Failure-Mode Classification in Dental Adhesion – Training Code

This repository contains the PyTorch training code used to fine-tune a modern vision backbone (e.g. DINOv3 ConvNeXt, ViT, DeiT, Swin) for failure-mode classification on optical microscopy images from shear bond strength (SBS) tests.

The code is organized as a small Python package with:

• Dataset loading from a directory of images + annotations.pt
• HuggingFace vision backbone with a configurable classification head
• GPU-side data augmentations (torchvision v2)
• MixUp / CutMix regularization
• Optional EMA of model weights
• Optional SLURM job scheduling helper

---

Typical run (without SLURM scheduling):

python main.py \
    --data-root /your/data/path/ \
    --augment-data \
    --image-size 1024 \
    --batch-size 16 \
    --num-workers 8 \
    --val-frac 0.2 \
    --output-dir checkpoints/baseline \
    --model-name facebook/dinov3-convnext-base-pretrain-lvd1689m \
    --head-type mlp \
    --mode freeze_then_unfreeze \
    --layers-to-unfreeze stages.3 \ 
    --freeze-epochs 5 \
    --num-epochs 200 \
    --base-lr 0.0003 \
    --discriminative-lr \
    --weight-decay 0.01 \
    --label-smoothing 0.1 \ 
    --tta \
    --ema \
    --mix-mode mixup_cutmix \
    --mixup-alpha 0.2 \
    --cutmix-alpha 1.0 \