DeiT-Small RGB Carbohydrate Regression - Stage 2 Fine-Tuned
This model predicts dish-level carbohydrate content from overhead RGB images from Nutrition5K.
Model version
This repository contains the stage 2 fine-tuned version of the model.
The main checkpoint is:
Dataset
- Nutrition5K
- Overhead RGB only
- Train / validation / test split from prepared metadata CSV
Architecture
- DeiT-Small distilled backbone pretrained on ImageNet
- Stage 1: frozen backbone + regression head training
- Stage 2: partial unfreezing of later transformer layers
- Regression head with linear output for carbohydrate prediction
Target
