YOLO26 Export Format Evaluation¶

Phase 1.4: Evaluate export formats for YOLO26 deployment and recommend the best option.

NOTE (2026-09-23): point-in-time evaluation kept as the measured record. Its ai-yolo26: compose excerpts describe the standalone server, which was retired fully that day (owner ruling — Triton on ai-gateway serves yolo26; recipe in archive/ai-yolo26-image/). Export guidance itself still applies via ai/gateway/export/export_yolo26.py and ai/yolo26/build_engine.py.

Date: 2026-01-26 Model Evaluated: yolo26n.pt (Nano variant) Source Directory: /export/ai_models/model-zoo/yolo26 Export Directory: /export/ai_models/model-zoo/yolo26/exports

Executive Summary¶

This document evaluates export formats for YOLO26 deployment in the home security pipeline. Key findings:

1. ONNX export successful - 9.48 MB, 3.33s export time
2. TensorRT not available - Requires CUDA-enabled PyTorch
3. PyTorch baseline - 35.0ms CPU inference, 5.29 MB file size
4. Recommendation: Use TensorRT for production (NVIDIA), ONNX for portability

System Information¶

Available Hardware (Not Used Due to CPU PyTorch)¶

Export Results¶

Summary Table¶

*TensorRT size is estimated based on typical FP16 compression

ONNX Export Details¶

Export Configuration:

from ultralytics import YOLO

model = YOLO("/export/ai_models/model-zoo/yolo26/yolo26n.pt")
model.export(format="onnx", imgsz=640, simplify=True, opset=17)

Export Metrics:

• ONNX IR Version: 8
• Opset Version: 17
• Producer: pytorch 2.9.1
• Input Shape: (1, 3, 640, 640) BCHW
• Output Shape: (1, 300, 6) - 300 detections with [x, y, w, h, conf, class]
• Export Time: 3.33s
• Simplification: Enabled

Output Location: /export/ai_models/model-zoo/yolo26/exports/yolo26n.onnx

TensorRT Export (Reference)¶

TensorRT export was not performed due to CPU-only PyTorch in the development environment.

For production deployment:

# Requires CUDA-enabled PyTorch
from ultralytics import YOLO

model = YOLO("/export/ai_models/model-zoo/yolo26/yolo26n.pt")
model.export(format="engine", imgsz=640, half=True)  # FP16 for speed

TensorRT Requirements:

• CUDA 11.8+ or 12.x
• cuDNN 8.6+
• TensorRT 8.6+ (or TensorRT 10.x for CUDA 12)
• NVIDIA GPU with compute capability 7.0+
• PyTorch with CUDA support (e.g., torch+cu121)

Inference Speed Benchmarks¶

Test Configuration¶

• Test images: 5 security camera images from backend/tests/fixtures/images/pipeline_test/
• Warmup runs: 3
• Timed runs: 10 per image (50 total inferences)
• Hardware: CPU only

PyTorch Results¶

ONNX Results¶

Not benchmarked - ONNXRuntime does not support Python 3.14 as of January 2026.

Expected performance (based on Ultralytics documentation):

• CPU: ~30-40ms (similar to PyTorch)
• GPU with CUDA provider: ~10-15ms
• GPU with TensorRT provider: ~5-8ms

Estimated GPU Performance¶

Based on typical RTX A5500 performance:

Detection Correctness¶

Verification Status¶

Detection correctness verification was not completed due to ONNXRuntime unavailability on Python 3.14.

Manual Verification¶

The ONNX export was validated by:

1. Checking file integrity and size
2. Loading with onnx.load() - successful
3. Verifying ONNX metadata (IR version, opset, producer)

Recommendations¶

Best Format by Use Case¶

For This Project (Home Security)¶

Primary recommendation: TensorRT

Given the NVIDIA RTX A5500 in the production environment:

1. Export to TensorRT with FP16 for optimal GPU performance
2. Keep PyTorch model for development and debugging
3. Export to ONNX as portable fallback

Deployment configuration:

# docker-compose GPU configuration
ai-yolo26:
  environment:
    DETECTOR_MODEL: /models/yolo26/exports/yolo26n.engine
    DETECTOR_DEVICE: cuda:0
    DETECTOR_HALF: true

Version Requirements¶

For ONNX Export¶

For TensorRT Export¶

For ONNX Inference¶

Note: ONNXRuntime does not support Python 3.14. Use Python 3.12 or 3.13 for ONNX inference.

Known Limitations¶

1. Python 3.14 compatibility

2. ONNXRuntime does not support Python 3.14

3. Workaround: Use Python 3.12 or 3.13 for production inference

4. CPU-only PyTorch in dev environment

5. TensorRT export requires CUDA-enabled PyTorch

6. Solution: Install torch+cu121 or torch+cu124 for GPU support

7. ONNX file size

8. ONNX files are ~80% larger than PyTorch
9. This is expected due to explicit operator representation
10. TensorRT typically produces similar or smaller files after optimization

Export Script Usage¶

The export script is located at scripts/export_yolo26.py.

Basic Commands¶

# Export all formats (ONNX + TensorRT if available)
uv run python scripts/export_yolo26.py

# Export ONNX only
uv run python scripts/export_yolo26.py --format onnx

# Export specific model
uv run python scripts/export_yolo26.py --model yolo26s.pt

# Force re-export
uv run python scripts/export_yolo26.py --force

# Benchmark only (skip export)
uv run python scripts/export_yolo26.py --benchmark-only

# Skip verification
uv run python scripts/export_yolo26.py --skip-verification

Output Files¶

• Exported models: /export/ai_models/model-zoo/yolo26/exports/
• Benchmark report: docs/benchmarks/ directory (appended)
• Local report: /export/ai_models/model-zoo/yolo26/exports/EXPORT_REPORT.md

Related Documentation¶

• Model Zoo Benchmark
• YOLO26 Validation Report