FoodBev — ConvNeXt-Base food classifier
convnext_base.fb_in1k fine-tuned on Food-101 (101 dish classes). Stage 1 of the
FoodBev food-image-recognition and beverage-pairing project, a TUM Deep Learning and
Decision Making course project.
Results
Measured on the Food-101 validation split (25,250 images, 250 per class), which is the
standard benchmark test set. It was held out of every checkpoint and early-stopping
decision and evaluated exactly once, after training finished.
| Metric | Value |
|---|---|
| Top-1 accuracy | 91.65% |
| Top-5 accuracy | 98.47% |
| Macro F1 | 0.9166 |
| Expected calibration error | 0.0791 |
Selected from a controlled 14-run comparison across four families (ConvNeXt, EfficientNet, ViT, ResNet) and three scale tiers, all trained under one shared recipe: batch size 64, 60 epochs, AdamW, ImageNet-1k-only pretraining.
Details
| Backbone | convnext_base.fb_in1k |
| Parameters | 87.7M |
| Input resolution | 224x224 |
| Embedding dimension | 1024 |
| Classes | 101 |
Files
export.pt— weights plus the class-name list, input resolution and embedding dimension needed to rebuild the model. (upload pending)
Note on using this for pairing
Top-1 is 91.65% but top-5 is 98.47%. Roughly one photo in twelve gets the
wrong dish under a hard argmax, while the correct dish is in the top-5 about
sixty-four times out of sixty-five. Downstream consumers should prefer the probability
distribution over the single label.
- Downloads last month
- -