PyTorch

PyTorch#

Integral Driver: PyTorch#

Pytorch-based integral implementations.

Drivers#

Driver: PyTorch#

Collection of PyTorch-based integral drivers.

class dxtb._src.integral.driver.pytorch.driver.BaseIntDriverPytorch(*args, **kwargs)[source]#

Bases: PytorchImplementation, IntDriver

PyTorch-based integral driver.

Note

Currently, only the overlap integral is implemented.

setup(positions, **kwargs)[source]#

Run the libcint-specific driver setup.

Return type:

None

Parameters:

positions (Tensor) – Cartesian coordinates of all atoms (shape: (..., nat, 3)).

Parameters:
abstract setup_eval_funcs()[source]#

Specification of the overlap (gradient) evaluation functions (eval_ovlp and eval_ovlp_grad).

Return type:

None

eval_ovlp: OverlapFunction#

Function for overlap calculation.

eval_ovlp_grad: OverlapFunction#

Function for overlap gradient calculation.

family: Literal['PyTorch']#

Label for integral implementation family.

ihelp: IndexHelper#

Helper class for indexing.

numbers: Tensor#

Atomic numbers of the atoms in the system.

par: ParamModule#

Representation of parametrization of xTB method.

class dxtb._src.integral.driver.pytorch.driver.IntDriverPytorch(*args, **kwargs)[source]#

Bases: BaseIntDriverPytorch

PyTorch-based integral driver.

The overlap evaluation function implements a custom backward function containing the analytical overlap derivative.

Note

Currently, only the overlap integral is implemented.

setup_eval_funcs()[source]#

Specification of the overlap (gradient) evaluation functions (eval_ovlp and eval_ovlp_grad).

Return type:

None

eval_ovlp: OverlapFunction#

Function for overlap calculation.

eval_ovlp_grad: OverlapFunction#

Function for overlap gradient calculation.

family: Literal['PyTorch']#

Label for integral implementation family.

ihelp: IndexHelper#

Helper class for indexing.

numbers: Tensor#

Atomic numbers of the atoms in the system.

par: ParamModule#

Representation of parametrization of xTB method.

class dxtb._src.integral.driver.pytorch.driver.IntDriverPytorchLegacy(*args, **kwargs)[source]#

Bases: BaseIntDriverPytorch

PyTorch-based integral driver with old loop-based version of the full matrix build. The newer version partially vectorizes over the centers of the orbitals (unique pair algorithm).

Note

Currently, only the overlap integral is implemented.

setup_eval_funcs()[source]#

Specification of the overlap (gradient) evaluation functions (eval_ovlp and eval_ovlp_grad).

Return type:

None

family: Literal['PyTorch']#

Label for integral implementation family.

ihelp: IndexHelper#

Helper class for indexing.

numbers: Tensor#

Atomic numbers of the atoms in the system.

par: ParamModule#

Representation of parametrization of xTB method.

class dxtb._src.integral.driver.pytorch.driver.IntDriverPytorchNoAnalytical(*args, **kwargs)[source]#

Bases: BaseIntDriverPytorch

PyTorch-based integral driver without analytical derivatives.

Note

Currently, only the overlap integral is implemented.

setup_eval_funcs()[source]#

Specification of the overlap (gradient) evaluation functions (eval_ovlp and eval_ovlp_grad).

Return type:

None

family: Literal['PyTorch']#

Label for integral implementation family.

ihelp: IndexHelper#

Helper class for indexing.

numbers: Tensor#

Atomic numbers of the atoms in the system.

par: ParamModule#

Representation of parametrization of xTB method.

Implementations#

Implementation: Overlap#

PyTorch-based overlap implementations.

class dxtb._src.integral.driver.pytorch.overlap.OverlapPytorch(device=None, dtype=None, uplo='l', cutoff=50.0)[source]#

Bases: OverlapIntegral, IntegralPytorch

Overlap integral from atomic orbitals.

Use the build() method to calculate the overlap integral. The returned matrix uses a custom autograd function to calculate the backward pass with the analytical gradient. For the full gradient, i.e., a matrix of shape (..., norb, norb, 3), the get_gradient() method should be used.

Parameters:
build(driver)[source]#

Overlap calculation of unique shells pairs, using the McMurchie-Davidson algorithm.

Return type:

Tensor

Parameters:

driver (BaseIntDriverPytorch) – The integral driver for the calculation.

Returns:

Overlap integral matrix of shape (..., norb, norb).

Return type:

Tensor

Parameters:

driver (BaseIntDriverPytorch)

get_gradient(driver)[source]#

Overlap gradient calculation of unique shells pairs, using the McMurchie-Davidson algorithm.

Return type:

Tensor

Parameters:

driver (BaseIntDriverPytorch) – Integral driver for the calculation.

Returns:

Overlap gradient of shape (..., norb, norb, 3).

Return type:

Tensor

Parameters:

driver (BaseIntDriverPytorch)

cutoff: Tensor | float | int | None#

Real-space cutoff for integral calculation in Bohr. Defaults to constants.defaults.INTCUTOFF.

family: Literal['PyTorch']#

Label for integral implementation family.

uplo: Literal['n', 'u', 'l']#

Whether the matrix of unique shell pairs should be create as a triangular matrix (l: lower, u: upper) or full matrix (n). Defaults to l (lower triangular matrix).

Contents