Troubleshooting

Common issues and solutions.

CUDA OOM (Out of Memory)

Symptom: RuntimeError: CUDA out of memory

Fixes: 1. Reduce batch_size. 2. Enable gradient_checkpointing for the model. 3. Use mixed precision (fp16). 4. Clear cache: torch.cuda.empty_cache() (use sparingly).

GDAL/Rasterio Installation

Symptom: ImportError: /usr/lib/libgdal.so...

Fixes: - Use conda for easiest geospatial dependency management. - Or use pip install ununennium[geo] which attempts to pull binary wheels.

NaN Loss

Symptom: Loss becomes nan.

Fixes: - Check learning rate (too high?). - Check for division by zero in custom losses/metrics. - Enable detect_anomaly=True in PyTorch for traceback.