mlcolvar.core.loss.GeneratorLoss

class mlcolvar.core.loss.GeneratorLoss(r: int, eta: float, friction: Tensor, alpha: float, descriptors_derivatives: SmartDerivatives | Tensor = None, n_dim: int = 3, u_stat: bool = True)

Bases: Module

Computes the loss function to learn a representation for the resolvent of the infinitesimal generator

__init__(r: int, eta: float, friction: Tensor, alpha: float, descriptors_derivatives: SmartDerivatives | Tensor = None, n_dim: int = 3, u_stat: bool = True)

Computes the loss to learn a representation on which the resolvent of the infinitesimal generator can be learned

Parameters:

• r (int) – Number of eigenfunctions wanted, i.e., number of outputs of model.

• eta (float) – Hyperparameter for the shift to define the resolvent, i.e., $(eta I-_mathcal{L})^{-1}$

• friction (torch.Tensor) – Langevin friction, i.e., $sqrt{k_B*T/(gamma*m_i)}$

• alpha (float) – Hyperparamer that scales the contribution of orthonormality loss to the total loss, i.e., L = L_ef + alpha*L_ortho

• descriptors_derivatives (Union[SmartDerivatives, torch.Tensor], optional) – Derivatives of descriptors wrt atomic positions (if used) to speed up calculation of gradients, by default None. Can be either:

  • A SmartDerivatives object to save both memory and time, see also mlcolvar.core.loss.committor_loss.SmartDerivatives

  • A torch.Tensor with the derivatives to save time, memory-wise could be less efficient

• ref_idx (torch.Tensor, optional) – Reference indeces for the unshuffled dataset for properly handling batching/splitting/shuffling when descriptors derivatives are provided, by default None. Ref_idx can be generated automatically using SmartDerivatives or by setting create_ref_idx=True when initializing a DictDataset. See also mlcolvar.core.loss.utils.smart_derivatives.SmartDerivatives

• n_dim (int) – Number of dimensions, by default 3.

• u_stat (bool) – Do we use U-statistics to compute the loss

Methods

__init__(r, eta, friction, alpha[, ...])	Computes the loss to learn a representation on which the resolvent of the infinitesimal generator can be learned
forward(input, output, weights[, ref_idx])	Define the computation performed at every call.

forward(input: Tensor, output: Tensor, weights: Tensor, ref_idx: Tensor = None) → Tuple[Tensor, Tensor, Tensor]

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Attributes

T_destination
call_super_init
dump_patches
training