1 /* 2 * Copyright (C) 2018 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 #ifndef NLP_SAFT_COMPONENTS_COMMON_MOBILE_EMBEDDING_NETWORK_H_ 18 #define NLP_SAFT_COMPONENTS_COMMON_MOBILE_EMBEDDING_NETWORK_H_ 19 20 #include <vector> 21 22 #include "lang_id/common/embedding-network-params.h" 23 #include "lang_id/common/fel/feature-extractor.h" 24 25 namespace libtextclassifier3 { 26 namespace mobile { 27 28 // Classifier using a hand-coded feed-forward neural network. 29 // 30 // No gradient computation, just inference. 31 // 32 // Based on the more general nlp_saft::EmbeddingNetwork (without ::mobile). 33 // 34 // Classification works as follows: 35 // 36 // Discrete features -> Embeddings -> Concatenation -> Hidden+ -> Softmax 37 // 38 // In words: given some discrete features, this class extracts the embeddings 39 // for these features, concatenates them, passes them through one or more hidden 40 // layers (each layer uses Relu) and next through a softmax layer that computes 41 // an unnormalized score for each possible class. Note: there is always a 42 // softmax layer at the end. 43 class EmbeddingNetwork { 44 public: 45 // Constructs an embedding network using the parameters from model. 46 // 47 // Note: model should stay alive for at least the lifetime of this 48 // EmbeddingNetwork object. 49 explicit EmbeddingNetwork(const EmbeddingNetworkParams *model); 50 ~EmbeddingNetwork()51 virtual ~EmbeddingNetwork() {} 52 53 // Runs forward computation to fill scores with unnormalized output unit 54 // scores. This is useful for making predictions. 55 void ComputeFinalScores(const std::vector<FeatureVector> &features, 56 std::vector<float> *scores) const; 57 58 // Same as above, but allows specification of extra extra neural network 59 // inputs that will be appended to the embedding vector build from features. 60 void ComputeFinalScores(const std::vector<FeatureVector> &features, 61 const std::vector<float> &extra_inputs, 62 std::vector<float> *scores) const; 63 64 private: 65 // Constructs the concatenated input embedding vector in place in output 66 // vector concat. 67 void ConcatEmbeddings(const std::vector<FeatureVector> &features, 68 std::vector<float> *concat) const; 69 70 // Pointer to the model object passed to the constructor. Not owned. 71 const EmbeddingNetworkParams *model_; 72 73 // Network parameters. 74 75 // One weight matrix for each embedding. 76 std::vector<EmbeddingNetworkParams::Matrix> embedding_matrices_; 77 78 // embedding_row_size_in_bytes_[i] is the size (in bytes) of a row from 79 // embedding_matrices_[i]. We precompute this in order to quickly find the 80 // beginning of the k-th row from an embedding matrix (which is stored in 81 // row-major order). 82 std::vector<int> embedding_row_size_in_bytes_; 83 84 // concat_offset_[i] is the input layer offset for i-th embedding space. 85 std::vector<int> concat_offset_; 86 87 // Size of the input ("concatenation") layer. 88 int concat_layer_size_ = 0; 89 90 // One weight matrix and one vector of bias weights for each layer of neurons. 91 // Last layer is the softmax layer, the previous ones are the hidden layers. 92 std::vector<EmbeddingNetworkParams::Matrix> layer_weights_; 93 std::vector<EmbeddingNetworkParams::Matrix> layer_bias_; 94 }; 95 96 } // namespace mobile 97 } // namespace nlp_saft 98 99 #endif // NLP_SAFT_COMPONENTS_COMMON_MOBILE_EMBEDDING_NETWORK_H_ 100