flash.core¶

flash.core.adapter ¶

`Adapter`	The `Adapter` is a lightweight interface that can be used to encapsulate the logic from a particular provider within a `Task`.
`AdapterTask`	The `AdapterTask` is a `Task` which wraps an `Adapter` and forwards all of the hooks.

`ClassesOutput`	A `Output` which applies an argmax to the model outputs (either logits or probabilities) and converts to a list.
`ClassificationOutput`	A base class for classification outputs.
`ClassificationTask`
`FiftyOneLabelsOutput`	A `Output` which converts the model outputs to FiftyOne classification format.
`LabelsOutput`	A `Output` which converts the model outputs (either logits or probabilities) to the label of the argmax classification.
`LogitsOutput`	A `Output` which simply converts the model outputs (assumed to be logits) to a list.
`PredsClassificationOutput`	A `ClassificationOutput` which gets the `PREDS` from the sample.
`ProbabilitiesOutput`	A `Output` which applies a softmax to the model outputs (assumed to be logits) and converts to a list.

`FlashBaseFinetuning`	FlashBaseFinetuning can be used to create a custom Flash Finetuning Callback.
`FineTuningHooks`	Hooks to be used in Task and FlashBaseTuning.
`Freeze`
`FreezeUnfreeze`
`NoFreeze`
`UnfreezeMilestones`

Visualizes predictions from a model with a FiftyOne Output in the FiftyOne App.

type _sphinx_paramlinks_flash.core.integrations.icevision.transforms.IceVisionTransformAdapter.transform: List[Callable]

`BenchmarkConvergenceCI`	Specialized callback only used during testing Keeps track metrics during training.
`CheckDependenciesMeta`
`ModuleWrapperBase`	The `ModuleWrapperBase` is a base for classes which wrap a `LightningModule` or an instance of `ModuleWrapperBase`.
`DatasetProcessor`	The `DatasetProcessor` mixin provides hooks for classes which need custom logic for producing the data loaders for each running stage given the corresponding dataset.
`Task`	A general Task.

`FlashRegistry`	This class is used to register function or `functools.partial` class to a registry.
`ExternalRegistry`	The `ExternalRegistry` is a `FlashRegistry` that can point to an external provider via a getter function.
`ConcatRegistry`	The `ConcatRegistry` can be used to concatenate multiple registries of different types together.

`LARS`	Extends SGD in PyTorch with LARS scaling.
`LAMB`	Extends ADAM in pytorch to incorporate LAMB algorithm from the paper: Large batch optimization for deep learning: Training BERT in 76 minutes.
`LinearWarmupCosineAnnealingLR`	Sets the learning rate of each parameter group to follow a linear warmup schedule between warmup_start_lr and base_lr followed by a cosine annealing schedule between base_lr and eta_min.

Modified version of pytorch_lightning.utilities.argparse.from_argparse_args() which populates valid_kwargs from pytorch_lightning.Trainer.