README.md

metadata

title: Plant Disease Assistant
emoji: 🌱
colorFrom: green
colorTo: yellow
sdk: gradio
sdk_version: 5.0.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Diagnose 22 crop diseases from a leaf. DINOv2-L on MI300X.
tags:
  - plant-disease
  - agriculture
  - dinov2
  - amd
  - mi300x
  - rocm
  - image-classification

🌱 Plant Disease Assistant

Snap a leaf, name the disease, get a fix.

Upload a photo of a plant leaf and this Space will identify which of 22 crop diseases it has, rate the model's confidence, and return a structured treatment + prevention guide.

What's under the hood

Component	Details
Classifier	DINOv2-Large (304M params, ViT-L/14), fine-tuned with a linear head
Accuracy	97.06% top-1, 0.9713 macro F1 on the held-out test split
Dataset	CCMT Crop Pest and Disease Detection — cashew, cassava, maize, tomato
Hardware	Fine-tuned on a single AMD Instinct MI300X (192 GB HBM3) via the AMD Developer Cloud
Framework	PyTorch 2.x + ROCm, timm
Knowledge base	Hand-curated treatment, prevention, and severity notes per class

The classifier runs on CPU here for accessibility — inference takes a few seconds per image. The original training run used ROCm on MI300X.

Crops & diseases covered

• Cashew — anthracnose, gumosis, leaf miner, red rust, healthy
• Cassava — bacterial blight, brown spot, green mite, mosaic, healthy
• Maize — fall armyworm, grasshopper, leaf beetle, leaf blight, leaf spot, streak virus, healthy
• Tomato — leaf blight, leaf curl, septoria leaf spot, verticilium wilt, healthy

How it was built

This is Track 2 of a multi-track entry in the lablab.ai AMD Developer Hackathon:

• Track 2 — Fine-tune DINOv2-L on CCMT for plant disease classification (this Space)
• Track 3 — Fine-tune Llama 3.2 11B Vision (LoRA) on the same data for conversational diagnosis (adapter on HF)
• Build in Public — Documented the journey end-to-end on social

Both tracks were trained on the same MI300X droplet, demonstrating that a single AMD GPU can comfortably handle both a 304M-param classifier and an 11B-param vision-language model in the same workflow.

Limitations

• Trained on a single dataset (CCMT) — performance on field photos with very different lighting, angles, or unseen crops will degrade.
• The treatment guidance is informational only and not a substitute for advice from a qualified agronomist or extension officer.
• CPU inference is intentionally slow (~5–10s/image). The original GPU pipeline runs in milliseconds.

License

Apache 2.0. Model weights and code are open. CCMT dataset licensing applies to the training data only.

Acknowledgements

• AMD for the Developer Cloud credits and MI300X access
• Meta for DINOv2
• The CCMT dataset authors
• lablab.ai for organizing the hackathon