Macro F1
Computes F1 independently for each class and averages the results, so the larger nature class cannot dominate the score.
01
Select an image
02
Prediction
The ranked probabilities will appear here.
Predicted class—
Confidence—
Inference—
03
Model evaluation dashboard
Test-set results generated by the local RTX experiment. The deployment model is selected using macro F1 while respecting the browser model-size limit.
Selected for deployment
—
Evaluation results are loading.
Controlled model comparison
All candidates use the same stratified split, test set, seed, and evaluation framework.
| Model | Accuracy | Macro F1 | Weighted F1 | Top-2 | ROC-AUC | ECE | Size | Latency |
|---|---|---|---|---|---|---|---|---|
Selected model by class
Precision, recall, and F1 reveal performance hidden by overall accuracy.
| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
Robustness checks
Macro F1 under controlled image corruptions.
How to interpret these metrics
Balanced accuracy
Averages recall across classes and is more informative than raw accuracy when class frequencies differ.
ROC-AUC
Measures how well predicted probabilities separate each class from the others across thresholds.
Expected calibration error
Measures the gap between confidence and observed correctness. Lower is better.
04
What this project proves
From scratch
Simple CNN and AlexNet-style CNN architectures are trained directly on the target data.
Transfer learning
Frozen and partially fine-tuned MobileNetV2 models test practical feature reuse.
Evaluation
Macro F1, calibration, robustness, error analysis, Grad-CAM, latency, and model size determine the final model.
Deployment
The selected PyTorch model is exported to ONNX and served through ONNX Runtime Web on GitHub Pages.
05