/*
 * To change this license header, choose License Headers in Project Properties.
 * To change this template file, choose Tools | Templates
 * and open the template in the editor.
 */

/* 
 * File:   Rbm.cpp
 * Author: jens
 * 
 * Created on 21. Oktober 2019, 21:28
 */

#include <cassert>
#include "Rbm.hpp"
#include "matutils.hpp"

using namespace Matutils;

Rbm::Rbm(size_t numVisible, size_t numHidden)
: m_params()
, m_whv(numVisible, numHidden)
, m_bhv(1, numHidden)
, m_bv(1, numVisible)
{
	assert(numVisible > 0);
	assert(numHidden > 0);
}

Rbm::Rbm(const Rbm& orig)
: m_params(orig.m_params)
, m_whv(orig.m_whv)
, m_bhv(orig.m_bhv)
, m_bv(orig.m_bv)
{
}

Rbm::~Rbm()
{
}

size_t Rbm::numHidden() const
{
	return m_bhv.size();
}

size_t Rbm::numVisible() const
{
	return m_bv.size();
}

Rbm::Params& Rbm::params()
{
	return m_params;
}

arma::mat Rbm::rms_error(arma::mat diffErr)
{
	arma::mat diffErr_squared = diffErr % diffErr;
	return arma::sum(diffErr_squared, 1) * 1.0 / diffErr_squared.n_cols;
}

double Rbm::rms_error_accu(arma::mat diffErr)
{
	arma::mat diffErr_squared = diffErr % diffErr;
	return arma::accu(diffErr_squared) / diffErr_squared.n_elem;
}

arma::mat Rbm::toHiddenProbs(const arma::mat& visible) const
{
	return Rbm::prob(v_to_h(visible));
}

arma::mat Rbm::toVisibleProbs(const arma::mat& hidden) const
{
	return Rbm::prob(h_to_v(hidden));
}

void Rbm::weightsAssign(const arma::mat& w, const arma::mat& bhv, const arma::mat& bv)
{
	m_whv.submat(0, 0, w.n_rows - 1, w.n_cols - 1) = w;
	m_bhv.submat(0, 0, bhv.n_rows - 1, bhv.n_cols - 1) = bhv;
	m_bv.submat(0, 0, bv.n_rows - 1, bv.n_cols - 1) = bv;
}

void Rbm::weightsInit(double stddev, double mu)
{
	uniform(m_whv, stddev, mu);
	uniform(m_bhv, stddev, mu);
	uniform(m_bv, stddev, mu);
}

void Rbm::fromJson(Json::Value rbm)
{
	std::cout << "Importing Rbm" << std::endl;
	m_params.fromJson(rbm["params"]);
}

Json::Value Rbm::toJson() const
{
	std::cout << "Exporting Rbm" << std::endl;
	Json::Value rbm;
	rbm["params"] = m_params.toJson();
	return rbm;
}

void Rbm::gibbs_vh(arma::mat &v_probs, arma::mat &h_probs) const
{
	for (int i=0; i < m_params.numGibbs; i++)
	{
		// Create hidden representation given v
		h_probs = prob(v_to_h(v_probs));

		// Create visible reconstruction (a fantasy...) given hid
		v_probs = prob(h_to_v(h_probs));
	}
}

void Rbm::gibbs_hv(arma::mat &h_probs, arma::mat &v_probs) const
{
	for (int i=0; i < m_params.numGibbs; i++)
	{
		// Create visible reconstruction (a fantasy...) given hid
		v_probs = prob(h_to_v(h_probs));

		// Create hidden representation given v
		h_probs = prob(v_to_h(v_probs));
	}
}

void Rbm::contrastiveDivergence(arma::mat const &v_states, arma::mat &dw, arma::mat &dbhv, arma::mat &dbv)
{
	arma::mat v_probs(v_states);
	arma::mat h_states = v_to_h(v_states);
	arma::mat h_probs;
	
	// Sample hidden
	if (m_params.doGaussianHidden)
	{
		h_probs = h_states;
		h_states = h_probs + arma::randn(h_probs.n_rows, h_probs.n_cols);
	}
	else if (m_params.doRaoBlackwell)
	{
		h_probs = prob(v_to_h(v_states));
		h_states = h_probs;
	}
	else
	{
		h_probs = prob(v_to_h(v_states));
		h_states = sample(h_probs);
	}
	
	// Update weights (positive phase)
	dw = v_states.t() * h_states;
	dbv = sum(v_states, 0);
	dbhv = sum(h_states, 0);

	// Gibbs sampling with training params
	for (int i=0; i < m_params.numGibbs; i++)
	{
		// Create visible reconstruction (a fantasy...) given hid
		if (m_params.gibbsDoSampleHidden)
		{
			v_probs = prob(h_to_v(sample(h_probs)));
		}
		else
		{
			v_probs = prob(h_to_v(h_probs));
		}

		// Create hidden representation given v
		if (m_params.doGaussianHidden)
		{
			h_probs = v_to_h(v_probs);
		}
		else if (m_params.gibbsDoSampleVisible)
		{
			h_probs = prob(v_to_h(sample(v_probs)));
		}
		else
		{
			h_probs = prob(v_to_h(v_probs));
		}
	}

	// Update weights (negative phase)
	dw -= v_probs.t() * h_probs;
	dbv -= sum(v_probs, 0);
	dbhv -= sum(h_probs, 0);
}

void Rbm::train(arma::mat const &batch, IListener* pListener)
{
	Status status;
	double dProgress = 100.0/(batch.n_rows*m_params.numEpochs);
	double progress = 0;
	int lastProgress = -100;
	int batchRowIndex = 0;
	
	arma::mat grad_bias_v(arma::zeros(1, m_bv.n_cols));
	arma::mat grad_bias_hv(arma::zeros(1, m_bhv.n_cols));
	arma::mat grad_weight_hv(arma::zeros(m_whv.n_rows, m_whv.n_cols));
	arma::mat momentum_whv = arma::zeros(m_whv.n_rows, m_whv.n_cols);
	arma::mat momentum_bias_v(arma::zeros(1, m_bv.n_cols));
	arma::mat momentum_bias_hv(arma::zeros(1, m_bhv.n_cols));
	arma::mat penalty_weights = arma::zeros(m_whv.n_rows, m_whv.n_cols);

	int trainingSizeRemain = batch.n_rows;

	bool shouldAbort = false;
	while (trainingSizeRemain && !shouldAbort)
	{
		int miniBatchSizeActual = std::min(m_params.miniBatchSize, trainingSizeRemain);
		arma::mat miniBatch = batch.rows(batchRowIndex, batchRowIndex+miniBatchSizeActual-1);
		trainingSizeRemain -= miniBatchSizeActual;
		batchRowIndex += miniBatchSizeActual;
		int scaler = std::min(m_params.miniBatchSize, (int)batch.n_rows);
		double learning_rate = m_params.learningRate/scaler;
		double weight_decay = m_params.weightDecay/scaler;

		arma::mat v_states(miniBatch);

		// Create hidden layer base on training data
		if (m_params.doSampleBatch)
		{
			// When the hidden units are being driven by data, always use stochastic binary states
			v_states = sample(miniBatch);
		}
		
		for (int epoch=0; epoch < m_params.numEpochs; epoch++)
		{
			// Contrastive divergence learning: calculate gradients
			contrastiveDivergence(v_states, grad_weight_hv, grad_bias_hv, grad_bias_v);
			
			// Adjust weight and biases
			penalty_weights = weight_decay*arma::sign(m_whv);
			
			status.L1 = accu(abs(m_whv));
			status.L2 = accu(m_whv % m_whv);
			momentum_bias_v = m_params.momentum*momentum_bias_v + grad_bias_v;
			momentum_bias_hv = m_params.momentum*momentum_bias_hv + grad_bias_hv;
			momentum_whv = m_params.momentum*momentum_whv + grad_weight_hv - status.L2*penalty_weights;
			
			m_bv += learning_rate*momentum_bias_v;
			m_bhv += learning_rate*momentum_bias_hv;
			m_whv += learning_rate*momentum_whv;
			
			progress += dProgress*miniBatchSizeActual;
			status.progress = (int)(progress + 0.5);

			// Update status
			if (status.progress != lastProgress)
			{
				lastProgress = status.progress;
				
				// Calculate error
				status.err = rms_error_accu(miniBatch - prob(h_to_v(prob(v_to_h(v_states)))));
				if (pListener)
				{
					if(!pListener->onProgress(this, status))
					{
						shouldAbort = true;
						break;
					}
				}
			}
			
		} // Number of epochs
			
	} // number of mini batches

	// Update final status
	status.err_total = rms_error_accu(batch - prob(h_to_v(prob(v_to_h(batch)))));

	if (pListener)
	{
		pListener->onProgress(this, status);
	}
}

arma::mat Rbm::prob(const arma::mat &src)
{
	return 1 / (1 + (arma::exp(-src)));
}

arma::mat Rbm::v_to_h(const arma::mat &visible) const
{
	return visible * m_whv  + arma::repmat(m_bhv, visible.n_rows, 1);
}

arma::mat Rbm::h_to_v(const arma::mat &hidden) const
{
	return hidden * m_whv.t() + arma::repmat(m_bv, hidden.n_rows, 1);
}

const arma::mat& Rbm::whv() const
{
	return m_whv;
}

const arma::mat& Rbm::bv() const
{
	return m_bv;
}

const arma::mat& Rbm::bh() const
{
	return m_bhv;
}