InputPreprocessing

The input_preprocessing key of the config.

Defines how input tensors are prepared before being provided to the model. Each training dataset split (train, val, test) can have its own preprocessing configuration, as can input data at inference ("predict") time. All take their base configuration from "default", which can be overridden for each group.

Field	Type	Required	Default	Description
default	DefaultInputPreprocessing	yes		Default input preprocessing configuration
train	TrainInputPreprocessing	no	null
val	ValTestSplitInputPreprocessing	no	null
test	ValTestSplitInputPreprocessing	no	null
predict	PredictInputPreprocessing	no	null

DefaultInputPreprocessing

Controls how windows are processed before being provided to the model.

This class applies default rules to all training-time splits (train/val/test), as well as the predict preprocessor at inference time. Most values can be overriden by the non-default preprocessors.

CONCEPTUAL OVERVIEW:

Raw windows are not passed directly into the model during forward passes. They are first cropped or tiled into "inputs", which may subsequently be transformed by one or more transformation functions, serially. Each "input" is treated as an independent item in a batch to the model.

A config author can define input_size to specify the edge length of each square crop, as well as input_mode, which determines whether the model will see a single random crop or the full grid of cropped inputs. Desired behavior here is variable based on the model, the data split, and whether we are performing fine-tuning or inference.

The size of the windows being provided to preprocessors is determined as follows:

• training time: labeled_data_prep.window_preparer
• inference time: prediction_requests.partitioners

SCENARIOS:

Window size == input_size
═════════════════════════════════════════════════════════

input_mode has no practical effect — one input either way.

window
┌──────────┐
│          │
│  input   │  input_size × input_size
│ (=window)│
│          │
└──────────┘

Window size > input_size
════════════════════════════════════════════════════════════════════

input_mode: all_tiles             input_mode: single_random_crop
─────────────────────             ──────────────────────────────

Tile the full window with         Take ONE random input from
a grid of inputs (complete        the window (used during
spatial coverage).                training for augmentation
                                  via random positioning).

window                            window
┌───────────────────────────┐     ┌───────────────────────────┐
│ ┌──────────┬──────────┐   │     │░░░░░░░░░░░░░░░░░░░░░░░░░░│
│ │          │          │   │     │░░░░░░░░░░░░░░░░░░░░░░░░░░│
│ │ input 1  │ input 2  │   │     │░░░░┌──────────┐░░░░░░░░░░│
│ │          │          │   │     │░░░░│          │░░░░░░░░░░│
│ ├──────────┼──────────┤   │     │░░░░│  random  │░░░░░░░░░░│
│ │          │          │   │     │░░░░│  input   │░░░░░░░░░░│
│ │ input 3  │ input 4  │   │     │░░░░│          │░░░░░░░░░░│
│ │          │          │   │     │░░░░└──────────┘░░░░░░░░░░│
│ └──────────┴──────────┘   │     │░░░░░░░░░░░░░░░░░░░░░░░░░░│
└───────────────────────────┘     └───────────────────────────┘

Each input is input_size          Input is input_size × input_size
× input_size. Edge inputs         at a random position.
may overlap at predict time       ░ = window pixels not seen
(see PredictInputPreprocessing    by the model in this pass.
.overlap_pixels).

Window size < input_size
═════════════════════════════════════════════════════════

INVALID — raises an error at runtime. The window must be at least as large as input_size.

NOTE:

input_mode is inherited by the val and test splits (unless overridden per-split), however:

• train always uses single_random_crop
• predict always tiles

Field	Type	Required	Default	Description
input_size	integer (≥ 1)	yes		Side length of the square input fed to the model.
input_mode	InputMode	yes		How windows are converted to model inputs. all_tiles: tile the full window with a grid of inputs (complete spatial coverage). single_random_crop: take one random input from the window (training augmentation via random positioning). Inherited by val/test splits unless overridden per-split. Train always uses single_random_crop regardless of this setting. Predict always tiles regardless of this setting.
transforms	array of Transform	no	null	The transforms to apply to the input tensors before they are provided to the model.

TrainInputPreprocessing

Training split must always use single random crop.

Field	Type	Required	Default	Description
transforms	array of Transform	no	null	Transforms to apply to inputs. Replaces default transforms; takes precedence over additional_transforms.
additional_transforms	array of Transform	no	null	Additional transforms to apply beyond the default transforms.
sampler	TrainingSampler	no	null	Samples down the data to a smaller set based on the given strategy.
input_mode	"single_random_crop"	no	"single_random_crop"

ValTestSplitInputPreprocessing

Field	Type	Required	Default	Description
transforms	array of Transform	no	null	Transforms to apply to inputs. Replaces default transforms; takes precedence over additional_transforms.
additional_transforms	array of Transform	no	null	Additional transforms to apply beyond the default transforms.
sampler	TrainingSampler	no	null	Samples down the data to a smaller set based on the given strategy.
input_mode	InputMode	no	null	How windows are converted to model inputs. Overrides the default value for this split.

PredictInputPreprocessing

Preprocessing overrides for the predict (inference) split.

overlap_pixels controls how much adjacent inputs overlap when tiling a window. Used to reduce boundary artifacts in dense prediction (i.e. segmentation, per-pixel regression, embeddings)::

overlap_pixels = 0                         overlap_pixels > 0
────────────────────────────────────      ───────────────────────────────────────

window                                    window
┌─────────────────────────────────┐       ┌─────────────────────────────────────┐
│ ┌──────────────┬──────────────┐ │       │ ┌──────────────┬────┬──────────────┐ │
│ │              │              │ │       │ │              │····│              │ │
│ │              │              │ │       │ │              │····│              │ │
│ │   input 1    │   input 2    │ │       │ │   input 1    │····│   input 2    │ │
│ │              │              │ │       │ │              │····│              │ │
│ │              │              │ │       │ │              │····│              │ │
│ ├──────────────┼──────────────┤ │       │ ├──────────────┼────┼──────────────┤ │
│ │              │              │ │       │ │··············│····│··············│ │
│ │              │              │ │       │ ├──────────────┼────┼──────────────┤ │
│ │   input 3    │   input 4    │ │       │ │              │····│              │ │
│ │              │              │ │       │ │              │····│              │ │
│ │              │              │ │       │ │   input 3    │····│   input 4    │ │
│ └──────────────┴──────────────┘ │       │ │              │····│              │ │
└─────────────────────────────────┘       │ │              │····│              │ │
                                          │ └──────────────┴────┴──────────────┘ │
No shared pixels between                  └─────────────────────────────────────┘
adjacent inputs.
                                          ···· = overlap region seen by
                                          adjacent inputs; predictions
                                          merged by the postprocessor.

Field	Type	Required	Default	Description
overlap_pixels	integer (≥ 0)	no	0	Overlap between adjacent inputs when tiling; also used when merging inference outputs. Not relevant for vector-based tasks.
transforms	array of Transform	no	null	Transforms to apply to inputs. Replaces default transforms; takes precedence over additional_transforms.
additional_transforms	array of Transform	no	null	Additional transforms to apply beyond the default transforms.

InputMode

Allowed values: "single_random_crop", "all_tiles"

Transform

One of the following, selected by its name field:

name	Type
manual	EscapeHatchComponent
olmoearth_normalize	OlmoEarthNormalize
random_flip	RandomFlip
sentinel1_to_decibels	Sentinel1ToDecibels

TrainingSampler

One of the following, selected by its name field:

name	Type
random	RandomSampler

OlmoEarthNormalize

Field	Type	Required	Default	Description
name	"olmoearth_normalize"	yes

RandomFlip

Field	Type	Required	Default	Description
name	"random_flip"	yes
x	boolean	no	true	Flip the image horizontally
y	boolean	no	true	Flip the image vertically

Sentinel1ToDecibels

Convert Sentinel-1 RTC data from linear to decibel scale.

Field	Type	Required	Default	Description
name	"sentinel1_to_decibels"	yes

RandomSampler

Field	Type	Required	Default	Description
name	"random"	yes
num_samples	integer (≥ 0)	yes		The number of samples to take
replace	boolean	no	false	Whether sampled items can be sampled again