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)
