beverlyhillscop/cotton-disease-qwen25vl-3b

Cotton Disease Detection - Qwen2.5-VL Models

Projekt-Übersicht

Vision-Language Model für Cotton Disease Detection auf Drohnen (Jetson Orin 64GB). Binäre Klassifikation: healthy vs diseased

---

Ergebnisse

Model	Accuracy	Inference	VRAM	Status
7B FP16	98%	291ms	15.4 GB	Getestet
7B INT4	98%	336ms	5.5 GB	Getestet
3B FP16	98%	264ms	7.13 GB	PRODUCTION
3B INT4	-	-	-	CUDA 13.1 nicht supported

Empfehlung: 3B FP16 für Jetson Deployment

---

Verzeichnisstruktur

/workspace/cotton-disease-detection/

• data/raw/Cotton Disease/train/ - Training Daten
• data/raw/Cotton Disease/val/ - Validation/Test Daten
• scripts/train_qwen25vl_lora_v2.py - 7B Training Script
• scripts/train_qwen25vl_3b_simple.py - 3B Training Script (EMPFOHLEN)
• scripts/test_3b_fp16.py - 3B FP16 Test Script
• outputs/qwen25vl_merged_v2/ - 7B Merged Model (~16GB)
• outputs/qwen25vl_3b_merged/ - 3B Merged Model (~7GB) PRODUCTION
• outputs/qwen25vl_q4_k_m.gguf - 7B GGUF INT4 (4.46 GB)

---

Training Commands

7B Model:

python3 /workspace/cotton-disease-detection/scripts/train_qwen25vl_lora_v2.py

• Base: Qwen/Qwen2.5-VL-7B-Instruct
• Time: ~55 min (A100)
• Batch: 4

3B Model (Empfohlen):

python3 /workspace/cotton-disease-detection/scripts/train_qwen25vl_3b_simple.py

• Base: Qwen/Qwen2.5-VL-3B-Instruct
• Time: ~157 min (A100)
• Batch: 8

---

Testing

python3 /workspace/cotton-disease-detection/scripts/test_3b_fp16.py

---

Jetson Deployment

Kopieren auf Jetson Orin 64GB:

• 3B FP16: outputs/qwen25vl_3b_merged/ (~7GB)
• Alternative: outputs/qwen25vl_q4_k_m.gguf (4.46GB für llama.cpp)

Erwartete Jetson Performance:

• 3B FP16 PyTorch: ~400-500ms
• 3B FP16 TensorRT: ~100-150ms (geschätzt)

---

LoRA Config

r=16, alpha=32, dropout=0.05 targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

---

Changelog

2025-01-29: 3B Model (98% Accuracy, 264ms, 7.13GB) 2025-01-29: 7B Model v2 (98% Accuracy)