v1/README.md

Billiards YOLOv8n Model - Exported Formats (1280x1280)

Successfully Exported Models

CoreML

• File: billiards-yolov8n.mlpackage
• Input Size: 1280x1280 RGB
• Size: ~6.0 MB
• Status: ✅ Successfully exported
• Platform: iOS/macOS
• Description: CoreML package ready for Apple device inference

ONNX

• File: billiards-yolov8n.onnx
• Input Size: 1280x1280 (1x3x1280x1280 BCHW)
• Format: ONNX opset 19
• Size: 12.2 MB
• Status: ✅ Successfully exported
• Platform: Cross-platform
• Description: Universal format, can be used with ONNXRuntime or converted to other formats

TFLite (Float32)

• File: billiards-yolov8n_float32.tflite
• Input Size: 1280x1280
• Size: 12 MB
• Status: ✅ Successfully exported
• Platform: Android, Edge devices, TensorFlow Lite runtime
• Description: Full precision TFLite model

TFLite (Float16)

• File: billiards-yolov8n_float16.tflite
• Input Size: 1280x1280
• Size: 6.0 MB
• Status: ✅ Successfully exported
• Platform: Android, Edge devices, TensorFlow Lite runtime
• Description: Half precision TFLite model (smaller, slightly less accurate)

Model Details

• Base Model: YOLOv8n (Nano)
• Training: Custom billiards dataset
• Input Size: 1280x1280x3 RGB
• Task: Object detection
• Output Shape: (1, 6, 33600) - 6 classes, 33600 detections at 1280 resolution

Conversion Notes

CoreML

• Converted directly from PyTorch using Ultralytics export
• Ready for use with Core ML framework on iOS/macOS
• Optimized for Apple Neural Engine

ONNX

• Exported with opset 19 for broad compatibility
• Can be used as an intermediate format for other conversions
• Works with ONNXRuntime on all platforms

TFLite

• Converted via ONNX → TensorFlow SavedModel → TFLite pipeline using Docker
• Used onnx2tf tool with TensorFlow 2.19
• Two variants provided:
  • Float32: Full precision, larger file
  • Float16: Half precision, 50% smaller, minimal accuracy loss

Usage Examples

CoreML (Swift/iOS)

let model = try billiards_yolov8n()
let input = billiards_yolov8nInput(image: pixelBuffer)
let output = try model.prediction(input: input)

ONNX (Python)

import onnxruntime as ort
session = ort.InferenceSession("billiards-yolov8n.onnx")
output = session.run(None, {"images": input_tensor})

TFLite (Python)

import tensorflow as tf
interpreter = tf.lite.Interpreter("billiards-yolov8n_float32.tflite")
interpreter.allocate_tensors()
# Run inference...

File Sizes Summary

Format	File	Size
PyTorch	billiards-yolov8n.pt	5.9 MB
CoreML	billiards-yolov8n.mlpackage	~6.0 MB
ONNX	billiards-yolov8n.onnx	12.2 MB
TFLite (FP32)	billiards-yolov8n_float32.tflite	12 MB
TFLite (FP16)	billiards-yolov8n_float16.tflite	6.0 MB