I intend to use a virtual machine instead of dual-booting my existing PC, simply because I don’t see any other use for Linux apart from Hailo model training. Since I have only one PC that I also need for other work, I’m considering using a virtual machine — unless retraining can be done directly on the Raspberry Pi.
What would be the minimum requirements needed to retrain models for specific datasets?
Please note we do not validate our software in a virtual machine. If you run into any issues we will not be able to support you. I would recommend that you check with your colleagues whether you can use another machine remotely if your machine is not sufficient.
That very much depends on the model. You will also need to convert a model to the HEF format. The Hailo AI Software Suite Docker requires a x86 machine running Ubuntu 22.04 or 24.0. You can find the System Requirements in the Hailo Dataflow Compiler User Guide.
While you can train and convert models without a NVIDIA GPU, it will require significant more time, and some model optimizations require a NVIDIA GPU.
@KlausK I aim to use Ultralytics model to add more objects into the model for detection. So what sort of resource would I need for re training?
While it’s possible to retrain a YOLO model on a mid-range x86 CPU without an NVIDIA GPU, the process may take more than a day to complete.
I recommend using a system equipped with an NVIDIA GPU with at least 16 GB of VRAM, 32 GB or more of system RAM, and SSD storage for optimal performance.
A better GPU and more RAM will allow you to shorten the training time.
@KlausK Thanks for the spec you have outlined how long it would take.
Ultralytics lets you train the model using a PC via Python, but it does not output the HEF file format. Is it possible to just train using my PC via Ultraonics and then use Linux for the final format conversion?
Models can be trained in any deep learning framework of your choice. After training, export the model to an ONNX or TFLite format, and then use the Hailo Dataflow Compiler to convert it into the HEF format for deployment on the Hailo device.
@KlausK So if we were to train the model as per the Ultralytics examples using Python, and then convert it to HEF, is this conversion process resource hungry or could I use a simple PC with Linux for that?
@KlausK
Sorry to repeat, but what sort of resource would be required to convert an Ultralytics model to HEF format,
The Hailo AI Software Suite Docker requires a x86 machine running Ubuntu 22.04 or 24.0. You can find the System Requirements in the Hailo Dataflow Compiler User Guide.
Thanks, but shows the requirement for the full training, does it take a long time to convert a model to Hef?
Converting a model takes anywhere from 10 minutes to a couple of hours. It depends on the size of the model, what optimization algorithm is used, whether you use performance mode and your machine.