1 /* 2 * Copyright (C) 2017 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 #include "common/softmax.h" 18 19 #include <limits> 20 21 #include "common/fastexp.h" 22 #include "util/base/logging.h" 23 24 namespace libtextclassifier { 25 namespace nlp_core { 26 27 float ComputeSoftmaxProbability(const std::vector<float> &scores, int label) { 28 if ((label < 0) || (label >= scores.size())) { 29 TC_LOG(ERROR) << "label " << label << " outside range " 30 << "[0, " << scores.size() << ")"; 31 return 0.0f; 32 } 33 34 // Standard softmax formula for label's probability is 35 // 36 // exp(scores[label]) / sum_i exp(scores[i]) 37 // 38 // We compute the mathematically equivalent 39 // 40 // 1 / (1 + sum_{i != label} exp(scores[i] - scores[label])) 41 // 42 // which saves two calls to exp(). 43 const float label_score = scores[label]; 44 float denominator = 1.0f; // Contribution of i == label. 45 for (int i = 0; i < scores.size(); ++i) { 46 if (i == label) continue; 47 const float delta_score = scores[i] - label_score; 48 49 // TODO(salcianu): one can optimize the test below, to avoid any float 50 // operation: extract exponent (via bit mask + shift) and check it's >= 4. 51 if (fabs(delta_score) >= 16.0f) { 52 if (delta_score > 0.0f) { 53 // If delta_score >= 16, the denominator (e^delta_score + other positive 54 // terms) is very big and its inverse can be approximated with 0. 55 return 0.0f; 56 } else { 57 // If delta_score <= -16, then e^delta_score < 1.2e-7. Even if we have 58 // 1000 such labels i, their sum is < 1.2e-4 (which gets summed with 59 // 1.0f for i == label). Hence, we can approximate each such label with 60 // 0 and skip the call to VeryFastExp and the update to denominator. 61 continue; 62 } 63 } 64 65 // At this point, delta_score is in (-16.0, 16.0). For such values, vfexp 66 // works fine: no under/overflows (we have tests for that in fastexp_test). 67 // Also, even for 1000 labels, denominator will not overflow. 68 denominator += VeryFastExp(delta_score); 69 } 70 return 1.0f / denominator; 71 } 72 73 std::vector<float> ComputeSoftmax(const std::vector<float> &scores) { 74 std::vector<float> softmax; 75 std::vector<float> exp_scores; 76 exp_scores.reserve(scores.size()); 77 softmax.reserve(scores.size()); 78 79 // Find max value in "scores" vector and rescale to avoid overflows. 80 float max = std::numeric_limits<float>::min(); 81 for (const auto &score : scores) { 82 if (score > max) max = score; 83 } 84 float denominator = 0; 85 for (auto &score : scores) { 86 // See comments above in ComputeSoftmaxProbability for the reasoning behind 87 // this approximation. 88 const float exp_score = score - max < -16.0f ? 0 : VeryFastExp(score - max); 89 exp_scores.push_back(exp_score); 90 denominator += exp_score; 91 } 92 93 for (int i = 0; i < scores.size(); ++i) { 94 softmax.push_back(exp_scores[i] / denominator); 95 } 96 return softmax; 97 } 98 99 } // namespace nlp_core 100 } // namespace libtextclassifier 101
