Skip to content

Installation

Requirements
Python 3.8 or newer
PyTorch 1.13 or newer (CPU or CUDA build)
Other numpy, scipy, safetensors, rasterio, affine, rich, installed automatically

Install

pip install synapse-sr

pip install git+https://github.com/SharadhNaidu/synapse-sr.git

git clone https://github.com/SharadhNaidu/synapse-sr.git
cd synapse-sr
pip install -e ".[all,test]"

Install the PyTorch build that matches your hardware first if you need a specific CUDA version. See pytorch.org.

Optional extras

Extra Installs Enables
synapse-sr[stac] pystac-client fetch_sentinel2 for downloading a scene
synapse-sr[xarray] xarray Result.to_xarray() and xarray inputs
synapse-sr[all] both of the above
synapse-sr[cuda] mamba-ssm the fused CUDA selective-scan kernel

Platforms

Platform Pro scan backend Recommended model
Google Colab, Kaggle (GPU) triton, compiled on first use; nothing extra to install Pro
Linux + NVIDIA GPU triton; fused if mamba-ssm is installed Pro
Windows + NVIDIA GPU pytorch (exact, slower) Pro or Flash
CPU only, laptops, integrated graphics pytorch Flash
macOS (Intel or Apple silicon), ARM Linux pytorch Flash

All backends give the same numbers to about 1e-5 relative error; only the speed differs. See Choosing a model and a device.

Optional: the fused mamba-ssm kernel

mamba-ssm compiles CUDA code and must match your PyTorch and CUDA versions:

pip install mamba-ssm --no-build-isolation

If it is installed but was built for a different PyTorch, synapse-sr warns and uses Triton or the PyTorch path instead of failing. SYNAPSE_SR_DISABLE_FUSED=1 and SYNAPSE_SR_DISABLE_TRITON=1 switch those backends off.

Check the installation

synapse-sr --env

This prints the synapse-sr, Python, PyTorch, numpy, rasterio and GDAL versions, the CPU thread count, whether CUDA, bfloat16 and Apple MPS are available, and which scan backend Pro will use (--json for a machine-readable version). synapse-sr --models lists the registered checkpoints and whether they are published.