Description
Hello, I am working on a program that utilizes TensorRT as part of a demo suite that we run. Specifically, we run code that is fairly similar to the SSD sample that is provided, and we run it against the SSD model that is provided with the sample.
Two years ago (2020/2019) this application worked fine, and we had inference and bounding boxes display.
Now, the application does not work - the confidence values for the image are always very low, like less than 1%. It’s actually quite frustrating because the only things that have changed between now and then is the CUDA version and TensorRT version. I was previously using CUDA 10.2 and TensorRT 6.x. I changed exactly one line of code to make it work between TensorRT 6.x and TensorRT 8.x - I changed a line to read “createNetworkV2(0)” instead of “createNetwork()”
Short of either backtracing to older versions of TensorRT and CUDA, or tracing the program line-by-line against the SSD sample, I’m kind of at a loss here. Are there any additional steps that are required for me to migrate the code?