Skip to content

Colab and Kaggle

The quick-start notebook runs unchanged on both platforms:

Open in Colab Open in Kaggle

It downloads a real Sentinel-2 scene, super-resolves it, and walks through crop monitoring, urban analysis and flood / landslide change detection.

Google Colab

  1. Open the notebook with the badge above.
  2. Runtime → Change runtime type → T4 GPU (optional; the CPU runtime works too, more slowly).

  3. Run the first cell:

    %pip install -q "synapse-sr[stac,xarray]" matplotlib

    No restart is needed. Colab already ships PyTorch and Triton. 4. Check the environment:

    import sys
!{sys.executable} -m synapse_sr --env        # works in any Jupyter; the shorter `!synapse-sr --env` on Colab and Kaggle

    On a GPU runtime pro_scan_backend reads triton. Triton compiles its kernel once per session, which takes a few seconds on the first scene.

Kaggle

  1. Open the notebook with the badge above, or Create → New notebook → File → Import notebook and paste the GitHub URL of notebooks/quickstart.ipynb.
  2. In the right-hand panel: Session options → Accelerator → GPU T4 x2 or P100, and Internet → On. Internet access is needed for the one-time weight download and for fetch_sentinel2.
  3. Run the cells in order.

Without internet on Kaggle

Add the weights as a Kaggle dataset: download synapse-pro-v2.safetensors from Hugging Face and upload it. Then load it with Pro.from_pretrained(weights="/kaggle/input/<dataset>/synapse-pro-v2.safetensors"). No network access is needed after that.

Using your own files

from google.colab import files
up = files.upload()                                  # pick your Sentinel-2 GeoTIFF
r = model.super_resolve(next(iter(up)))
r.save("result_2m.tif"); files.download("result_2m.tif")

Or mount Google Drive: from google.colab import drive; drive.mount("/content/drive").

Add the GeoTIFF as a dataset (Add data → Upload), then:

r = model.super_resolve("/kaggle/input/<dataset>/scene.tif")
r.save("/kaggle/working/scene_2m.tif")              # appears under Output

Speed and memory

Runtime Model Notes
Colab / Kaggle GPU Pro Triton scan kernel; the default of 8 tiles per batch fits a T4 (16 GB)
Colab / Kaggle CPU Flash (default) works well: on Kaggle's free 2-core CPU runtime the whole quick-start notebook (install, three scene downloads, three super-resolutions, all plots) finishes in about 3 minutes

If the GPU runs out of memory, lower batch (tiles per forward pass) or tile:

from synapse_sr import Pro

model = Pro.from_pretrained()
r = model.super_resolve("scene.tif", batch=2)          # 2 tiles per forward pass