It looks like the documentation hasn’t been produced yet- but looking at the files in /assets/models, the .json file seems faily straightforward (although some of the items such as “postprocess” I’m not sure about)- but where does the .key file come from ??
I’ve got a custom .engine object detection model (exported from Yolov5- Onxx format also available) so if anyone can give me some pointers to use this in CVEDIA, I’d be most grateful
The model format need to be supported by our backends (onnx, openvino, tensorrt, etc) – call listnndevices within files folder to see what is supported
The model input / output need to be supported by our pre and post processing stack
The current version supports YoloX and YoloV4 as postprocessing stacks (you can see what’s available when cvedia-rt loads, it will transverse the plugins folder loading one by one)
Then to run your own model you need to create a .json file with the configs, you can grab one that already exists as an example, then change model_file key in the solution / instance to point to your model, for example, to run a local tensorrt model:
Keep in mind that tensorrt models need to be converted using the same cuda compute and tensorrt versions of the device you will be running them, otherwise it won’t work.
As for the .key file, this file only exists for models CVEDIA provides as security measure.
It would be great to have a simple path to use YoloV5 models in CVEDIA.
From my perspective- CVEDIA is an excellent tool for ‘Computer Vision made easy’, and YoloV5 has a similar motto/philosophy and is well documented. easy & reliable- so it would make perfect sense to support YoloV5 models in CVEDIA