deepfold.losses.geometry.compute_fape

deepfold.losses.geometry.compute_fape(pred_frames: Rigid, target_frames: Rigid, frames_mask: Tensor, pred_positions: Tensor, target_positions: Tensor, positions_mask: Tensor, length_scale: float, l1_clamp_distance: float | None = None, eps: float = 1e-08) → Tensor[source]

Computes FAPE loss.

Parameters:

  • pred_frames – Rigid object of predicted frames. [*, N_frames]

  • target_frames – Rigid object of ground truth frames. [*, N_frames]

  • frames_mask – Binary mask for the frames. [*, N_frames]

  • pred_positions – Predicted atom positions. [*, N_pts, 3]

  • target_positions – Ground truth positions. [*, N_pts, 3]

  • positions_mask – Positions mask. [*, N_pts]

  • length_scale – Length scale by which the loss is divided.

  • l1_clamp_distance – Cutoff above which distance errors are disregarded.

  • eps – Small value used to regularize denominators.

Returns:

FAPE loss tensor.