NutriConsistNet (V2)

Predicts total calories, mass, fat, carb and protein of a plate from ONE overhead RGB photo.

• Backbone: ResNet-50 (ImageNet pretrained)
• Structure: mass x per-gram density decomposition (totals = density * mass)
• Trained with an Atwater energy-consistency loss: calories = 4protein + 4carb + 9*fat (soft penalty, lambda = 0.1)
• V2 additionally supervises the density head and anchors mass more strongly
• Dataset: Nutrition5k, overhead RGB subset, official train/test split
• Input: one 320x320 RGB overhead photo, ImageNet normalization (no depth needed)
• Output order: [calories(kcal), mass(g), fat(g), carb(g), protein(g)]

Test results (official Nutrition5k test split, single-frame protocol)

Calories MAE: 44.7 kcal (17.5%) Mass MAE: 26.2 g (13.2%) Fat MAE: 3.4 g (26.5%) Carb MAE: 5.0 g (25.2%) Protein MAE: 4.3 g (24.6%) Atwater violation of final predictions: 8.3 kcal

How to load

import numpy as np, torch, torch.nn as nn, torchvision.models as tvm
# paste the model class from the training notebook, then:
means = np.load("train_means.npy")
model = NutriConsistNetV2(means)      # for the v2 file
# model = NutriConsistNet()           # for the e2 file
model.load_state_dict(torch.load("nutriconsistnet_v2.pt", map_location="cpu"))
model.eval()