Warning ignoring invalid prediction with bbox

[YoungjaeDev]

Hello I'm using the improved code, not the original code, but I'm not modifying the slide prediction core part
By the way, the warning appears on the slide dictionary. Is it supposed to appear? Or is there a problem?

04/14/2022 08:07:08 - WARNING - sahi.model - ignoring invalid prediction with bbox: [711, 0, 773, 0]

Thank you!

[fcakyon]

It means either you have implemented a ned detection pipeline incorrectly or your model is predicting invalid boxes. I.e a box shaped [711, 0, 773, 0] is predicted which is invalid.

[YoungjaeDev]

@fcakyon
So, is it right that my code is wrong?

[YoungjaeDev]

@fcakyon

The reason for the revision was that the yolov5 model currently supports only the form of AutoShape without supporting DetecMultiBackend If turn off the param autoshape, an error occurs

model = yolov5.load(self.model_path, device=self.device, autoshape=False)

[fcakyon]

@youngjae-avikus i am the maintainer of the yolov5 package. I am open for new ideas to implement in the yolov5 package. Can you explain the issue a bit more detailed? Why do you want to turn off the autoshape? What is the disadvantage of autoshape?

As far as I know you should always use autoshape=True for inference. It should be false only for the training.

[YoungjaeDev]

@fcakyon

First of all, can you sahi as an tensorrt engine file from the type loaded with Autoshape?

[fcakyon]

@youngjae-avikus yes, yolov5 can load tensorrt engine models in AutoShape mode. If you are having any errors loading tensorrt engine models with Autoshape=True, then open an issue in yolov5 repo and I will fix it.

[YoungjaeDev]

Ah, but Error below when loading engine from file 'inference_for_yolov5.ipynb'

detection_model = Yolov5DetectionModel(
    model_path=yolov5_model_path,
    confidence_threshold=0.3,
    device="cuda:0", # "cpu" or 'cuda:0'
    image_size=640
)

result = get_prediction("demo_data/small-vehicles1.jpeg", detection_model)

[fcakyon]

@youngjae-avikus this error should be fixed with this PR: fcakyon/yolov5-pip#108