Problem getting published inference time on single GPU: Compilation of "script/view_tranformation_cuda" or TensorRT necessary?

[Seyd2]

Using model M0 I'm able to achieve 2 fps using the provided checkpoints. I just realized that the repo contains a folder with compilable code for the view transformer and would like to ask if the compilation is necessary to get to achieve the published 50 fps on Tesla T4? If not, is a TensorRT implementation necessary to get the published fps?

Else I could be running the test wrong. I'm using run_fastbev.sh without slurm as:

function test {
    GPUS=$1
    EXPNAME=$2
    RESUME=${3:-work_dirs/fastbev/exp/$EXPNAME/epoch_20.pth}

    echo test; sleep 0.5s
    python ./tools/test.py \
        configs/fastbev/exp/$EXPNAME.py \
        $RESUME \
        --eval bbox \
        --out work_dirs/fastbev/exp/$EXPNAME/results/results.pkl \
        --launcher pytorch \
}

test 1 paper/original_configs/fastbev_m0_r18_s256x704_v200x200x4_c192_d2_f4

with comamnd line:

./tools/fast_bevrun.sh