Evaluate exported `.joint` object detection model on COCO

[mikel-brostrom]

I want to run inference on my exported INT8 .joint model locally (on my laptop). The idea here is to evaluate its mAP performance. I only see one way of doing this:

pulsar run\
    my_model.joint\
    --input resnet18_export_data/images/cat.jpg\
    --output_gt inference_results

And then read the generated .npy file. I want to avoid reading/writing from file during evaluation as it is a very time consuming operation. Is there a way of running inference using python? I don't see anybody wanting to deploy a model without knowing its performance... How do you calculated the performance of your exported INT8 .joint models?

[BUG1989]

Simulation running speed is very slow, we usually only run one picture to test on X86. Accuracy evaluation such as mAP? We all with do it on the board.

[mikel-brostrom]

Could you provide INT8 model COCO results so that people know what level of performance degradation is to be expected on the quantized models @BUG1989 ?