Fp16 models return no detections

[fnando1995]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hello, I used the validate and detect files for different models, changing fp32 and fp16 for tensorrt framework. I notice that fp16 models return NO detections (0 detects). Does anyone can point out what could this be? I make many imports ,changing fp, batchsize, model version, image size; but all fp16 models do not return detections, only fp32.

Additional

No response

[glenn-jocher]

@fnando1995 this is a CUDA/conda/windows bug that affects some install configurations. Most of the time this is caused by CUDA 11 incompatibilities and will be resolved by downgrading to CUDA 10, i.e. 10.2.

[fnando1995]

Hello @glenn-jocher, I tested in this 2 environments

Environment 1:
ubuntu 21.04
cuda 11.4
python3-pip

Environment 2:
Jetson nano jetpack 4.6
cuda 10.2
python3-pip

Any thoughts?

[DavidBaldsiefen]

Did you run the detect.py and val.py with the --half flag? export.py creates engines which need FP16 inputs in order to work.

[fnando1995]

@DavidBaldsiefen That was the problem!, but I was doing the change directly but, the argsparse default for half was changing it. Thanks.