Galileo Transformers Models

Self-contained HuggingFace model checkpoints for Galileo.

Each checkpoint subfolder ships remote code for model, processor, and custom pipeline loading with trust_remote_code=True. No external galileo package is required at inference time.

Available checkpoints

Usage

Galileo operates on native patch grids (default patch_size: 8 in preprocessor_config.json). Stack shapes are (H, W, T, C); no fixed 224×224 resize is applied.

from transformers import pipeline
import numpy as np

MODEL = "/path/to/GALILEO-transformers/galileo-nano-patch8"

pipe = pipeline(
    task="galileo-feature-extraction",
    model=MODEL,
    trust_remote_code=True,
)

# 10-band Sentinel-2 stack at native spatial size
s2 = np.random.randn(64, 64, 1, 10).astype(np.float32)
features = pipe(s2=s2, pool=True, return_tensors=True)

Sentinel-1 only:

s1 = np.random.randn(64, 64, 1, 2).astype(np.float32)
features = pipe(s1=s1, pool=True, return_tensors=True)

Test CLI

conda activate rsgen
python test_galileo.py
python test_galileo.py --model galileo-tiny-patch8
python test_galileo.py --model galileo-base-patch8 --no-pool

Dependencies

• transformers
• torch
• einops

Per-folder contents

Each checkpoint folder is self-contained:

• config.json — HF config with auto_map and custom_pipelines
• model.safetensors — converted encoder weights
• preprocessor_config.json — processor settings
• modeling_galileo.py — config + encoder + GalileoEncoderModel
• processing_galileo.py — GalileoProcessor
• pipeline_galileo.py — GalileoImageFeatureExtractionPipeline