ApplyKeras.h

#pragma once
 
//
// A simple class to enable:
// (1) reading a (simple) keras model from a text file (prepared in the training scripts in the print_layers function)
// (2) Applying it to a given sparse mat, or a line to get the embedding layer (currently always one before the last)
// (3) serialize/deserialize it via MedSerialize (to enable upper classes put it inside a model)
//
// Currently only supports : 
// (1) dense (fully connected, with bias) , with linear or relu or sigmoid activation
// (2) LeakyReLU activation (typically right AFTER a dense layer)
// (3) Dropout (nothing to do, may need to multiply by a factor)
// (4) batch normalization
// (5) Softmax or Sigmoid
//
 
#include <SerializableObject/SerializableObject/SerializableObject.h>
#include <MedSparseMat/MedSparseMat/MedSparseMat.h>
#include <MedMat/MedMat/MedMat.h>
#include <External/Eigen/Core>
 
//===================================================================================================
 
enum KerasLayerTypes { K_UNKNOWN = 0, K_DENSE, K_LEAKY, K_DROPOUT, K_BN , K_ACTIVATION};
enum KerasActivations { A_UNKNOWN = 0, A_LINEAR, A_SIGMOID , A_RELU, A_LEAKY, A_SOFTMAX};
 
//===================================================================================================
class KerasLayer : public SerializableObject {
 
public:
 
    int type = K_UNKNOWN;
    string name = "";
 
    // for dense
    int in_dim = 0;
    int out_dim = 0;
    int n_bias = 0;
    int activation = A_UNKNOWN;
 
    // bn
    int dim;
 
    // dropout
    float drop_rate;
 
    // leaky
    float leaky_alpha;
 
    // weights (holding also transposed version for ease of use with Eigen)
    MedMat<float> wgts, twgts;
 
    // biases
    vector<float> bias;
 
    // helpers
    unordered_map<string, int> name_to_type = { { "dense" , K_DENSE } ,{ "leaky", K_LEAKY } ,{ "dropout" , K_DROPOUT },{ "batch_normalization" , K_BN } , {"activation", K_ACTIVATION} };
    unordered_map<string, int> name_to_activation = { { "linear" , A_LINEAR } ,{ "relu", A_RELU } ,{ "leaky" , A_LEAKY },{ "sigmoid" , A_SIGMOID }, {"softmax", A_SOFTMAX} };
 
 
    // appliers for a single sample
    int apply_sparse(vector<pair<int, float>> &sline, vector<float> &output) const;
    int apply_sparse(map<int, float> &sline, vector<float> &output) const;
    int apply(vector<float> &in, vector<float> &out) const;
    int apply_bn(vector<float> &in, vector<float> &out) const;
    int apply_activation(vector<float> &in, vector<float> &out) const; // for in place send same vector for in/out
 
    // appliers for a batch of samples
    int apply(MedMat<float> &in, MedMat<float>& out) const;
    int apply_bn(MedMat<float> &in, MedMat<float> &out) const;
    int apply_activation(MedMat<float> &in, MedMat<float> &out) const; // for in place send same vector for in/out
 
    // initialization from string
    int init(map<string, string>& _map);
 
    ADD_CLASS_NAME(KerasLayer)
    ADD_SERIALIZATION_FUNCS(type, name, in_dim, out_dim, n_bias, activation, dim, drop_rate, leaky_alpha, wgts, twgts, bias)
};
 
 
//===================================================================================================
class ApplyKeras : public SerializableObject {
 
public:
    vector<KerasLayer> layers;
 
    int init_from_text_file(string layers_file);
 
    int apply_sparse(vector<pair<int, float>> &sline, vector<float> &output, int to_layer) const;
    int apply_sparse(map<int, float> &sline, vector<float> &output, int to_layer) const;
    int apply(vector<float>& line, vector<float> &output, int to_layer) const;
    int apply(vector<float>& line, vector<float> &output) const { return apply(line, output, (int)(layers.size() - 1)); }
    int apply(MedMat<float> &line, MedMat<float>& output, int to_layer) const;
    int apply(MedMat<float>& line, MedMat<float> &output) const { return apply(line, output, (int)(layers.size() - 1)); }
 
    int get_all_embeddings(MedSparseMat &smat, int to_layer, MedMat<float> &emat);
 
    int get_output_dimension() {
        if (layers.size() > 0)
            return layers.back().out_dim;
        return 0;
    }
 
    ADD_CLASS_NAME(ApplyKeras)
    ADD_SERIALIZATION_FUNCS(layers)
};
 
MEDSERIALIZE_SUPPORT(KerasLayer)
MEDSERIALIZE_SUPPORT(ApplyKeras)