CNN Food Classifier

The model lives at Hugging Face Spaces and lets anyone upload a food photo to get instant predictions across 101 categories — from apple pie to waffles. It identifies the most likely dish and surfaces the top 3 predictions with confidence scores.

The architecture is built on MobileNetV2 pretrained on ImageNet, which already knows how to detect edges, textures, and shapes from 1.2M images. Rather than training from scratch — which would require far more data and compute — I reused those learned features and trained only what needed to be food-specific.

Stage 1 freezes the entire MobileNetV2 base and trains a lightweight classification head on top: GlobalAveragePooling → Dense(256, ReLU) → Dropout(0.3) → Dense(101, softmax). This converges quickly and reaches ~62% validation accuracy in 20 epochs at lr=0.0001.

Stage 2 unfreezes the top 30 layers of the base and fine-tunes the full model end-to-end at lr=1e-5 — gentle enough to refine the pretrained weights toward food-specific features without overwriting them. This pushes validation accuracy to ~68.3%.

The training pipeline uses tf.data with mixed float16 precision throughout for ~30% faster throughput on Apple M1 Metal GPU. The app is deployed as a Gradio interface on Hugging Face Spaces — users upload a photo, the model preprocesses it to 224×224, normalises pixel values, and returns the top 3 predicted classes with confidence scores.