I know TensorRT takes time to optimize the NN doing precision calibration, fusing etc. I am getting following build times.
~ 1400 layers: 5-10 mins
~ 2800 layers: 2.5 hrs
~ 14k layers: Haven’t finished yet (3 days and counting)
GPU: TITAN V
May I know the reason for such an exponential relation between number of layers and build time. Also is 14k layers feasible? Would be a tragedy if it fails from hereon.
Also any way I can speed up the process? (Tried changing max_workspace_size but in my experience that hardly changes the build time taken).
Thanks in advance!