Captum for PyTorch Model Interpretability
Captum is designed to offer interpretability for PyTorch models, especially for those in scientific research. It has multiple key features:
- Multi-modal: Supports models in various modalities.
- Built on PyTorch: Compatible with most PyTorch models and easy to incorporate.
- Extensible: Enables users to implement and benchmark new algorithms.
Its use cases involve understanding model predictions in scientific research with complex data such as images and text, as well as uncovering significant words or phrases in text-based models in natural language processing. For installation, the conda method is recommended for a smoother experience, though the pip option is also provided. To guarantee reproducibility, it is suggested to fix the random seeds as demonstrated in the examples.
