Troubleshooting#

`Dataset directory not found` / files missing#

Datasets must live under ./data/ from the directory where you run torchgeo-bench. The runner does not honour GEOBENCH_ROOT or GEOBENCH_V2_ROOT environment variables — paths are fixed:

V1: data/classification_v1.0/<name>/
V2: data/geobenchv2/<name>/
EuroSAT: data/eurosat/

Re-run torchgeo-bench download … to fetch missing data. If your data lives elsewhere, symlink data/ to the real location.

`ModuleNotFoundError: geobench`#

The legacy geobench package is no longer a dependency. V1 datasets are read directly from HDF5 (the internal GeoBenchv1 loader in src/torchgeo_bench/datasets/geobench_v1.py); V2 dispatches to upstream geobench_v2.datasets.GeoBench<X>. Make sure your environment matches the pinned geobenchv2 version in pyproject.toml.

CUDA out of memory#

$ torchgeo-bench run dataset.batch_size=32
$ # or run on CPU
$ torchgeo-bench run device=cpu

For segmentation, also try

$ torchgeo-bench run \
    eval.segmentation.cache_dtype=float32 \
    eval.segmentation.cache_features=false

if RAM (rather than GPU memory) is the bottleneck.

GPU run crashes immediately#

The default config is device: cuda:0, so the first documented run uses the GPU. uv sync installs the latest torch, whose bundled CUDA and kernel architectures may not match your GPU or driver. Two distinct failures:

RuntimeError: The NVIDIA driver on your system is too old — the installed torch was built against a newer CUDA than your driver supports.
CUDA error: no kernel image is available for execution on the device (cudaErrorNoKernelImageForDevice) — torch’s CUDA 12.8 wheels dropped Volta (sm_70) kernels, so they fail on a V100 even though CUDA initialises. Older GPUs need a cu126 (or earlier) build, e.g. torch==2.7.1+cu126 + torchvision==0.22.1+cu126 from https://download.pytorch.org.

Either way you can fall back to CPU (slower, but always works):

$ torchgeo-bench run dataset.names=[m-eurosat] device=cpu

CPU is fine for the small V1 splits, but large V2 datasets (e.g. benv2 / BigEarthNet) can take far longer — prefer a working GPU for those.

`KeyError: 's2'` on a V2 dataset#

A known V2 issue: geobench_v2.rearrange_bands expects modality keys ('s2', 's1', …) that aren’t present when a flat band list is requested. Workaround: use dataset.bands=all for affected V2 datasets.

`eurosat-spatial` reports `Dataset not found`#

torchgeo-bench download eurosat fetches EuroSAT plus the standard eurosat-{train,val,test}.txt splits, but the eurosat-spatial dataset uses torchgeo.datasets.EuroSATSpatial, which needs its own spatial split files. Those download automatically on the first CLI run that uses eurosat-spatial; the plain download eurosat command does not provision them, so its slow test skips until that first run.

V1 slow tests skip after the auto-download#

The per-dataset auto-download (triggered by running a V1 dataset such as dataset.names=[m-eurosat]) writes the webdataset layout under data/classification_v1.0_wds/. The V1 slow integration tests instead read the legacy HDF5 layout under data/classification_v1.0/ and skip if only the _wds data is present. Fetch the legacy bundle with torchgeo-bench download geobench_v1 to run them.

Build / docs warnings#

If you build the docs locally without internet access, expect ~9 WARNING: failed to reach any of the inventories messages from sphinx.ext.intersphinx. These are network reachability errors, not real issues — the GitHub Pages build runner has network access and resolves these inventories cleanly.

Troubleshooting#

Dataset directory not found / files missing#

ModuleNotFoundError: geobench#