d3rlpy.augmentation.pipeline.DrQPipeline

class d3rlpy.augmentation.pipeline.DrQPipeline(augmentations=None, n_mean=1)

Data-reguralized Q augmentation pipeline.

References

• Kostrikov et al., Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels.

Parameters:

• augmentations (list(d3rlpy.augmentation.base.Augmentation or str)) – list of augmentations or augmentation types.
• n_mean (int) – the number of computations to average

augmentations

list of augmentations.

Type:list(d3rlpy.augmentation.base.Augmentation)

n_mean

the number of computations to average

Type:int

Methods

append(augmentation)

Append augmentation to pipeline.

Parameters:augmentation (d3rlpy.augmentation.base.Augmentation) – augmentation.

get_augmentation_params()

Returns augmentation parameters.

Parameters:deep (bool) – flag to deeply copy objects. Returns:list of augmentation parameters. Return type:list(dict)

get_augmentation_types()

Returns augmentation types.

Returns:list of augmentation types. Return type:list(str)

get_params(deep=False)

Returns pipeline parameters.

Returns:piple parameters. Return type:dict

process(func, inputs, targets)

Runs a given function while augmenting inputs.

Parameters:

• func (callable) – function to compute.
• inputs (dict) – inputs to the func.
• target (list(str)) – list of argument names to augment.

Returns:

the computation result.

Return type:

torch.Tensor

transform(x)

Returns observation processed by all augmentations.

Parameters:x (torch.Tensor) – observation tensor. Returns:processed observation tensor. Return type:torch.Tensor