Polli Linnaeus Documentation Hub

Welcome to the official documentation for Polli Linnaeus, an open-source deep learning framework designed for taxonomic image classification and biodiversity monitoring applications. This hub provides a central point for accessing all documentation resources, whether you're looking to quickly use a pre-trained model, train your own custom models, or dive deep into the framework's architecture.

Getting Started

New to Polli Linnaeus? These guides will help you get up and running.

• Project README: Start here for a general overview of the project, its goals, and quick installation/usage notes.
• Installation Guide: Detailed instructions for installing Polli Linnaeus and its dependencies.
• Getting Started Tutorial: A beginner-friendly introduction to the basic functionalities of Linnaeus.

Using Pre-trained Models

Leverage our suite of pre-trained models for your classification tasks.

• Model Zoo: Discover available pre-trained models, their taxonomic scope, and links to Hugging Face Hub.
• Running Inference with Pre-trained Models: A step-by-step tutorial on how to use our pre-trained models for inference on your own images.
• Inference Overview: Learn about the core components of the Linnaeus inference pipeline, including the LinnaeusInferenceHandler.

Training Models

Guides for researchers and developers looking to train or fine-tune models.

• Training Your First Polli Linnaeus Model: A comprehensive tutorial on preparing your custom dataset (HDF5 format), configuring experiments, and launching training runs.
• Training Overview: An overview of the training system, including data loading, augmentations, multi-task learning, and more.
• Data Loading for Training: Detailed information on how Linnaeus handles data, with a focus on the HDF5 format.
• Phase 2: Abstention Training with RL (Experimental): Learn how to fine-tune pre-trained models to learn abstention behavior using Reinforcement Learning.

Datasets

Understanding the data used to train and evaluate models.

• Official Dataset Provenance (ibrida-v0-r1): Detailed information on the provenance and filtering logic used to create the initial batch of pre-trained models from the iNaturalist Open Data. This is crucial for understanding model scope and limitations.

Advanced Topics

For users who want to explore more specialized features of Linnaeus.

• Advanced Topics Index: Links to guides on:
  • Automatic Batch Sizing
  • Development Features & Debug Flags
  • Hierarchical Approaches (Taxonomy-Guided Label Smoothing, Hierarchical Heads)
  • Taxonomy Representation
  • Training Progress Tracking

Developer Documentation

Resources for those looking to extend or contribute to Polli Linnaeus.

• Model System Overview: Understand the architectural design, component registries, and how to create new model architectures.
• CI & Docker Guide: Learn about the continuous integration setup and Docker image architecture.
• Docker Build Guide: Detailed documentation on building and maintaining the Linnaeus Docker images.
• Development Guides: (Link to docs/dev/ directory if it contains an index or relevant files - e.g., docs/dev/01_training_loop_and_progress.md) Explore notes on the training loop, scheduling system, metrics, and other internal design choices.
• Contributing Guidelines: (To be created) How to contribute to the Polli Linnaeus project.

Known Limitations

• Known Limitations: Current limitations and known issues in Polli Linnaeus, along with recommended workarounds.

---

If you can't find what you're looking for, please consider opening an issue on our GitHub repository.