PromethEO

From pretrained EO representations to simple, reusable building blocks.

Essence

Two things only:

1. Datasets produce Predictors (a typed container of modalities: s1, s2, meteo, dem, latlon, timestamps, optional label).
2. Models consume Predictors and return embeddings (per time, pooled, or token sequence) and optional task head outputs.

Everything else (training loop, collation, pooling) is lightweight glue so you can swap either side.

What It Supports

• Pixel time series (H=W=1)
• Multi‑temporal image patches / stacks (H,W > 1)
• Any subset of modalities (missing ones are None)

Included Examples

• Dataset: WorldCerealDataset / WorldCerealLabelledDataset – illustrative only. Downstream projects should implement their own dataset class that returns a Predictors instance per sample.
• Model: Presto – pixel‑centric pretrained encoder (works for patches too). Additional models just need a thin wrapper.

Minimal Flow

from torch.utils.data import DataLoader
from prometheo.datasets import WorldCerealLabelledDataset
from prometheo.models import Presto
from prometheo import finetune
from prometheo.finetune import Hyperparams
from torch.nn import BCEWithLogitsLoss

train_ds = WorldCerealLabelledDataset(df, task_type="binary", num_outputs=1)
train_dl = DataLoader(train_ds, batch_size=8, shuffle=True)
val_dl   = DataLoader(train_ds, batch_size=8)

model = Presto(num_outputs=1)
finetune.run_finetuning(
    model,
    train_dl,
    val_dl,
    experiment_name="demo",
    output_dir="./outputs",
    loss_fn=BCEWithLogitsLoss(),
    hyperparams=Hyperparams(max_epochs=3, batch_size=8)
)

Inference

from prometheo.datasets import WorldCerealDataset
from prometheo.models import Presto
from prometheo.models.presto import load_presto_weights

sample = WorldCerealDataset(df)[0]
model  = load_presto_weights(Presto(), "weights.pt")
emb    = model(sample)  # (B, H, W, T|1, D)

Add Your Own Dataset

Return a Predictors object in __getitem__ – nothing else required.

Add Your Own Model

Wrap it so forward(predictors, eval_pooling=...) -> tensor or (tokens, mask, meta).

Install

pip install .

Dev mode:

pip install -e .[dev]

License

See LICENSE.

Acknowledgement

Built around an initial WorldCereal (crop mapping) use case; intentionally task‑agnostic.