INFTY Documentation¶
INFTY is an optimization-centric toolkit for Continual AI. It provides plug-and-play optimizers and visualization tools for diagnosing optimization behavior in continual-learning systems.
The library is designed for research workflows where users need to replace or augment an existing PyTorch optimizer without rewriting the whole training pipeline.
Documentation map¶
Get started¶
| Page | Purpose |
|---|---|
| Installation | Install INFTY from PyPI or from source. |
| Quick Start | Learn the minimum optimizer-wrapping pattern and a small inline example. |
| Examples | Find runnable minimal scripts, visualization examples, and formal PILOT launcher scripts. |
Use INFTY¶
| Page | Purpose |
|---|---|
| User Guide | Understand optimizer families, closures, continual-learning workflows, and visualization utilities. |
| API Reference | Look up public classes, functions, arguments, return values, and expected closure contracts. |
| Troubleshooting | Resolve common installation, closure, training, and plotting issues. |
Project & Development¶
| Page | Purpose |
|---|---|
| Version Policy | Understand INFTY's public beta status, compatibility promises, and supported Python versions. |
| Developer Guide | Extend the library with new optimizers, plots, examples, and documentation. |
| Changelog | Track user-facing documentation, API, and packaging changes across releases. |
| Citation | Cite INFTY and related algorithm papers. |
Main concepts¶
INFTY follows four design ideas:
- Optimizer wrapping: most INFTY optimizers wrap a standard PyTorch optimizer such as
torch.optim.SGDortorch.optim.Adam. - Closure-based training: the training loss is supplied through a closure that returns
(logits, loss_list). - Continual-learning awareness: several optimizers explicitly handle flatness, gradient conflicts, or zeroth-order updates, which are common concerns in continual learning.
- Diagnostics as first-class utilities: the package includes tools for loss landscapes, Hessian ESD, gradient-conflict curves, and trajectory visualization.
Recommended reading order¶
New users should read:
Users integrating INFTY into an existing training framework should additionally read:
Contributors should read:
Current status¶
INFTY is in public beta. The documented infty.optim and infty.plot APIs are intended for stable incremental updates, while experiment integrations and benchmark scripts may evolve faster. For long-lived use, pin the package version or commit hash and review the version policy before upgrading.