Less accuracy while using engine file

[Hemanth-TS]

💡 Your Question

I've trained a custom dataset with 4000 image , tried both pth and fp16. pth is way too slow to be used in realtime in our case, engine file is faster but way less accurate than it should be. I thought the accuracy would be similar to yolov8 but this is having some false detections even in the raw pth file, is there anyway to improve it? is there anyway to make fp32 engine file ?

Versions

No response