EmbeddingGenerator.h

#pragma once
#include <MedProcessTools/MedProcessTools/FeatureGenerator.h>
#include <MedEmbed/MedEmbed/MedEmbed.h>
#include <MedEmbed/MedEmbed/ApplyKeras.h>
#include <MedSparseMat/MedSparseMat/MedSparseMat.h>
 
//
// Allow the usage of a trained embedding to generate features
// needed are the .scheme and .layers file created using the Embedding app and Keras script.
// This version only APPLIES and does not LEARN the embedding.
//
 
class EmbeddingGenerator : public FeatureGenerator {
public:
 
    // Feature Descrption
    string signalName;
    int signalId;
 
    // parameters
    string f_scheme = "";
    string f_layers = "";
 
    // serialized parameters
    string name_prefix = "Embedding"; // used when naming the features, names will be FTR_<i>.<name_prefix>.col_<j>
    int to_layer = 9; // default embedding layer in our keras generated models
 
    // basic embedding structures, serialized (init from scheme, layers files upon creation from init() )
    EmbedMatCreator emc;
    ApplyKeras embedder;
 
    // helper variables
    int e_dim = 0; // we get it from embedder, but serialize it for simpler code
 
 
    // Signal to determine allowed time-range (e.g. current stay/admission for inpatients)
 
                                                        // Constructor/Destructor
    EmbeddingGenerator() : FeatureGenerator() {};
    ~EmbeddingGenerator() {};
 
    int init(map<string, string>& mapper);
    //void init_defaults();
 
    // Naming
    void set_names();
 
    // preparing a batch result of model results if an internal MedModel generator is used
    void prepare(MedFeatures & features, MedPidRepository& rep, MedSamples& samples);
 
    // Learn a generator
    int _learn(MedPidRepository& rep, const MedSamples& samples, vector<RepProcessor *> processors);
 
    // generate a new feature
    int _generate(PidDynamicRec& rec, MedFeatures& features, int index, int num, vector<float *> &_p_data);
 
    int generate_by_rec(PidDynamicRec& rec, MedFeatures& features, int index, int num, vector<float *> &_p_data);
 
    // Signal Ids
    void set_signal_ids(MedSignals& sigs);
 
    // Init required tables
    void init_tables(MedDictionarySections& dict);
 
    void get_required_signal_categories(unordered_map<string, vector<string>> &signal_categories_in_use) const;
 
    // in case of selection
    int filter_features(unordered_set<string>& validFeatures) { return 1; }; // TODO: improve to push only selected columns
 
    // Serialization
    ADD_CLASS_NAME(EmbeddingGenerator)
    ADD_SERIALIZATION_FUNCS(generator_type, tags, names, req_signals, signalName, iGenerateWeights, name_prefix, to_layer, e_dim, emc, embedder)
};
 
MEDSERIALIZE_SUPPORT(EmbeddingGenerator);