AlphaAkib/food-nutrition-multitask

Food Nutrition Multi-Task Model

ConvNeXt-Tiny backbone fine-tuned jointly for:

• Nutrition regression: total_calories, total_mass, total_fat, total_carb, total_protein
• Multi-label ingredient recognition (80 ingredient classes)

Trained on mmathys/food-nutrients (Nutrition5k-derived).

Test-set MAE: {'total_calories': 39.68, 'total_mass': 23.8, 'total_fat': 2.81, 'total_carb': 3.56, 'total_protein': 3.51} Test-set ingredient micro-F1: 0.881

See the accompanying notebook for the model definition (FoodNutritionModel) and how to load model_weights.pt + config.json for inference.