search

NVIDIA-AI-IOT / jetson_benchmarks

Jetson Benchmark
MIT License
372 stars 71 forks source link

FPS of unet segmentation is lower than expected #13

Closed pavloshargan closed 3 years ago

pavloshargan commented 3 years ago

I am using Jetson NX and trying to speed up the segmentation model(unet-mobilenet-512x512). I converted my tensorflow model to tensorRT with FP16 precision mode. And the speed is lower than I expected. Before the optimization i had 7FPS on inference with .pb frozen graph. After tensorRT oprimization I have 14FPS.

I have runned the benchmark on my Jetson NX, and unet 256x256 segmentation speed (speed of those .uff model, that is provided in repo) is really 146 FPS. I thought, the speed of my unet512x512 should be 4 times slower in the worst case. Maybe I should run inference in other way(without tensorflow) or change some converting parameters? Is it possible to find the pipeline with end-to-end way for optimizing and inferencing tensorflow models? I am looking for the solutions to get speed of my model(unet-mobilenet-512x512) close to 30FPS

ak-nv commented 3 years ago

Since this question is more on DL model optimization, could you please ask this on NVIDIA Developed Forum for Jetson devices.

Closing this as you could achieve expected performance with provided models. Please re-open if needed.