Skip to content

Frequently Asked Questions (FAQ)

A: First, ensure your CUDA version is compatible with your PyTorch version. You can check the compatibility table on the PyTorch website. If issues persist, try installing the corresponding PyTorch version first, then install torchhydro.

Q: Is GPU required?

A: No, torchhydro can run on CPU environments, but using a GPU will significantly speed up training.

Q: What data formats are supported?

A: Currently, we support CAMELS dataset format, NetCDF format, and common CSV formats. For other formats, you can refer to the documentation to create custom data loaders.

Q: How to handle missing values?

A: torchhydro provides various missing value handling strategies, including mean filling, forward filling, and interpolation. You can use the MissingValueFiller transformer to handle missing values.

Q: What should I do if the model loss doesn't converge?

A: Try the following approaches: - Adjust the learning rate - Check if the data preprocessing is appropriate - Use gradient clipping - Try different model architectures

Q: How to choose the appropriate model architecture?

A: It depends on your specific task. LSTM is suitable for time series data, CNN for spatial data, and Transformer excels in long sequence tasks. We recommend starting with simpler models.

Q: How to improve training speed?

A: Consider: - Using GPU training - Increasing batch size - Using data preloading - Optimizing data preprocessing pipeline

Q: How to handle out-of-memory issues?

A: Try: - Reducing batch size - Using data generators - Reducing model complexity - Using gradient accumulation

5. Other Questions

Q: How can I contribute to the project?

A: Please refer to our contribution guide. We welcome all forms of contributions, including code, documentation, and issue reporting.

Q: Where can I get help?

A: You can: - Check the detailed documentation - Submit an Issue on GitHub - Join our community discussions