Skip to content

allenact.embodiedai.aux_losses.losses#

[view_source]

Defining the auxiliary loss for actor critic type models.

Several of the losses defined in this file are modified versions of those found in https://github.com/joel99/habitat-pointnav-aux/blob/master/habitat_baselines/

MultiAuxTaskNegEntropyLoss#

class MultiAuxTaskNegEntropyLoss(AbstractActorCriticLoss)

[view_source]

Used in multiple auxiliary tasks setting. Add a negative entropy loss over all the task weights.

AuxiliaryLoss#

class AuxiliaryLoss(AbstractActorCriticLoss)

[view_source]

Base class of auxiliary loss. Any auxiliary task loss should inherit from it, and implement the get_aux_loss function.

InverseDynamicsLoss#

class InverseDynamicsLoss(AuxiliaryLoss)

[view_source]

Auxiliary task of Inverse Dynamics from Auxiliary Tasks Speed Up Learning PointGoal Navigation (Ye, 2020) https://arxiv.org/abs/2007.04561 originally from Curiosity-driven Exploration by Self-supervised Prediction (Pathak, 2017) https://arxiv.org/abs/1705.05363

InverseDynamicsLoss.__init__#

 | __init__(subsample_rate: float = 0.2, subsample_min_num: int = 10, *args, **kwargs)

[view_source]

Subsample the valid samples by the rate of subsample_rate, if the total num of the valid samples is larger than subsample_min_num.

TemporalDistanceLoss#

class TemporalDistanceLoss(AuxiliaryLoss)

[view_source]

Auxiliary task of Temporal Distance from Auxiliary Tasks Speed Up Learning PointGoal Navigation (Ye, 2020) https://arxiv.org/abs/2007.04561

CPCALoss#

class CPCALoss(AuxiliaryLoss)

[view_source]

Auxiliary task of CPC|A from Auxiliary Tasks Speed Up Learning PointGoal Navigation (Ye, 2020) https://arxiv.org/abs/2007.04561 originally from Neural Predictive Belief Representations (Guo, 2018) https://arxiv.org/abs/1811.06407