RedenFlow help
Export formats, and which one you want
A trained model is a file, and which file depends entirely on where it is going to run. RedenFlow writes six of them.
The list you see on your own machine is shorter than the one below, and that is on purpose: a format only appears if it can actually be produced where you are sitting. TensorRT needs an NVIDIA card in the machine, CoreML needs a Mac. An option that would fail is worse than an option that is not offered.
| Format | Plan | What it is for |
|---|---|---|
| YOLO PyTorch.pt | Free | The model exactly as training produced it. Works anywhere Python and Ultralytics do, and it is what the Deploy wizard puts in the package. |
| ONNX.onnx | Pro | The one to pick when you do not know where it will end up. Almost every runtime on every platform reads it. |
| TensorRT.engine | Pro | NVIDIA hardware, and the fastest option on it. Only offered on a machine with an NVIDIA card, because an engine is built for the exact chip it runs on. |
| CoreML.mlmodel | Pro | iPhone, iPad and Mac. Only offered on a Mac, because that is where the converter runs. |
| TensorFlow Lite.tflite | Pro | Android and small boards. Small files, modest hardware. |
| OpenVINOfolder | Pro | Intel processors and their accelerators. Worth trying on an industrial PC with no graphics card in it. |
Smaller and faster: FP16 and INT8
Two checkboxes sit under the format list, and both trade a little accuracy for size and speed.
FP16, half precision
Halves the size of the numbers in the model. Smaller file, usually faster, and an accuracy cost most jobs never notice. Recommended by default for TensorRT.
INT8 quantisation
Smaller and faster again, with a larger accuracy cost and a calibration step. Worth it when the model would otherwise be too slow to keep up with the camera.
There is also an image size, and options to simplify the model and to allow a dynamic input shape.
You may not need to export at all
If the model is going onto a device rather than into your own code, the Deploy wizard does this for you. It writes a folder that runs, and on a Jetson or a Raspberry Pi it carries a script that builds the fast version of the model on the device itself.
Questions people ask
Which one should I pick?
If it is going on a specific device, pick the one for that device: TensorRT for NVIDIA, CoreML for Apple, TFLite for Android, OpenVINO for Intel. If you do not know yet, pick ONNX, because everything reads it.
Why is TensorRT missing from my list?
The list only shows formats that can actually be produced on the machine you are sitting at. TensorRT needs an NVIDIA card present, and CoreML only appears on a Mac. This is deliberate: an option that would fail is worse than an option that is not there.
What does FP16 change?
It halves the size of the numbers inside the model. The file gets smaller and it usually runs faster, with a very slight loss of accuracy that most jobs never notice.
And INT8?
Smaller and faster again, and a larger accuracy cost. It needs a calibration step, and it is worth it on hardware where the model would otherwise be too slow to keep up with the camera.
Does the free plan let me export at all?
Yes, as YOLO PyTorch (.pt) and as a YOLO dataset. That is a real, usable model, not a crippled one. The other five formats are Pro.
Do I need to export at all if I am using Deploy?
No. The Deploy wizard handles it: the package carries the model and, on a Jetson or a Raspberry Pi, a script that builds the fast version on the device itself.
Need a format that is not here? Ask.