torchgeo.trainers#

TorchGeo provides LightningModules for a number of common tasks in geospatial and geotemporal deep learning.

Supervised Learning#

Supervised learning tasks have both inputs and labeled outputs.

Supervised Learning Tasks#
Input	Output	Task
\(\mathbb{R}^{T \times C}\)	\(\mathbb{R}^{T \times C}\)	TemporalRegressionTask
\(\mathbb{R}^{C \times H \times W}\)	\(\mathbb{N}\) or \(\mathbb{N}^C\)	ClassificationTask
\(\mathbb{R}^{C \times H \times W}\)	\(\mathbb{R}\) or \(\mathbb{R}^C\)	RegressionTask
\(\mathbb{R}^{C \times H \times W}\)	\(\mathbb{N}^{H \times W}\)	SemanticSegmentationTask
\(\mathbb{R}^{C \times H \times W}\)	\(\mathbb{R}^{H \times W}\)	PixelwiseRegressionTask
\(\mathbb{R}^{C \times H \times W}\)	\(\mathbb{R}^{O \times 4}\)	ObjectDetectionTask
\(\mathbb{R}^{C \times H \times W}\)	\(\mathbb{N}^{O \times H \times W}\)	InstanceSegmentationTask
\(\mathbb{R}^{2 \times C \times H \times W}\)	\(\mathbb{N}^{H \times W}\)	ChangeDetectionTask
\(\mathbb{R}^{T \times C \times H \times W}\)	\(\mathbb{N}^{H \times W}\)	SpatioTemporalSegmentationTask

Self-supervised learning (SSL) tasks have inputs and create their own labeled outputs.

Self-Supervised Learning Tasks#
Input	Task
\(\mathbb{R}^{C \times H \times W}\) or \(\mathbb{R}^{T \times C \times H \times W}\)	BYOLTask
\(\mathbb{R}^{C \times H \times W}\)	MAETask
\(\mathbb{R}^{C \times H \times W}\) or \(\mathbb{R}^{T \times C \times H \times W}\)	MoCoTask
\(\mathbb{R}^{C \times H \times W}\) or \(\mathbb{R}^{T \times C \times H \times W}\)	SimCLRTask

Tasks that do not relate to learning.

Non-Learning Tasks#
Input	Task
\(\mathbb{R}^{C \times H \times W}\)	IOBenchTask

Abstract base classes that all other tasks inherit from.

Base Classes#
Task
BaseTask

Mixins that support code reuse across multiple tasks.

Mixins#
Output	Mixin
\(\mathbb{N}\) or \(\mathbb{N}^C\) or \(\mathbb{N}^{H \times W}\)	ClassificationMixin
\(\mathbb{R}\) or \(\mathbb{R}^C\) or \(\mathbb{R}^{T \times C}\) or \(\mathbb{R}^{H \times W}\)	RegressionMixin