Skip to content

allenact.base_abstractions.experiment_config#

[view_source]

Defines the ExperimentConfig abstract class used as the basis of all experiments.

FrozenClassVariables#

class FrozenClassVariables(abc.ABCMeta)

[view_source]

Metaclass for ExperimentConfig.

Ensures ExperimentConfig class-level attributes cannot be modified. ExperimentConfig attributes can still be modified at the object level.

ExperimentConfig#

class ExperimentConfig(, metaclass=FrozenClassVariables)

[view_source]

Abstract class used to define experiments.

Instead of using yaml or text files, experiments in our framework are defined as a class. In particular, to define an experiment one must define a new class inheriting from this class which implements all of the below methods. The below methods will then be called when running the experiment.

ExperimentConfig.tag#

 | @abc.abstractmethod
 | tag() -> str

[view_source]

A string describing the experiment.

ExperimentConfig.training_pipeline#

 | @abc.abstractmethod
 | training_pipeline(**kwargs) -> TrainingPipeline

[view_source]

Creates the training pipeline.

Parameters

  • kwargs : Extra kwargs. Currently unused.

Returns

An instantiate TrainingPipeline object.

ExperimentConfig.machine_params#

 | @abc.abstractmethod
 | machine_params(mode="train", **kwargs) -> Union[MachineParams, Dict[str, Any]]

[view_source]

Parameters used to specify machine information.

Machine information includes at least (1) the number of processes to train with and (2) the gpu devices indices to use.

mode : Whether or not the machine parameters should be those for "train", "valid", or "test". kwargs : Extra kwargs.

Returns

A dictionary of the form{"nprocesses": ..., "gpu_ids": ..., ...}. Here nprocesses must be a non-negative integer, gpu_ids must be a sequence of non-negative integers (if empty, then everything will be run on the cpu).

ExperimentConfig.create_model#

 | @abc.abstractmethod
 | create_model(**kwargs) -> nn.Module

[view_source]

Create the neural model.

ExperimentConfig.make_sampler_fn#

 | @abc.abstractmethod
 | make_sampler_fn(**kwargs) -> TaskSampler

[view_source]

Create the TaskSampler given keyword arguments.

These kwargs will be generated by one of ExperimentConfig.train_task_sampler_args, ExperimentConfig.valid_task_sampler_args, or ExperimentConfig.test_task_sampler_args depending on whether the user has chosen to train, validate, or test.

ExperimentConfig.train_task_sampler_args#

 | train_task_sampler_args(process_ind: int, total_processes: int, devices: Optional[List[int]] = None, seeds: Optional[List[int]] = None, deterministic_cudnn: bool = False) -> Dict[str, Any]

[view_source]

Specifies the training parameters for the process_indth training process.

These parameters are meant be passed as keyword arguments to ExperimentConfig.make_sampler_fn to generate a task sampler.

Parameters

  • process_ind : The unique index of the training process (0 ≤ process_ind < total_processes).
  • total_processes : The total number of training processes.
  • devices : Gpu devices (if any) to use.
  • seeds : The seeds to use, if any.
  • deterministic_cudnn : Whether or not to use deterministic cudnn.

Returns

The parameters for make_sampler_fn

ExperimentConfig.valid_task_sampler_args#

 | valid_task_sampler_args(process_ind: int, total_processes: int, devices: Optional[List[int]] = None, seeds: Optional[List[int]] = None, deterministic_cudnn: bool = False) -> Dict[str, Any]

[view_source]

Specifies the validation parameters for the process_indth validation process.

See ExperimentConfig.train_task_sampler_args for parameter definitions.

ExperimentConfig.test_task_sampler_args#

 | test_task_sampler_args(process_ind: int, total_processes: int, devices: Optional[List[int]] = None, seeds: Optional[List[int]] = None, deterministic_cudnn: bool = False) -> Dict[str, Any]

[view_source]

Specifies the test parameters for the process_indth test process.

See ExperimentConfig.train_task_sampler_args for parameter definitions.