DRYTorch

Reproducible machine learning experiments with PyTorch.

Design

Don't Repeat Yourself (DRY) principles: replicable, documented, reusable.

• Reproducibility: experimental isolation to prevent unintended dependencies, data leakage, and misconfiguration.
• Modularity: flexible protocols preserving type inference in custom implementations.
• Decoupled Tracking: execution independent of tracking events (logging, plotting, and storing metadata).
• Optional Dependencies: support for external libraries (Hydra, W&B, TensorBoard, etc.) but minimal requirements.
• Self-Documentation: automatic metadata extraction and standardization.
• Ready-to-use: high-level implementations for advanced applications and workflows.

Installation

Requirements:

• The library only requires recent versions of PyTorch and NumPy.
• PyYAML and tqdm are recommended.

pip:

pip install drytorch

uv:

uv add drytorch

Package Structure

Modules are organized into the following subpackages:

• core: internal routines and the interfaces for library and custom components.
• lib: reusable implementations and abstract classes of the protocols.
• tracker: optional tracker plugins that integrate via the event system.
• contrib: community-driven extensions and support for external libraries.
• utils: general utilities independent to the framework.

Documentation

Read the full documentation on Read the Docs →

The documentation includes:

• Tutorials: walkthrough through the core features.
• API Reference: detailed API documentation.
• Architecture Overview: design structure.

See also

• Changelog