fp32 vs FP16 vs int8 , both fp32 & fp16 having same results & fps?

[akashAD98]

i have seen your implementation,its by default supports fp16, & I tested the results on frames, there are lots of FP, so what i thout if i used FP32 (which is used in darknet) then there will be improvment in accuracuy.
But FPS & accuracy of both fp16 & fp32 is same.

i turned off fp16 using as per your suggetions #450
builder.fp16_mode = False & converted yolo.onnx into yolo.trt ,
& both speed (FPS) & results are almost same.

i think its not using fp32 . can you explain me how to use fp32

[jkjung-avt]

If you are using TensorRT 7+, you'll need to comment out (or delete) this line of code instead.

tensorrt_demos/yolo/onnx_to_tensorrt.py

Line 137 in fb2f075

config.set_flag(trt.BuilderFlag.FP16)

[akashAD98]

yes it's working fine. but no effect on accuracy, still it's the same. I thought due to FP32 little bit of accuracy will improve.

[jkjung-avt]

The accuracy issue (mAP drop for TensorRT YOLO engines) has been summarized/tracked in #315.