WardrobeWizard
End-to-end ML — training, evaluation, deployment
The Problem
Building a clothing classification system that actually works requires training a real model — not wrapping an API call. The challenge was creating an accurate 18-category classifier with limited training data using transfer learning.
Architecture
TensorFlow-based training pipeline using MobileNetV2 as the base architecture with ImageNet pre-trained weights. Custom classification head for 18 clothing categories. Django web application for inference and user interaction.
Key Decisions
Chose MobileNetV2 for its efficiency-accuracy tradeoff — lightweight enough for deployment while maintaining classification quality.
Used transfer learning with ImageNet weights to compensate for limited domain-specific training data.
Implemented comprehensive data augmentation (rotation, flip, zoom, brightness) to improve model generalization.
Trained for 25 epochs with learning rate scheduling to avoid overfitting on the small dataset.
Outcome
Working classifier across 18 clothing categories deployed in a Django web application. End-to-end ML pipeline demonstrating real model training, not API integration.