Image classification of 158,846 food photos across 251 categories with TinyNet, a custom CNN designed to stay under one million parameters (997,763 in the Optuna-tuned variant). A self-supervised pretraining stage — reconstructing images with randomly masked black boxes, U-Net style — provides encoder weights that make training faster and more stable. The best model reaches 45.33% validation accuracy (micro-average F1 0.4533) over 251 classes.
Final project for the Supervised Learning course, MSc in Artificial Intelligence (University of Milano-Bicocca), with Andrea Borghesi.
The self-supervised pretext task: a random black box masks the image and the network learns to reconstruct it. Source: report, Fig. 2.
| Model | Validation accuracy |
|---|---|
| TinyNet + SSL encoder weights (best) | 45.33% |
| TinyNet, random init | 45.31% |
| Optuna-tuned TinyNet (no SSL / SSL) | 43.83% / 43.93% |
Micro-average F1 of the best model: 0.4533 (251 classes). Dataset split: 118,475 train (74.6%) / 11,994 val (7.5%) / 28,377 unlabeled test images used for the self-supervised task (17.9%). All numbers from the report, Sections 2 and 7.
- TinyNet: 5 convolutional blocks (Conv2d + GELU + BatchNorm + MaxPool) and 2 fully-connected layers, tuned with Optuna to 997,763 parameters (report, Table 2).
- Self-supervised pretraining: black-box masking + reconstruction for 60 epochs with a U-Net-inspired decoder; the encoder is then transferred to the classifier.
- Data augmentation: random rotations, translations, flips and affine transforms — no color transformations, to preserve food appearance.
- SSL weights improve convergence speed and stability; the final accuracy gain is small (+0.02%), which the report attributes to both variants having enough epochs to converge.
Training and validation curves of the SSL-initialized network, 150 epochs, no overfitting signs. Source: report, Fig. 4.
Accuracy heatmap across all 251 categories: a clear diagonal, with confusions concentrated among similar dishes. Source: report, Fig. 6.
The final project lives in Final_Project/: main.py is a
jupytext percent-format notebook with
the full pipeline, htuning.py runs the Optuna hyperparameter search.
Weekly assignments are under Assignments/A01–A09.
pip install torch torchvision optuna scikit-learn matplotlib jupytext
jupytext --to notebook Final_Project/main.py # or open main.py directly in VS Code / Jupyter
python Final_Project/htuning.py # Optuna searchFull write-up: Supervised_Learning__Final_project_.pdf — Mirko Morello, Andrea Borghesi, January 2025.