An error occurred with int8

[chrichard]

I use deepstream 6.2 and yolov8, and when I convert to int8, the conversion process is normal. But at the time of object detection, no targets were detected. In the DeepStream 6.2 environment, other conversions are normal, such as YOLO8 to FP16, or YOLO7 to INT8.

[chrichard]

Today I updated deepstream-yolo. Further testing found that in the Deepstream 6.2 environment, if batch_size=1, the int8 model can output the detection results. However, when batch_size=16, there is no detection box output.

[botjeboon]

I'm having the exact same problem, everything runs fine but no output. I didn't change the batch_size (so it's still set to 1). You're saying that it didn't work when you had batch_size=16, but it did work when batch_size=1?

Was there anything else you might have changed?

[chrichard]

If it is FP16 mode, batch_size whether it is 1 or 16, it is normal。But in int8 mode, batch_size only 1 is normal

[marcoslucianops]

The DeepStream is a bit confusing about batch-size. According to NVIDIA, batch-size should be equal to the number of sources on primary gie and streammux.