Platform: TX2 Jetpack4.4/4.2
TensorRT: 7.1/5.0
CUDA: 10.2/10.0
Hello, about half mode plan file:
I created TRT plan file for yolov4-tiny, respectively with TRT5 and TRT7.
I set the Input layer as nvinfer1::DataType::kHALF, and set bingbuffers who will receive TRT inference’s output, also __half.
And I prepocess my data to __half before executing with created plan file.
Q:
When I run in TRT7 Jetpack4.4, I got the correct bboxes in number 4~5.
While I run int TRT5 Jetpack4.2, I got the wrong bboxes in number about 100.
But if set input layer as nvinfer1::kFLOAT, and host buffer as float* too, TRT5 will get the correct bboxes!
Why? Thanks!