SSAGES::ANN Class Reference

Artificial Neural Network Method. More...

#include <ANN.h>

Inheritance diagram for SSAGES::ANN:

Public Member Functions
	ANN (const MPI_Comm &world, const MPI_Comm &comm, const Eigen::VectorXi &topol, Grid< Eigen::VectorXd > *fgrid, Grid< unsigned int > *hgrid, Grid< double > *ugrid, const std::vector< double > &lowerb, const std::vector< double > &upperb, const std::vector< double > &lowerk, const std::vector< double > &upperk, double temperature, double weight, unsigned int nsweep)
	Constructor. More...
void	PreSimulation (Snapshot *snapshot, const class CVManager &cvmanager) override
	Method call prior to simulation initiation. More...
void	PostIntegration (Snapshot *snapshot, const class CVManager &cvmanager) override
	Method call post integration. More...
void	PostSimulation (Snapshot *snapshot, const class CVManager &cvmanager) override
	Method call post simulation. More...
void	SetPrevWeight (double h)
	Set previous history weight. More...
void	SetOutput (const std::string &outfile)
	Set name of output file. More...
void	SetOutputOverwrite (bool overwrite)
	Set overwrite flag on output file. More...
void	SetConvergeIters (unsigned int citers)
	Set number of iterations after which we turn on full weight. More...
void	SetMaxIters (unsigned int iters)
	Set maximum number of training iterations per sweep. More...
void	SetMinLoss (double loss)
	Set minimum loss function value (should be zero for production). More...
void	ReadBias (const std::string &, const std::string &)
	Load network state and bias from file.
Public Member Functions inherited from SSAGES::Method
	Method (unsigned int frequency, const MPI_Comm &world, const MPI_Comm &comm)
	Constructor. More...
void	SetCVMask (const std::vector< unsigned int > &mask)
	Sets the collective variable mask. More...
virtual	~Method ()
	Destructor.
Public Member Functions inherited from SSAGES::EventListener
	EventListener (unsigned int frequency)
	Constructor. More...
unsigned int	GetFrequency () const
	Get frequency of event listener. More...
virtual	~EventListener ()
	Destructor.

Static Public Member Functions
static ANN *	Build (const Json::Value &json, const MPI_Comm &world, const MPI_Comm &comm, const std::string &path)
	Build a derived method from JSON node. More...
Static Public Member Functions inherited from SSAGES::Method
static Method *	BuildMethod (const Json::Value &json, const MPI_Comm &world, const MPI_Comm &comm, const std::string &path)
	Build a derived method from JSON node. More...
Static Public Member Functions inherited from SSAGES::EventListener
static unsigned int	GetWalkerID (const MPI_Comm &world, const MPI_Comm &comm)
	Get walker ID number of specified communicator. More...
static unsigned int	GetNumWalkers (const MPI_Comm &world, const MPI_Comm &comm)
	Get total number of walkers in the simulation. More...
static bool	IsMasterRank (const MPI_Comm &comm)
	Check if current processor is master. More...

Private Member Functions
void	TrainNetwork ()
	Trains the neural network.
void	WriteBias ()
	Writes out the bias to file.

Private Attributes
Eigen::VectorXi	topol_
	Neural network topology.
unsigned int	citers_
	Number of iterations after which we turn on full weight.
nnet::neural_net	net_
	Neural network.
Grid< Eigen::VectorXd > *	fgrid_
	Force grid.
Grid< unsigned int > *	hgrid_
	Histogram grid.
Grid< double > *	ugrid_
	Unbiased histogram grid.
std::string	outfile_
	Output filename.
bool	preloaded_
	Is the network preloaded?
bool	overwrite_
	Overwrite outputs?
unsigned int	sweep_
unsigned int	nsweep_
double	pweight_
double	weight_
double	temp_
double	kbt_
Eigen::MatrixXd	hist_
Eigen::MatrixXd	bias_
Eigen::MatrixXd	rbias_
std::vector< double >	lowerb_
std::vector< double >	upperb_
std::vector< double >	lowerk_
std::vector< double >	upperk_

Additional Inherited Members
Protected Attributes inherited from SSAGES::Method
mxx::comm	world_
	Global MPI communicator.
mxx::comm	comm_
	Local MPI communicator.
std::vector< unsigned int >	cvmask_
	Mask which identifies which CVs to act on.

Detailed Description

Artificial Neural Network Method.

Implementation of the Artificial Neural Network Method based on [3]

Definition at line 35 of file ANN.h.

Constructor & Destructor Documentation

ANN()

SSAGES::ANN::ANN	(	const MPI_Comm &	world,
		const MPI_Comm &	comm,
		const Eigen::VectorXi &	topol,
		Grid< Eigen::VectorXd > *	fgrid,
		Grid< unsigned int > *	hgrid,
		Grid< double > *	ugrid,
		const std::vector< double > &	lowerb,
		const std::vector< double > &	upperb,
		const std::vector< double > &	lowerk,
		const std::vector< double > &	upperk,
		double	temperature,
		double	weight,
		unsigned int	nsweep
	)

Constructor.

Parameters

world	MPI global communicator.
comm	MPI local communicator.
topol	Topology of network.
fgrid	Grid containing biasing forces.
hgrid	Grid containing histogram.
ugrid	Grid containing unbiased histogram.
lowerb	Lower bounds for CVs.
upperb	Upper bounds for CVs.
lowerk	Lower bound restraints for CVs.
upperk	Upper bound restraints for CVs.
temperature	Temperature of the simulation.
weight	Relative weight of the statistics in sweep.
nsweep	Number of iterations in the sweep.

Constructs an instance of Artificial Neural Network method.

Definition at line 34 of file ANN.cpp.

                                  :
    Method(1, world, comm), topol_(topol), sweep_(0), nsweep_(nsweep),
    citers_(0), net_(topol), pweight_(1.), weight_(weight), temp_(temperature),
    kbt_(0), fgrid_(fgrid), hgrid_(hgrid), ugrid_(ugrid), hist_(), bias_(),
    lowerb_(lowerb), upperb_(upperb), lowerk_(lowerk), upperk_(upperk),
    outfile_("ann.out"), preloaded_(false), overwrite_(true)
    {
        // Create histogram grid matrix.
        hist_.resize(hgrid_->size(), hgrid_->GetDimension());
 
        // Fill it up.
        size_t i = 0;
        for(auto it = hgrid_->begin(); it != hgrid_->end(); ++it)
        {
            auto coord = it.coordinates();
            for(size_t j = 0; j < coord.size(); ++j)
                hist_(i, j) = coord[j];
            ++i;
        }
 
        // Initialize restraint reweight vector.
        rbias_.resize(hgrid_->size(), hgrid_->GetDimension());
        rbias_.fill(0);
 
        // Initialize FES vector.
        bias_.resize(hgrid_->size(), 1);
        net_.forward_pass(hist_);
        bias_.array() = net_.get_activation().col(0).array();
    }

References SSAGES::Grid< T >::begin(), SSAGES::Grid< T >::end(), SSAGES::GridBase< T >::GetDimension(), hgrid_, hist_, net_, and SSAGES::GridBase< T >::size().

Referenced by Build().

Here is the call graph for this function:

Here is the caller graph for this function:

Member Function Documentation

Build()

ANN * SSAGES::ANN::Build ( const Json::Value &  json, const MPI_Comm &  world, const MPI_Comm &  comm, const std::string &  path  )

static

Build a derived method from JSON node.

Parameters

json	JSON Value containing all input information.
world	MPI global communicator.
comm	MPI local communicator.
path	Path for JSON path specification.

Returns

Pointer to the Method built. nullptr if an unknown error occurred.

This function builds a registered method from a JSON node. The difference between this function and "Build" is that this automatically determines the appropriate derived type based on the JSON node information.

Note

Object lifetime is the caller's responsibility.

Definition at line 287 of file ANN.cpp.

    {
        ObjectRequirement validator;
        Value schema;
        CharReaderBuilder rbuilder;
        CharReader* reader = rbuilder.newCharReader();
 
        reader->parse(JsonSchema::ANNMethod.c_str(),
                      JsonSchema::ANNMethod.c_str() + JsonSchema::ANNMethod.size(),
                      &schema, nullptr);
        validator.Parse(schema, path);
 
        // Validate inputs.
        validator.Validate(json, path);
        if(validator.HasErrors())
            throw BuildException(validator.GetErrors());
        
        // Grid. 
        auto* fgrid = Grid<VectorXd>::BuildGrid(json.get("grid", Json::Value()));
        auto* hgrid = Grid<unsigned int>::BuildGrid(json.get("grid", Json::Value()));
        auto* ugrid = Grid<double>::BuildGrid(json.get("grid", Json::Value()));
 
        // Topology. 
        auto nlayers = json["topology"].size() + 2;
        VectorXi topol(nlayers);
        topol[0] = fgrid->GetDimension();
        topol[nlayers-1] = 1;
        for(int i = 0; i < static_cast<int>(json["topology"].size()); ++i)
            topol[i+1] = json["topology"][i].asInt();
        
        auto weight = json.get("weight", 1.).asDouble();
        auto temp = json["temperature"].asDouble();
        auto nsweep = json["nsweep"].asUInt();
 
        // Assume all vectors are the same size. 
        std::vector<double> lowerb, upperb, lowerk, upperk;
        for(int i = 0; i < static_cast<int>(json["lower_bound_restraints"].size()); ++i)
        {
            lowerk.push_back(json["lower_bound_restraints"][i].asDouble());
            upperk.push_back(json["upper_bound_restraints"][i].asDouble());
            lowerb.push_back(json["lower_bounds"][i].asDouble());
            upperb.push_back(json["upper_bounds"][i].asDouble());
        }
 
        auto* m = new ANN(world, comm, topol, fgrid, hgrid, ugrid, lowerb, upperb, lowerk, upperk, temp, weight, nsweep);
 
        // Set optional params.
        m->SetPrevWeight(json.get("prev_weight", 1).asDouble());
        m->SetOutput(json.get("output_file", "ann.out").asString());
        m->SetOutputOverwrite( json.get("overwrite_output", true).asBool());
        m->SetConvergeIters(json.get("converge_iters", 0).asUInt());
        m->SetMaxIters(json.get("max_iters", 1000).asUInt());
        m->SetMinLoss(json.get("min_loss", 0).asDouble());
 
        if(json.isMember("net_state") && json.isMember("load_bias"))
            m->ReadBias(json["net_state"].asString(), json["load_bias"].asString());
 
        return m;
    }

References ANN(), SSAGES::Grid< T >::BuildGrid(), Json::Requirement::GetErrors(), Json::Requirement::HasErrors(), Json::ObjectRequirement::Parse(), and Json::ObjectRequirement::Validate().

Here is the call graph for this function:

PostIntegration()

void SSAGES::ANN::PostIntegration ( Snapshot *  snapshot, const class CVManager &  cvmanager  )

overridevirtual

Method call post integration.

Parameters

snapshot	Pointer to the simulation snapshot.
cvmanager	Collective variable manager.

This function will be called after each integration step.

Implements SSAGES::Method.

Definition at line 113 of file ANN.cpp.

    {
        if(snapshot->GetIteration() && snapshot->GetIteration() % nsweep_ == 0)
        {
            // Switch to full blast.
            if(citers_ && snapshot->GetIteration() > citers_)
                pweight_ = 1.0;
            
            TrainNetwork();
            if(IsMasterRank(world_))
                WriteBias();
        }
 
        // Get CV vals.
        auto cvs = cvmanager.GetCVs(cvmask_);
        auto n = cvs.size();
 
        // Determine if we are in bounds.
        RowVectorXd vec(n);
        std::vector<double> val(n);
        bool inbounds = true;
        for(size_t i = 0; i < n; ++i)
        {
            val[i] = cvs[i]->GetValue();
            vec[i] = cvs[i]->GetValue();
            if(val[i] < hgrid_->GetLower(i) || val[i] > hgrid_->GetUpper(i))
                inbounds = false;
        }
 
        // If in bounds, bias. 
        VectorXd derivatives = VectorXd::Zero(n);
        if(inbounds)
        {
            // Record histogram hit and get gradient. 
            // Only record hits on master processes since we will 
            // reduce later. 
            if(IsMasterRank(comm_))
                hgrid_->at(val) += 1;
            //derivatives = (*fgrid_)[val];
            net_.forward_pass(vec);
            derivatives = net_.get_gradient(0);
        }
        else
        {
            if(IsMasterRank(comm_))
            {
                std::cerr << "ANN (" << snapshot->GetIteration() << "): out of bounds ( ";
                for(auto& v : val)
                    std::cerr << v << " "; 
                std::cerr << ")" << std::endl;
            }
        }
 
        // Restraints.
        for(size_t i = 0; i < n; ++i)
        {
            auto cval = cvs[i]->GetValue();
            if(cval < lowerb_[i])
                derivatives[i] += lowerk_[i]*cvs[i]->GetDifference(lowerb_[i]);
            else if(cval > upperb_[i])
                derivatives[i] += upperk_[i]*cvs[i]->GetDifference(upperb_[i]);
        }
 
        // Apply bias to atoms. 
        auto& forces = snapshot->GetForces(); 
        auto& virial = snapshot->GetVirial();
 
        for(size_t i = 0; i < cvs.size(); ++i)
        {
            auto& grad = cvs[i]->GetGradient();
            auto& boxgrad = cvs[i]->GetBoxGradient();
            
            // Update the forces in snapshot by adding in the force bias from each
            // CV to each atom based on the gradient of the CV.
            for (size_t j = 0; j < forces.size(); ++j)
                forces[j] -= derivatives[i]*grad[j];
            
            virial += derivatives[i]*boxgrad;
        }
    }

References SSAGES::GridBase< T >::at(), citers_, SSAGES::Method::comm_, SSAGES::Method::cvmask_, SSAGES::CVManager::GetCVs(), SSAGES::Snapshot::GetForces(), SSAGES::Snapshot::GetIteration(), SSAGES::GridBase< T >::GetLower(), SSAGES::GridBase< T >::GetUpper(), SSAGES::Snapshot::GetVirial(), hgrid_, SSAGES::EventListener::IsMasterRank(), lowerb_, lowerk_, net_, pweight_, TrainNetwork(), SSAGES::Method::world_, and WriteBias().

Here is the call graph for this function:

PostSimulation()

void SSAGES::ANN::PostSimulation ( Snapshot *  snapshot, const class CVManager &  cvmanager  )

overridevirtual

Method call post simulation.

Parameters

snapshot	Pointer to the simulation snapshot.
cvmanager	Collective variable manager.

This function will be called after the end of the simulation run.

Implements SSAGES::Method.

Definition at line 194 of file ANN.cpp.

    {
    }

PreSimulation()

void SSAGES::ANN::PreSimulation ( Snapshot *  snapshot, const class CVManager &  cvmanager  )

overridevirtual

Method call prior to simulation initiation.

Parameters

snapshot	Pointer to the simulation snapshot.
cvmanager	Collective variable manager.

This function will be called before the simulation is started.

Implements SSAGES::Method.

Definition at line 76 of file ANN.cpp.

    {
        auto ndim = hgrid_->GetDimension();
        kbt_ = snapshot->GetKb()*temp_;
        
        // Zero out forces and histogram. 
        VectorXd vec = VectorXd::Zero(ndim);
        std::fill(hgrid_->begin(), hgrid_->end(), 0);
 
        if(preloaded_)
        {
            net_.forward_pass(hist_);
            bias_.array() = net_.get_activation().col(0).array();
            TrainNetwork();
        }
        else
            std::fill(ugrid_->begin(), ugrid_->end(), 1.0);
 
        // Fill in the reweighting matrix for restraints.
        auto cvs = cvmanager.GetCVs(cvmask_);
        auto ncvs = cvs.size();
 
        for(int i = 0; i < hist_.rows(); ++i)
        {
            auto cval = hist_.row(i);
            for(size_t j = 0; j < ncvs; ++j)
            {
                if(cval[j] < lowerb_[j])
                    rbias_(i,j) = lowerk_[j]*(cval[j] - lowerb_[j])*(cval[j] - lowerb_[j]);
                else if(cval[j] > upperb_[j])
                    rbias_(i,j) = upperk_[j]*(cval[j] - upperb_[j])*(cval[j] - upperb_[j]);
            }
        }
 
        std::fill(fgrid_->begin(), fgrid_->end(), vec);
    }

References SSAGES::Grid< T >::begin(), SSAGES::Method::cvmask_, SSAGES::Grid< T >::end(), fgrid_, SSAGES::CVManager::GetCVs(), SSAGES::GridBase< T >::GetDimension(), SSAGES::Snapshot::GetKb(), hgrid_, hist_, lowerb_, lowerk_, net_, preloaded_, temp_, TrainNetwork(), and ugrid_.

Here is the call graph for this function:

SetConvergeIters()

void SSAGES::ANN::SetConvergeIters ( unsigned int  citers )

inline

Set number of iterations after which we turn on full weight.

Parameters

citers

Number of iterations before full weight

Definition at line 175 of file ANN.h.

        {
            citers_ = citers;
        }

References citers_.

SetMaxIters()

void SSAGES::ANN::SetMaxIters ( unsigned int  iters )

inline

Set maximum number of training iterations per sweep.

Parameters

iters

Maximum iterations per sweep

Definition at line 184 of file ANN.h.

        {
            auto params = net_.get_train_params();
            params.max_iter = iters;
            net_.set_train_params(params);
        }

References net_.

SetMinLoss()

void SSAGES::ANN::SetMinLoss ( double  loss )

inline

Set minimum loss function value (should be zero for production).

Parameters

loss	Minimum loss function value

Definition at line 195 of file ANN.h.

        {
            auto params = net_.get_train_params();
            params.min_loss = loss; 
            net_.set_train_params(params);
        }

References net_.

SetOutput()

void SSAGES::ANN::SetOutput ( const std::string &  outfile )

inline

Set name of output file.

Parameters

outfile

Output file

Definition at line 157 of file ANN.h.

        {
            outfile_ = outfile;
        }

References outfile_.

SetOutputOverwrite()

void SSAGES::ANN::SetOutputOverwrite ( bool  overwrite )

inline

Set overwrite flag on output file.

Parameters

overwrite

Boolean if output file should be overwritten

Definition at line 166 of file ANN.h.

        {
            overwrite_ = overwrite;
        }

References overwrite_.

SetPrevWeight()

void SSAGES::ANN::SetPrevWeight ( double  h )

inline

Set previous history weight.

Parameters

h	History weight

Definition at line 148 of file ANN.h.

        {
            pweight_ = h;
        }

References pweight_.

Member Data Documentation

hist_

Eigen::MatrixXd SSAGES::ANN::hist_

private

Eigen matrices of grids.

Definition at line 73 of file ANN.h.

Referenced by ANN(), PreSimulation(), ReadBias(), TrainNetwork(), and WriteBias().

lowerb_

std::vector<double> SSAGES::ANN::lowerb_

private

Bounds

Definition at line 78 of file ANN.h.

Referenced by PostIntegration(), and PreSimulation().

lowerk_

std::vector<double> SSAGES::ANN::lowerk_

private

Bound restraints.

Definition at line 83 of file ANN.h.

Referenced by PostIntegration(), and PreSimulation().

pweight_

double SSAGES::ANN::pweight_

private

Previous and current histogram weight.

Definition at line 54 of file ANN.h.

Referenced by PostIntegration(), SetPrevWeight(), and TrainNetwork().

sweep_

unsigned int SSAGES::ANN::sweep_

private

Number of iterations per sweep.

Definition at line 43 of file ANN.h.

Referenced by TrainNetwork(), and WriteBias().

temp_

double SSAGES::ANN::temp_

private

System temperature and energy units.

Definition at line 59 of file ANN.h.

Referenced by PreSimulation().

---

The documentation for this class was generated from the following files:

• /home/michael/development/SSAGES/src/Methods/ANN.h
• /home/michael/development/SSAGES/src/Methods/ANN.cpp