ARG-NCTU/rtdetr-spscd_buoy_joint_1280_b4_e100

RT-DETR Training Results

Model Information

• Model: rtdetr-l.pt
• Dataset: spscd_plus_buoy_yolo/dataset.yaml
• Classes: 13
• Image Size: 1280×1280
• Batch Size: 4
• Epochs: 100
• Workers: 16

Training Configuration

• Initial Learning Rate: 0.0001
• Final Learning Rate Factor: 0.01
• Patience: 50
• Mixed Precision (AMP): Enabled

Training Time

• Total Duration: 35h 26m 38s (127,598 seconds)
• Average per Epoch: 21.3 minutes (1,276 seconds)
• Hardware: NVIDIA RTX 5090 24GB

Final Test Results

• mAP50: 0.9925
• mAP50-95: 0.8596

Files in this Repository

• best.pt - Best model checkpoint
• results.csv - Complete training metrics
• results.png - Training curves visualization
• confusion_matrix.png - Confusion matrix
• F1_curve.png, PR_curve.png - Performance curves
• samples/ - Sample validation predictions

Note: TensorBoard logs are included if available. You can also visualize training using results.csv and the curve images.

How to Use

from ultralytics import RTDETR

# Load model
model = RTDETR('best.pt')

# Run inference
results = model('your_image.jpg')

Visualizing Training Progress

• View training curves in results.png
• Analyze per-class performance in confusion matrices
• Check results.csv for detailed epoch-by-epoch metrics