iamcode6/amini-cocoa-dinov2-l-mi300x

vit_large_patch14_dinov2.lvd142m — Amini Cocoa Contamination (MI300X fine-tune)

Fine-tuned vit_large_patch14_dinov2.lvd142m on the Amini cocoa contamination dataset (3 classes: anthracnose, cssvd, healthy).

Trained on a single AMD Instinct MI300X using PyTorch + ROCm, as part of the AMD hackathon. This model uses on-the-fly bbox cropping with randomized context-padding to improve robustness against detector imprecision.

Results

Metric	This model (ViT-L / MI300X)
Test accuracy (TTA)	0.9284
Macro F1 (TTA)	0.9265
Standard acc	0.9274

TTA rounds: 10.

Training

• Backbone: vit_large_patch14_dinov2.lvd142m
• Precision: bf16 (native MI300X)
• Optimizer: AdamW, cosine schedule
• Augmentation: RandAugment + Mixup/CutMix + Random context-pad [0.0, 0.15]

See config.yaml for the full hyperparameter set.

Usage

import timm, torch

model = timm.create_model(
    "vit_large_patch14_dinov2.lvd142m",
    pretrained=False,
    num_classes=3,
    img_size=224,
)
ckpt = torch.load("best.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["state_dict"])
model.eval()

Artifacts

• best.pt — model weights + training config
• config.yaml — hyperparameters used for this run
• classification_report.txt — per-class precision / recall / F1
• confusion_matrix.csv — 3x3 confusion matrix
• metrics.json — standard + TTA scores

Source

Training code: https://github.com/genyarko/amd-merolav/tree/main/cocoa_amini_finetuning