes = logits.dimension(kClassDim);
    166 
    167 // These arrays are used to reduce along the class dimension, and broadcast
    168 // the resulting value to all classes.
    169 #if !defined(EIGEN_HAS_INDEX_LIST)
    170     Eigen::array<int, 1> along_class;
    171     along_class[0] = kClassDim;
    172     Eigen::array<int, 1> batch_only;
    173     batch_only[0] = batch_size;
    174     Eigen::array<int, 2> batch_by_one;
    175     batch_by_one[0] = batch_size;
    176     batch_by_one[1] = 1;
    177     Eigen::array<int, 2> one_by_class;
    178     one_by_class[0] = 1;
    179     one_by_class[1] = num_classes;
    180 #else
    181     Eigen::IndexList<Eigen::type2index<kClassDim> > along_class;
    182     Eigen::IndexList<int, Eigen::type2index<1> > batch_by_one;
    183     batch_by_one.set(0, batch_size);
    184     Eigen::IndexList<int> batch_only;
    185     batch_only.set(0, batch_size);
    186     Eigen::IndexList<Eigen::type2index<1>, int> one_by_class;
    187     one_by_class.set(1, num_classes);
    188 #endif
    189 
    190     // scratch = max_logits along classes.
    191     To32Bit(scratch).device(d) = To32Bit(logits).maximum(along_class);
    192 
    193     // backprop = logits - max_logits.
    194     To32Bit(backprop).device(d) =
    195         To32Bit(logits) -
    196         To32Bit(scratch).reshape(batch_by_one).broadcast(one_by_class);
    197 
    198     // scratch = sum(exp(logits - max_logits)) along classes.
    199     To32Bit(scratch).device(d) = To32Bit(backprop).exp().sum(along_class);
    200 
    201     //  sum(-labels *
    202     //     ((logits - max_logits) - log(sum(exp(logits - max_logits)))))
    203     //  along classes
    204     generator::SparseXentLossGenerator<T, Index> sparse_xent_loss_gen(
    205         sparse_xent_helpers::To32BitConst<T>(backprop),
    206         sparse_xent_helpers::To32BitConst<T>(scratch), To32Bit(labels),
    207         backprop.dimension(1) /* max_depth */);
    208     To32Bit(loss).device(d) =
    209         To32Bit(backprop).generate(sparse_xent_loss_gen).sum(along_class);
    210 
    211     // backprop: prob - labels, where
    212     //   prob = exp(logits - max_logits) / sum(exp(logits - max_logits))
    213     To32Bit(backprop).device(d) = To32Bit(backprop).exp();
    214     generator::SparseXentGradGenerator<T, Index> sparse_xent_grad_gen(
    215         sparse_xent_helpers::To32BitConst<T>(backprop),
    216         sparse_xent_helpers::To32BitConst<T>(scratch), To32Bit(labels),
    217         backprop.dimension(1) /* max_depth */);
    218     To32Bit(backprop).device(d) =
    219         To32Bit(backprop).generate(sparse_xent_grad_gen);
    220   }
    221 };
    222 
    223 }  // namespace functor
    224 
    225 }  // namespace tensorflow
    226 
    227 #endif  // TENSORFLOW_KERNELS_XENT_OP_H_
    228