How do I train my own classifier and import it into AddaxAI?
You have two main options for training your own classifier and using it in AddaxAI:
Option 1: Train with MEWC
The easiest way to train a compatible classifier is to use the MEWC pipeline, developed by researchers from the University of Tasmania. It’s built for species classification tasks and integrates smoothly with AddaxAI.
While it does require some familiarity with Python and Docker , it’s a solid entry point for anyone looking to integrate their own model. If you’re new to it, I recommend starting with the quick-start vignette.
If you follow the MEWC training workflow, your model should work out-of-the-box with AddaxAI, requiring only minimal setup.
Option 2: Use your own training method (PyTorch, TensorFlow, timm, YOLO, etc.)
Prefer to train using your own tools or architecture? That requires a custom inference script for integration.
Your script needs to tell AddaxAI:
- How to load your model (Line 197)
- How to crop the input image from MegaDetector bounding boxes (Line 233)
- How to preprocess and run inference on the image (Line 217)
Here’s an example you can use as a template:
classify_detections.py (DeepFaune example)
Relevant parts:
Once the script works, I can test it across our environments. If necessary, we’ll create a new one here:
AddaxAI environment build config
Final step: Share your model with others
Once your model runs successfully in AddaxAI, we’ll handle the admin side—adding metadata, citations, licenses, and so on. Then it can be published in the AddaxAI model zoo, ready for others to use. More on how to integrate your model here.