6.8 KiB

Raw Blame History

🚀 Model Retraining Guide - 4 Classes (Healthy Added)

📋 Perubahan Terbaru

Sistem klasifikasi telah diupdate untuk support 4 kelas daripada 3:

✅ Bacterialblight (Bercak Bakteri)
✅ Brownspot (Bercak Coklat)
✅ Healthy (BARU - Daun Sehat)
✅ Leafsmut (Jamur Daun)

File yang Sudah Diupdate

Backend: web_TA/scripts/rice_inference.py - CLASS_NAMES updated
Laravel: web_TA/app/Http/Controllers/ClassificationController.php - Added Healthy disease info
Mobile: mobile_TA/padi_app/lib/screen/result_screen.dart - UI support untuk Healthy class

🔧 Langkah Retrain Model

Step 1: Siapkan Dataset

Pastikan folder struktur ada:

rice leaf diseases dataset/
├── Bacterialblight/     (foto daun bercak bakteri)
├── Brownspot/           (foto daun bercak coklat)
├── Healthy/             (BARU - foto daun sehat)
└── Leafsmut/            (foto daun jamur)

Dataset Healthy harus contain:

Daun padi yang SEHAT (tanpa penyakit)
Minimal 100-200 foto berkualitas tinggi
Beragam kondisi pencahayaan & angle
Resolusi minimal 224x224 pixel

Step 2: Bersihkan Dataset (Optional tapi Recommended)

Hapus gambar duplikat/corrupted dari Jupyter notebook:

# Jalankan cell "Menghapus Gambar Duplikat" di notebook
PATH_DATASET = r"d:\PROJECT TA\rice leaf diseases dataset"
hapus_gambar_duplikat(PATH_DATASET)

Step 3: Training di Jupyter Notebook

File: rice leaf diseases dataset/rice_leaf_cnn_classification.ipynb

3a. Buka Notebook

cd "d:\PROJECT TA\rice leaf diseases dataset"
jupyter notebook rice_leaf_cnn_classification.ipynb

3b. Jalankan Cell Berturut-turut (PENTING - ikuti order!)

Cell 1: Setup & Imports

# Imports yang diperlukan
import os
import warnings
warnings.filterwarnings('ignore')

Cell 2: Define Classes

HARUS sesuai dengan folder di dataset:

classes = ['Bacterialblight', 'Brownspot', 'Healthy', 'Leafsmut']
IMG_SIZE = 224

Cell 3: Load Dataset

Scan folder dan count gambar per class
Output harus menunjukkan:
- Bacterialblight: N images
- Brownspot: N images
- Healthy: N images (BARU)
- Leafsmut: N images

Cell 4: Data Augmentation

Augment gambar untuk variance yang lebih baik

Cell 5: Model Building

Build MobileNetV2 transfer learning model
4 output classes (index 0-3)

Cell 6: Training

Train untuk 50-100 epochs
Monitor accuracy & loss
Tunggu sampai selesai (~30 menit - 2 jam)

Cell 7: Evaluation

Test model accuracy
Check confusion matrix

Cell 8: Save Model

model.save('rice_leaf_disease_model.keras')
# atau .h5 jika ingin format lama

✅ Validasi Model Baru

1. Check Model File Exists

cd "rice leaf diseases dataset"
ls -la rice_leaf_disease_model.keras
# File harus ada & > 30MB

2. Test dengan Python Script

cd web_TA
$json = @{image="[base64-image-here]"} | ConvertTo-Json

# Test classify
$json | python scripts/rice_inference.py classify --model-dir "../rice leaf diseases dataset"

Output harus format:

{
  "success": true,
  "predicted_class": "Healthy",
  "confidence": 0.95,
  "all_predictions": {
    "Bacterialblight": 0.01,
    "Brownspot": 0.02,
    "Healthy": 0.95,
    "Leafsmut": 0.02
  },
  "leafiness": 0.45
}

3. Health Check Endpoint

curl http://localhost:8000/api/health/
# Response:
# {
#   "healthy": true,
#   "health": {
#     "python_model": {
#       "status": "ok",
#       "classes": ["Bacterialblight", "Brownspot", "Healthy", "Leafsmut"]
#     }
#   }
# }

🧪 Test Mobile App

1. Upload Foto Daun Sehat

Open app
Click "Ubah Foto"
Upload foto daun sehat
Harus detect sebagai "Daun Sehat" (hijau color)
Severity: "Tidak Ada" (green badge)

2. Upload Foto Penyakit

Upload foto daun sakit
Harus detect penyakit specifik
Severity: Normal (red/orange/yellow)

3. Test Concurrent Requests

Multiple foto sekaligus
Tidak boleh ada error "gagal menghubungi model"

📊 Training Performance Tips

Meningkatkan Accuracy

Dataset Quality - Lebih penting dari quantity
- Foto clear, bien-lit
- Tidak ada gambar upside-down/rotated
- Consistent background (daun only, no tools)
Data Augmentation - Notebook sudah punya
- Rotation, Flip, Zoom
- Brightness/Contrast adjustment
Class Balance - Pastikan semua class balanced
- Ideal: ~200-300 per class
- Min: 50 per class
Training Parameters

# Already optimized di notebook:
IMG_SIZE = 224  # MobileNetV2 optimized size
learning_rate = 0.0001  # Small LR = stable learning
epochs = 50-100
batch_size = 32

Jika Accuracy Rendah

Tambah dataset (lebih banyak foto)
Improve data quality
Increase epochs (sampai loss plateau)
Check class imbalance (use confusion matrix)

🚨 Common Issues & Fixes

Issue: "Classes mismatch"

Error: IndexError: list index out of range saat inference

Solusi:

# Pastikan di notebook CLASS_NAMES updated:
classes = ['Bacterialblight', 'Brownspot', 'Healthy', 'Leafsmut']  # 4 classes

Issue: Model Predicts Wrong Class

Penyebab:

Dataset Healthy tidak representative
Too little training data
Model overfit