drugforge.ml.loss.SmoothL1Loss

class drugforge.ml.loss.SmoothL1Loss(*args: Any, **kwargs: Any)

Bases: SmoothL1Loss

__init__(loss_type=None)

Class for calculating smooth L1 loss, with various options

Parameters:

loss_type (str, optional) –

Which type of loss to use:

• None: vanilla smooth L1

• “step”: step smooth L1 loss

• “uncertainty”: smooth L1 loss with added uncertainty

Methods

__init__([loss_type])	Class for calculating smooth L1 loss, with various options
forward(pred, pose_preds, target, in_range, ...)	Dispatch method for calculating loss.
step_loss(pred, target[, in_range])	Step loss calculation.

forward(pred, pose_preds, target, in_range, uncertainty)

Dispatch method for calculating loss. All arguments should be passed regardless of actual loss function to keep an identical signature for this class. Data is passed to self.loss_function.

step_loss(pred, target, in_range=None)

Step loss calculation. For in_range < 0, loss is returned as 0 if pred < target, otherwise loss is calculated as normal. For in_range > 0, loss is returned as 0 if pred > target, otherwise loss is calculated as normal. For in_range == 0, loss is calculated as normal.

Parameters:

• pred (torch.Tensor) – Model prediction

• target (torch.Tensor) – Prediction target

• in_range (torch.Tensor, optional) – target’s presence in the dynamic range of the assay. Give a value of < 0 for target below lower bound, > 0 for target above upper bound, and 0 or None for inside range

Returns:

Calculated loss

Return type:

torch.Tensor