iamcode6/dinov2-l-ccmt-mi300x

DINOv2-Large — CCMT Crop & Disease (MI300X fine-tune)

Fine-tuned DINOv2-Large (304M params) on the CCMT crop-pest-and-disease dataset (22 classes across cashew, cassava, maize, tomato).

Trained on a single AMD Instinct MI300X using PyTorch + ROCm, as a submission to the lablab.ai AMD hackathon Track 2 — Fine-Tuning on AMD GPUs.

🌱 Try the live demo

This model is deployed as an interactive Gradio Space — upload a leaf photo and get an instant diagnosis with treatment guidance:

👉 [https://huggingface.co/spaces/iamcode6/merolav-space](https://huggingface.co/spaces/iamcode6/merolav-space)

Results

TTA rounds: 10.

Training

• —Backbone: DINOv2-L ViT-L/14 (self-supervised, LVD-142M pretrain)
• —Precision: bf16 (native MI300X)
• —Schedule: 2-phase — linear probe → full fine-tune with layer-wise LR decay
• —Optimizer: AdamW, cosine schedule, grad-clip 1.0
• —Augmentation: RandAugment + Mixup/CutMix + RandomErasing

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=22,
    img_size=224,
)
ckpt = torch.load("best.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["state_dict"])
model.eval()

Class index map is embedded inside the checkpoint under cfg; see the training repo for splits.json which defines the class_to_idx mapping.

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 — 22×22 confusion matrix
• —metrics.json — standard + TTA scores

Source

Training code: <https://github.com/genyarko/amd-merolav/tree/main/track2_finetuning>