1 /////////////////////////////////////////////////////////////////////// 2 // File: fullyconnected.h 3 // Description: Simple feed-forward layer with various non-linearities. 4 // Author: Ray Smith 5 // 6 // (C) Copyright 2014, Google Inc. 7 // Licensed under the Apache License, Version 2.0 (the "License"); 8 // you may not use this file except in compliance with the License. 9 // You may obtain a copy of the License at 10 // http://www.apache.org/licenses/LICENSE-2.0 11 // Unless required by applicable law or agreed to in writing, software 12 // distributed under the License is distributed on an "AS IS" BASIS, 13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 14 // See the License for the specific language governing permissions and 15 // limitations under the License. 16 /////////////////////////////////////////////////////////////////////// 17 18 #ifndef TESSERACT_LSTM_FULLYCONNECTED_H_ 19 #define TESSERACT_LSTM_FULLYCONNECTED_H_ 20 21 #include "network.h" 22 #include "networkscratch.h" 23 #include "tesstypes.h" 24 25 namespace tesseract { 26 27 // C++ Implementation of the Softmax (output) class from lstm.py. 28 class FullyConnected : public Network { 29 public: 30 TESS_API 31 FullyConnected(const std::string &name, int ni, int no, NetworkType type); 32 ~FullyConnected() override = default; 33 34 // Returns the shape output from the network given an input shape (which may 35 // be partially unknown ie zero). 36 StaticShape OutputShape(const StaticShape &input_shape) const override; 37 spec()38 std::string spec() const override { 39 std::string spec; 40 if (type_ == NT_TANH) { 41 spec += "Ft" + std::to_string(no_); 42 } else if (type_ == NT_LOGISTIC) { 43 spec += "Fs" + std::to_string(no_); 44 } else if (type_ == NT_RELU) { 45 spec += "Fr" + std::to_string(no_); 46 } else if (type_ == NT_LINEAR) { 47 spec += "Fl" + std::to_string(no_); 48 } else if (type_ == NT_POSCLIP) { 49 spec += "Fp" + std::to_string(no_); 50 } else if (type_ == NT_SYMCLIP) { 51 spec += "Fn" + std::to_string(no_); 52 } else if (type_ == NT_SOFTMAX) { 53 spec += "Fc" + std::to_string(no_); 54 } else { 55 spec += "Fm" + std::to_string(no_); 56 } 57 return spec; 58 } 59 60 // Changes the type to the given type. Used to commute a softmax to a 61 // non-output type for adding on other networks. ChangeType(NetworkType type)62 void ChangeType(NetworkType type) { 63 type_ = type; 64 } 65 66 // Suspends/Enables training by setting the training_ flag. Serialize and 67 // DeSerialize only operate on the run-time data if state is false. 68 void SetEnableTraining(TrainingState state) override; 69 70 // Sets up the network for training. Initializes weights using weights of 71 // scale `range` picked according to the random number generator `randomizer`. 72 int InitWeights(float range, TRand *randomizer) override; 73 // Recursively searches the network for softmaxes with old_no outputs, 74 // and remaps their outputs according to code_map. See network.h for details. 75 int RemapOutputs(int old_no, const std::vector<int> &code_map) override; 76 77 // Converts a float network to an int network. 78 void ConvertToInt() override; 79 80 // Provides debug output on the weights. 81 void DebugWeights() override; 82 83 // Writes to the given file. Returns false in case of error. 84 bool Serialize(TFile *fp) const override; 85 // Reads from the given file. Returns false in case of error. 86 bool DeSerialize(TFile *fp) override; 87 88 // Runs forward propagation of activations on the input line. 89 // See Network for a detailed discussion of the arguments. 90 void Forward(bool debug, const NetworkIO &input, const TransposedArray *input_transpose, 91 NetworkScratch *scratch, NetworkIO *output) override; 92 // Components of Forward so FullyConnected can be reused inside LSTM. 93 void SetupForward(const NetworkIO &input, const TransposedArray *input_transpose); 94 void ForwardTimeStep(int t, TFloat *output_line); 95 void ForwardTimeStep(const TFloat *d_input, int t, TFloat *output_line); 96 void ForwardTimeStep(const int8_t *i_input, int t, TFloat *output_line); 97 98 // Runs backward propagation of errors on the deltas line. 99 // See Network for a detailed discussion of the arguments. 100 bool Backward(bool debug, const NetworkIO &fwd_deltas, NetworkScratch *scratch, 101 NetworkIO *back_deltas) override; 102 // Components of Backward so FullyConnected can be reused inside LSTM. 103 void BackwardTimeStep(const NetworkIO &fwd_deltas, int t, TFloat *curr_errors, 104 TransposedArray *errors_t, TFloat *backprop); 105 void FinishBackward(const TransposedArray &errors_t); 106 107 // Updates the weights using the given learning rate, momentum and adam_beta. 108 // num_samples is used in the adam computation iff use_adam_ is true. 109 void Update(float learning_rate, float momentum, float adam_beta, int num_samples) override; 110 // Sums the products of weight updates in *this and other, splitting into 111 // positive (same direction) in *same and negative (different direction) in 112 // *changed. 113 void CountAlternators(const Network &other, TFloat *same, TFloat *changed) const override; 114 115 protected: 116 // Weight arrays of size [no, ni + 1]. 117 WeightMatrix weights_; 118 // Transposed copy of input used during training of size [ni, width]. 119 TransposedArray source_t_; 120 // Pointer to transposed input stored elsewhere. If not null, this is used 121 // in preference to calculating the transpose and storing it in source_t_. 122 const TransposedArray *external_source_; 123 // Activations from forward pass of size [width, no]. 124 NetworkIO acts_; 125 // Memory of the integer mode input to forward as softmax always outputs 126 // float, so the information is otherwise lost. 127 bool int_mode_; 128 }; 129 130 } // namespace tesseract. 131 132 #endif // TESSERACT_LSTM_FULLYCONNECTED_H_ 133