negtrain.h

#ifndef NEGTRAIN_H_
#define NEGTRAIN_H_

#include <random>

#include "negsamplingbase.h"

class NegTrain : public NegSamplingBase
{
public:
	static float *GetInitedCMParams(int vec_dim);

	// for energy with a matrix
	// probably not used
	static void InitMatrix(float *matrix, int dim0, int dim1);

	static void CloseVectors(float **vecs, int num_vecs, int vec_dim, int idx);

public:
	// use objs0 to predict objs1
	// e.g. objs0: documents, objs1: words
	//NegativeSamplingTrainer(ExpTable *exp_table, int num_negative_samples, int num_objs1,
	//	std::discrete_distribution<int> *negative_sample_dist);
	NegTrain(ExpTable *exp_table, int num_negative_samples, int num_objs1,
		int *obj_cnts);

	NegTrain(ExpTable *exp_table, int num_negative_samples,
		const char *freq_file);

	//NegativeSamplingTrainer(ExpTable *exp_table, int vec_dim, int num_objs, int num_negative_samples,
	//	std::discrete_distribution<int> *obj_sample_dist);

	~NegTrain();

	// obj0 -> obj1
	void TrainPair(int vec_dim, float *vec0, int obj1, float **vecs1, float alpha, float *tmp_neu1e,
		std::default_random_engine &generator, float gamma, bool update0 = true, bool update1 = true);

	// controled mix
	// dimention of vec0: vec_dim * 2
	void TrainPairCM(int vec_dim, float *vec0, int obj1, float **vecs1, float *cm_params, bool complement,
		float alpha, float *tmp_neu1e, float *tmp_cme, std::default_random_engine &generator,
		bool update0 = true, bool update1 = true, bool update_cm_params = true)
	{
		if (complement)
		{
			trainPairCMComplement(vec_dim, vec0, obj1, vecs1, cm_params, alpha, tmp_neu1e,
				tmp_cme, generator, update0, update1, update_cm_params);
		}
		else
		{
			trainPairCM(vec_dim, vec0, obj1, vecs1, cm_params, alpha, tmp_neu1e,
				tmp_cme, generator, update0, update1, update_cm_params);
		}

		for (int i = 0; i < vec_dim; ++i)
		{
			if (cm_params[i] < 0)
				cm_params[i] = 0.01f;
			else if (cm_params[i] > 1)
				cm_params[i] = 0.99f;
		}
	}

	void TrainPairMatrix(int dim0, int dim1, float *vec0, int obj1, float **vecs1, float *matrix, float alpha, float *tmp_neu1e,
		std::default_random_engine &generator, bool update0 = true, bool update1 = true, bool update_matrix = true);

	void CheckObject(int vec_dim, float *cur_vec, float **vecs1);

private:
	void trainPairCM(int vec_dim, float *vec0, int obj1, float **vecs1, float *cm_params,
		float alpha, float *tmp_neu1e, float *tmp_cme, std::default_random_engine &generator,
		bool update0, bool update1, bool update_cm_params);
	void trainPairCMComplement(int vec_dim, float *vec0, int obj1, float **vecs1, float *cm_params,
		float alpha, float *tmp_neu1e, float *tmp_cme, std::default_random_engine &generator,
		bool update0, bool update1, bool update_cm_params);

	float calcCMEnergy(int vec_dim, float *vec0, float *vec1, float *cm_params)
	{
		float result = 0;
		for (int i = 0; i < vec_dim; ++i)
		{
			int tmp = i << 1;
			result += vec1[i] * (vec0[tmp] * cm_params[i] + vec0[tmp + 1] * (1 - cm_params[i]));
		}
		return result;
	}

	float calcCMEnergyComplement(int vec_dim, float *vec0, float *vec1, float *cm_params)
	{
		float result = 0;
		for (int i = 0; i < vec_dim; ++i)
		{
			int tmp = i << 1;
			result += vec1[i] * (vec0[tmp] * (1 - cm_params[i]) + vec0[tmp + 1] * cm_params[i]);
		}
		return result;
	}

private:
	// use objs0 to predict objs1, e.g. objs0: documents, objs1: words
	int num_objs1_ = 0;

	std::discrete_distribution<int> negative_sample_dist_;
};

#endif