Inheritance diagram for ScratchPad:

Public Member Functions
void	setCheckPoint (SubSystem checkPointState)

std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor >	Loss (torch::Tensor t_seq, std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor > icfs, torch::Tensor H0, torch::Tensor kineticEnergy, torch::Tensor potentialEnergy, int n, int Np, int d)

std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor >	UpdateParamsNADAM (torch::Tensor t_seq, std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor > icfs, torch::Tensor velocities, torch::Tensor S, torch::Tensor totalEnergy, torch::Tensor kineticEnergy, torch::Tensor potentialEnergy, int epoch, int n, int Np, int d, double alpha, double epsilon, torch::Tensor beta)

std::pair< torch::Tensor, torch::Tensor >	mainTrain (torch::Tensor params, torch::Tensor t_seq, std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor > icfs, int num_epochs, torch::Tensor totalEnergy, torch::Tensor kineticEnergy, torch::Tensor potentialEnergy, int n, int Np, int d, double learn_rate, double momentum)

Public Member Functions inherited from PND
	PND ()

void	defineParams (int numberOfAtoms)

std::pair< torch::Tensor, torch::Tensor >	LossPreTrain (torch::Tensor t_seq, std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor > icfs, int n, int Np, int d)

std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor >	UpdatePreParamsNADAM (torch::Tensor t_seq, std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor > icfs, torch::Tensor velocities, torch::Tensor S, int epoch, int n, int Np, int d, double alpha=0.001, double epsilon=pow(10, -7), torch::Tensor beta=torch::tensor({0.999, 0.999}))

std::pair< torch::Tensor, torch::Tensor >	PreTrain (torch::Tensor t_seq, std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor > icfs, int num_epochs, int n, int Np, int d, double learn_rate=0.001, double momentum=0.99)

Public Attributes
SubSystem	checkPointState

Public Attributes inherited from PND
int	PreTrainEpochs

int	MainTrainEpochs

int	StepTrain

int	nodes

torch::Tensor	params

Detailed Description

scratch_pad.cpp is an examlpe to demonstrates how users implement loss functions, specify intial and boundary conditions, and conservation laws for their systems

Member Function Documentation

◆ Loss()

std::tuple<torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor> ScratchPad::Loss	(	torch::Tensor	t_seq,
		std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor >	icfs,
		torch::Tensor	totalEnergy,
		torch::Tensor	kineticEnergy,
		torch::Tensor	potentialEnergy,
		int	n,
		int	Np,
		int	d
	)

inlinevirtual

Define a Loss function for the network which is a set of DEs governing the evolution of the system. Only method declaration provided, method must be defined when user's working class inherits from PND

Parameters

t_seq	time sequences for which the pre-training is carried out
icfs	initial and final positions and velocites that are to be used for the pre-training task
totalEnergy	system total energy across time steps
kineticEnergy	system kinetic energy across time steps
potentialEnergy	system potential energy across time steps
n	Number of neurons in the NN
Np	Number of atoms in the MD system
d	dimensionality of the space (d = 3 when data in x,y,z co-ordinates is available)

Returns: returns a tuple of sum of gradients, mean squared error(MSE) loss, boundary loss, Least Action Loss, Hamiltonian Loss, momentum Loss

Reimplemented from PND.

◆ mainTrain()

std::pair<torch::Tensor, torch::Tensor> ScratchPad::mainTrain	(	torch::Tensor	params,
		torch::Tensor	t_seq,
		std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor >	icfs,
		int	num_epochs,
		torch::Tensor	totalEnergy,
		torch::Tensor	kineticEnergy,
		torch::Tensor	potentialEnergy,
		int	n,
		int	Np,
		int	d,
		double	learn_rate,
		double	momentum
	)

inlinevirtual

Routine to follow for main-training the NN, i.e., training for the target data. Users may decide to call their parameter updating algorithm implementation in each training iteration.

Parameters

params	inital parameters of the NN which will be trained
t_seq	time sequences for which the pre-training is carried out
icfs	initial and final positions and velocites that are to be used for the pre-training task
num_epochs	number of epochs for which training is to be carried out
totalEnergy	system total energy across time steps
kineticEnergy	system kinetic energy across time steps
potentialEnergy	system potential energy across time steps
n	number of neurons in the NN
Np	number of atoms in the MD system
d	number of atoms in the MD system
learn_rate	learning rate for optimizer
momentum	parameter to pass to the optimizer that uses momentum

Returns: the tuple of trained weights and mean loss value at the end of training

Reimplemented from PND.

◆ UpdateParamsNADAM()

std::tuple<torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor> ScratchPad::UpdateParamsNADAM	(	torch::Tensor	t_seq,
		std::tuple< torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor >	icfs,
		torch::Tensor	velocities,
		torch::Tensor	S,
		torch::Tensor	totalEnergy,
		torch::Tensor	kineticEnergy,
		torch::Tensor	potentialEnergy,
		int	epoch,
		int	n,
		int	Np,
		int	d,
		double	alpha,
		double	epsilon,
		torch::Tensor	beta
	)

inlinevirtual

Defines an optimizer for adjusting the parameters in each main training iteration. We provide the declaration and method must be defined when user's working class inherits from PND

Parameters

t_seq	time sequences for which the pre-training is carried out
icfs	initial and final positions and velocites that are to be used for the pre-training task
velocities
S
totalEnergy	system total energy across time steps
kineticEnergy	system kinetic energy across time steps
potentialEnergy	system potential energy across time steps
epoch	defines the number of steps
n	number of neurons in the NN
Np	number of atoms in the MD system
d	dimensionality of the space (d = 3 when data in x,y,z co-ordinates is available)
alpha	parameters for optimizers
epsilon	constant for stability
beta	parameters for optimizers

Returns: the tuple of trained weights and mean loss value at the end of training

Reimplemented from PND.

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

Example/scratch_pad.cpp

Public Member Functions

Public Attributes

Detailed Description

Member Function Documentation

◆ Loss()

◆ mainTrain()

◆ UpdateParamsNADAM()