Mohamed Amine Hajji / Projects

Gender Classification — Adience

Four transfer-learning experiments on the Adience benchmark, compared head-to-head — the winning EfficientNet-B2 + Mixup + TTA model reaches 98% test accuracy.

A gender-classification model on the Adience benchmark — real-world, unconstrained Flickr photos, not studio headshots — built with a teammate for an advanced AI course. Faces are detected and cropped with an OpenCV Haar Cascade pipeline, then four separate transfer-learning approaches (EfficientNet-B0, EfficientNet-B2, ConvNeXt-Tiny, and a 2-model ensemble) are trained and compared on the identical 12,981/3,247 train/test split rather than settling for the first result. The winning configuration — EfficientNet-B2 fine-tuned with Mixup and evaluated with 5-way Test-Time Augmentation — reaches 98% accuracy and F1-score on the held-out set; the errors that remain cluster around short hairstyles, off-angle faces, and genuinely ambiguous photos rather than being random.

Highlights

• Compared four transfer-learning pipelines (EfficientNet-B0, EfficientNet-B2, ConvNeXt-Tiny, and a 2-model ensemble) on an identical split — the ensemble, despite combining two strong backbones, didn’t beat the best single fine-tuned model.
• Best model — EfficientNet-B2 with Mixup, label smoothing, OneCycleLR, and 5-way Test-Time Augmentation at inference — reaches 98% accuracy and F1-score on 3,247 held-out images.
• Error analysis on the misclassified cases: mistakes cluster around short hairstyles, off-angle faces, and low-light photos rather than being random noise.

Stack: Python, PyTorch, OpenCV, EfficientNet-B2, scikit-learn