Model Card: EfficientNet V2-S — Wild Forest Animals

This document describes the image classifier used by the Wildlife Monitoring Dashboard (0HM340 Human–AI Interaction, TU/e). It follows the spirit of Model Cards for Model Reporting.

Model summary

Field	Value
Model family	EfficientNet V2-S (`torchvision.models.efficientnet_v2_s`)
Pre-training	ImageNet-1K (`EfficientNet_V2_S_Weights.IMAGENET1K_V1`)
Task	Multi-class image classification (single label per image)
Output	7 logits → softmax probabilities over fixed classes
Weights file	`efficientnet_v2_wild_forest_animals.pt` (state dict only; not shipped in the repo)
Input	RGB images resized to 224×224, ImageNet normalization
Framework	PyTorch

Intended use

Primary: Educational / research prototype for human–AI interaction and explainability (dashboard with ScoreCAM, LIME, nearest-neighbour views).
Deployment context: Simulated camera-trap workflow in a demo UI; not validated for real wildlife management, safety-critical decisions, or law enforcement.

Out-of-scope uses: Do not rely on this model for operational conservation decisions, species surveys with legal implications, or any setting where errors could cause harm without independent verification.

Output classes

Fixed label set (order matches the classifier head):

Index	Class
0	bear
1	deer
2	fox
3	hare
4	moose
5	person
6	wolf

The dashboard narrative may reference a specific national park; the model was not trained on data from that park only — see Training data.

Training data

Source: Roboflow project wild-forest-animals-and-person, workspace forestanimals, version 1, export format multiclass.
Local layout: wild-forest-animals-and-person-1/ with train/, valid/, test/ splits and _classes.csv per split (one-hot columns per class).
Label handling: Rows with multiple positive labels in the CSV use the first positive class only (single-label training).
Domain: Mixed camera-trap / wild-animal imagery bundled by the dataset authors; distribution across species, geography, lighting, and quality follows that dataset — not guaranteed to match any real park’s fauna or camera setup.

Training procedure

Implemented in train.py (see repository for exact defaults).

Setting	Default
Optimizer	Adam
Loss	Cross-entropy
Mixed precision	Enabled on CUDA (`autocast` + `GradScaler`)
Train augmentations	Random horizontal flip (p=0.5), then ToTensor, Resize 224, ImageNet normalize
Evaluation augmentations	ToTensor, Resize 224, ImageNet normalize
DataLoader shuffle (train)	Yes, with fixed generator seed
Reproducibility	`SEED = 42`; CUDNN deterministic mode enabled in training script

Default hyperparameters (CLI overridable): epochs 3, batch size 32, learning rate 1e-3. Example overrides: --epochs 5, --lr 0.0005, --batch-size 16.

Reported metrics: The training script prints validation loss/accuracy per epoch and test loss/accuracy after the last epoch. Exact numbers depend on run, hardware, and hyperparameters; record your own metrics when you train. Weights in the repo are not pinned to a single certified benchmark run.

Evaluation

Split: Held-out test folder from the Roboflow export.
Metric: Top-1 accuracy and cross-entropy loss on the test loader (see console output from train.py).
Limitations: No per-class confusion matrix or calibration analysis in the default pipeline; no external geographic or temporal holdout.

Ethical and fairness considerations

“Person” class: Predictions can affect privacy perceptions in camera-trap settings; treat as a coarse label, not identity or intent.
Wildlife labels: Errors could misrepresent which species are present; the UI supports manual verification — use it when stakes matter.
Deployment: Automated alerts (e.g. wolf/bear warnings in the demo) are illustrative; they should not replace expert assessment or park regulations.

Caveats and limitations

Domain shift: Performance will drop on images that differ strongly from the training distribution (new sensors, night IR, heavy occlusion, rare poses).
Single label: Images with multiple species only contribute one label during training; the model is not trained for multi-label detection.
Geographic / ecological claims: Class names refer to species types; the model does not prove an animal’s presence in any specific jurisdiction or ecosystem.
Weights: If you did not train the checkpoint yourself, treat reported behaviour as unknown until you evaluate on your data.
API keys / data download: Training and dashboard can auto-download data via Roboflow; use your own keys and comply with Roboflow terms in production-like setups.

How to reproduce

uv sync
uv run python train.py
# Optional: uv run python train.py --epochs 5 --lr 0.0005 --batch-size 16

This produces efficientnet_v2_wild_forest_animals.pt compatible with dashboard.py and main.py.

Citation / contact

Course / context: 0HM340 Human–AI Interaction, Eindhoven University of Technology.
Base architecture: Tan & Le, EfficientNetV2, 2021 (via torchvision).
For questions about this card or the codebase, refer to the project README.md and DEVELOPER_GUIDE.md.

Last updated to match repository layout and training script defaults; update this file if classes, dataset version, or training recipe change.

5.7 KiB Raw Blame History Unescape Escape