Model Card for AntID-Tutor Genus Classifier
This model card documents the EfficientNet-B4 model used by the AntID Tutor platform to classify ant genus from user-submitted images. It supports educational outreach in myrmecology and assists students in exploring ant biodiversity.
Model Details
Model Description
This model is a fine-tuned EfficientNet-B4 convolutional neural network trained to classify ant images into one of 42 known genera. It powers the genus identification feature in the AntID Tutor application.
- Developed by: Terry Stilwell
- Funded by: Old Dominion University (as part of CS895 AI for Health & Life Sciences project)
- Shared by: tstilwel (https://huggingface.co/tstilwel/antID-tutor-genus)
- Model type: Image classification (CNN)
- Language(s) (NLP): N/A (Vision task)
- License: Apache-2.0
- Finetuned from model: torchvision.models.efficientnet_b4 (pretrained on ImageNet)
Model Sources
- Repository: https://github.com/tstilwell91/cs895_project_ants
- Demo: http://localhost:5002 (local Flask web app)
Uses
Direct Use
This model is used in a web-based learning platform where students upload ant specimen images to receive genus-level predictions along with interpretive educational content.
Downstream Use
Can be embedded into broader biodiversity platforms or entomological mobile apps with species distribution mapping or educational content.
Out-of-Scope Use
- Not suitable for species-level identification.
- Not intended for medical, pest control, or legal decision-making.
- May not generalize well to highly distorted or poorly illuminated images.
Bias, Risks, and Limitations
The model is trained on a curated subset of labeled ant images and may underperform on genera not present in the training set. Misclassification may occur with rare, damaged, or low-resolution samples.
Recommendations
- Use high-quality, centered images.
- Combine predictions with expert validation.
- Future work may include species-level training and regional ecological integration.
How to Get Started with the Model
You can test the model locally on the ODU Wahab cluster using the included inference script:
# Request GPU node (if needed)
salloc -p gpu --gres gpu:1
# Load PyTorch module
module load container_env pytorch-gpu/2.5.1
# Run inference
cd inference/genus
crun -p ~/envs/myrmecid python inference.py \
--image casent0901862_h_1_med.jpg \
--model genus_best_model_full.pth \
--classes classes.json
Training Details
Training Data
- Curated ant images with labeled genus from AntWeb and ODU image sets.
- 42 ant genera.
- Balanced sampling and augmentations applied (resizing, normalization).
Training Procedure
- Framework: PyTorch
- Architecture: EfficientNet-B4
- Loss: Cross-Entropy
- Optimizer: Adam
- Epochs: 100
- Batch size: 64
- Learning Rate: 0.0005
Evaluation
Testing Data, Factors & Metrics
Testing Data
Held-out validation set from the same sources as training.
Factors
Image quality, angle (head/dorsal/profile), lighting.
Metrics
- Top-1 Accuracy
- Confusion Matrix
- SHAP Visual Explanations
Results
- Accuracy: ~92%
Environmental Impact
- Hardware Type: A100 GPU (university cluster)
- Cloud Provider: None (on-prem HPC)
- Compute Region: USA (Old Dominion University)
Technical Specifications
Model Architecture and Objective
- EfficientNet-B4 with modified classifier head
- Objective: multi-class genus prediction (42 classes)
Compute Infrastructure
- Hardware: NVIDIA A100 GPU
- Software: PyTorch 2.1.0, torchvision, Python 3.9
Citation
APA: Stilwell, T. (2025). AntID Tutor: A Genus-Level Classifier for Ant Images Using EfficientNet-B4. Old Dominion University.
BibTeX:
@misc{stilwell2025antid,
author = {Terry Stilwell},
title = {AntID Tutor: A Genus-Level Classifier for Ant Images Using EfficientNet-B4},
year = {2025},
howpublished = {\url{https://huggingface.co/tstilwel/antID-tutor-genus}}
}
Model Card Contact
- Terry Stilwell
- GitHub: https://github.com/tstilwell91
- Email: tstilwel@cs.odu.edu