[question]: Multiple inference for single frame

[ManishAwale]

While using process.json file, setting the inputs in inference_input field for frame data is working for single inference in a frame.
I want to run multiple inferences for each detected objects of a frame as my model doesn't support batch processing. Can this be achieved using process.json file? Does inference_input supports list of inputs for multiple inferences?

[miguelaeh]

Yes, you can simply add multiple stages to your pipeline, each stage will have a process.json and will run inference on every frame

[ManishAwale]

Does this mean for every object detected from previous stage, i need to add stages dynamically?
what i am trying to do is: first stage detects objects in each frame. In second stage, for every detected objects, i need to run my model individually. For e.g. if 4 objects are detected, i need to run inference 4 times individually and add them as a final frame output.

I can do this in a process.py using a loop. But process.py doesn't allow me to run onnx model in CUDAExecutionProvider. So, I wanted to know if this is possible using process.json

[miguelaeh]

Oh, I understand it now. That cannot be achieved with process.json right now. You would need to do it using a python file. Why you can't run the model with CUDA execution provider?

[ManishAwale]

While running the ONNX model with CUDA execution provider, it is showing segmentation fault and failing during initialization itself