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 16 #ifndef TENSORFLOW_CORE_KERNELS_LOGISTIC_LOSS_H_ 17 #define TENSORFLOW_CORE_KERNELS_LOGISTIC_LOSS_H_ 18 19 #include <cmath> 20 21 #include "tensorflow/core/kernels/loss.h" 22 #include "tensorflow/core/lib/core/errors.h" 23 24 namespace tensorflow { 25 26 class LogisticLossUpdater : public DualLossUpdater { 27 public: 28 // Adding vs. Averaging in Distributed Primal-Dual Optimization. 29 // Chenxin Ma, Virginia Smith, Martin Jaggi, Michael I. Jordan, Peter 30 // Richtarik, Martin Takac http://arxiv.org/abs/1502.03508 ComputeUpdatedDual(const int num_loss_partitions,const double label,const double example_weight,const double current_dual,const double wx,const double weighted_example_norm)31 double ComputeUpdatedDual(const int num_loss_partitions, const double label, 32 const double example_weight, 33 const double current_dual, const double wx, 34 const double weighted_example_norm) const final { 35 // Newton algorithm converges quadratically so 10 steps will be largely 36 // enough to achieve a very good precision 37 static const int newton_total_steps = 10; 38 double x = 0; 39 for (int i = 0; i < newton_total_steps; ++i) { 40 x = NewtonStep(x, num_loss_partitions, label, wx, example_weight, 41 weighted_example_norm, current_dual); 42 } 43 return 0.5 * (1 + tanh(x)) / label; 44 } 45 46 // Dual of logisitic loss function. 47 // https://en.wikipedia.org/wiki/Convex_conjugate ComputeDualLoss(const double current_dual,const double example_label,const double example_weight)48 double ComputeDualLoss(const double current_dual, const double example_label, 49 const double example_weight) const final { 50 // Dual of the logistic loss function is 51 // ay * log(ay) + (1-ay) * log (1-ay), where a is the dual variable. 52 const double ay = current_dual * example_label; 53 const double log_ay = (ay > 0) ? log(ay) : 0; 54 const double one_minus_ay = 1 - ay; 55 const double log_one_minus_ay = (one_minus_ay > 0) ? log(one_minus_ay) : 0; 56 return ((ay * log_ay) + (one_minus_ay * log_one_minus_ay)) * example_weight; 57 } 58 59 // Logistic loss for binary classification. 60 // https://en.wikipedia.org/wiki/Loss_functions_for_classification ComputePrimalLoss(const double wx,const double example_label,const double example_weight)61 double ComputePrimalLoss(const double wx, const double example_label, 62 const double example_weight) const final { 63 // Logistic loss: 64 // log(1 + e^(-ywx)) 65 // log(e^0 + e^(-ywx)) 66 // a + log(e^(0-a) + e^(-ywx - a)), where a is max(0, -ywx) 67 // https://hips.seas.harvard.edu/blog/2013/01/09/computing-log-sum-exp/ 68 const double y_wx = example_label * wx; 69 if (y_wx > 0) { 70 // 0 + log(e^(0) + e^(-ywx - 0)) 71 // log(1 + e^(-ywx)) 72 return log1p(exp(-y_wx)) * example_weight; 73 } 74 // -ywx + log(e^(ywx) + e^(-ywx + ywx)) 75 // log(e^(ywx) + e^(0)) - ywx 76 // log(1 + e^(ywx)) - ywx 77 return (log1p(exp(y_wx)) - y_wx) * example_weight; 78 } 79 80 // Derivative of logistic loss PrimalLossDerivative(const double wx,const double label,const double example_weight)81 double PrimalLossDerivative(const double wx, const double label, 82 const double example_weight) const final { 83 double inverse_exp_term = 0; 84 if (label * wx > 0) { 85 inverse_exp_term = exp(-label * wx) / (1 + exp(-label * wx)); 86 } else { 87 inverse_exp_term = 1 / (1 + exp(label * wx)); 88 } 89 return -inverse_exp_term * label * example_weight; 90 } 91 92 // The smoothness constant is 4 since the derivative of logistic loss, which 93 // is exp(-x) / (1 + exp(-x)) can be shown to 0.25-Lipschitz (its derivative 94 // is bounded by 0.25) SmoothnessConstant()95 double SmoothnessConstant() const final { return 4; } 96 97 // Converts binary example labels from 0.0 or 1.0 to -1.0 or 1.0 respectively 98 // as expected by logistic regression. ConvertLabel(float * const example_label)99 Status ConvertLabel(float* const example_label) const final { 100 if (*example_label == 0.0) { 101 *example_label = -1; 102 return Status::OK(); 103 } 104 if (*example_label == 1.0) { 105 return Status::OK(); 106 } 107 return errors::InvalidArgument( 108 "Only labels of 0.0 or 1.0 are supported right now. " 109 "Found example with label: ", 110 *example_label); 111 } 112 113 private: 114 // We use Newton algorithm on a modified function (see readme.md). NewtonStep(const double x,const int num_loss_partitions,const double label,const double wx,const double example_weight,const double weighted_example_norm,const double current_dual)115 double NewtonStep(const double x, const int num_loss_partitions, 116 const double label, const double wx, 117 const double example_weight, 118 const double weighted_example_norm, 119 const double current_dual) const { 120 const double tanhx = tanh(x); 121 const double numerator = -2 * label * x - wx - 122 num_loss_partitions * weighted_example_norm * 123 example_weight * 124 (0.5 * (1 + tanhx) / label - current_dual); 125 const double denominator = 126 -2 * label - num_loss_partitions * weighted_example_norm * 127 example_weight * (1 - tanhx * tanhx) * 0.5 / label; 128 return x - numerator / denominator; 129 } 130 }; 131 132 } // namespace tensorflow 133 134 #endif // TENSORFLOW_CORE_KERNELS_LOGISTIC_LOSS_H_ 135