waxmorph.torch.train.TrainResult

Contents

waxmorph.torch.train.TrainResult#

class waxmorph.torch.train.TrainResult(model, log)[source]#

Bases: object

Best post-update model and its training log.

The supplied optimizer state is restored alongside model. losses_total, losses_shape, and losses_l2 are length-n_epochs post-update histories; shape loss sums supervised frames and L2 is the unweighted learned-displacement regularizer. best_traj_pos and best_traj_pol have shape [t_rollout + 1, N, 3]; best_traj_c has shape [t_rollout + 1, N, C]. best_epoch is zero-based; other metadata keys are best_loss, sorted target_frames, and one config_<field> per TrainConfig field. JAX adds config_max_edges_factor.

Parameters:
Contents