AiNaType-v8x

AiNaType-v8x Release Notes

Overview

This release introduces the AiNaType-v8x model for wildlife image classification, featuring improved accuracy and robust generalization. The model has been trained for 100 epochs using a variety of augmentations and regularizations, with support for advanced features such as overlap masking and RandAugment.

Model Training Configuration

• Model: AiNaType-v8x
• Epochs: 100
• Batch size: 128
• Image size: 224x224
• Optimizer: SGD
• Augmentation: RandAugment, erasing, mosaic

See args.yaml for full reproducibility and hyperparameters.

---

Performance Metrics

1. Training & Validation Curves

• Train loss: Rapid decrease and stabilization below 0.1, indicating effective learning and minimal overfitting.
• Validation loss: Consistently low, with a slight increase at later epochs suggesting strong generalization.
• Top-1 Accuracy: Peaks at ~95%, maintaining above 94% for most epochs.

2. Confusion Matrix

• Live animal, dead animal, tracks, scat, other, background classes are well separated.
• Class accuracy rates:
  • Dead: 95%
  • Live animal: 96%
  • Other: 91%
  • Scat: 90%
  • Tracks: 98%

3. Sample Predictions