low memory usage on Jetson TX2 using SSD mobilenet caffe model

Hi,
I am using SSD mobilenet caffe model to run object detection. The Tegrastats shows the following while running the model…

RAM 2891/7846MB (lfb 266x4MB) CPU [87%@2035,69%@2035,96%@2035,87%@2035,78%@2035,82%@2035] BCPU@42.5C MCPU@42.5C GPU@38.5C PLL@42.5C Tboard@35C Tdiode@37.25C PMIC@100C thermal@40.6C VDD_IN 10133/10079 VDD_CPU 5652/5473 VDD_GPU 382/378 VDD_SOC 993/1043 VDD_WIFI 38/44 VDD_DDR 1651/1660

  1. The RAM is not utilized fully. How to make sure full RAM capacity is used.
  2. How can I increase the GPU usage.

I am just getting 5 FPS on the Jetson TX2.

Opencv - 3.4.1
Model - MobileNetSSD_deploy.caffemodel

Thanks.

Hi,

  1. If you have multiple input images, it’s recommended to inference it at once with a larger batchsize value.
    This will increase the memory usage but achieve a better performance.

  2. It’s recommended to use our TensorRT engine to compile your model into GPU-optimized.
    More information can be found here:
    https://developer.nvidia.com/tensorrt

Thanks.

Hi AastaLLL,
I dont want to use the TensorRT for the ssdmobilenet caffe model as it is known to have problems with certain nodes.

  1. Is there any way I can see the GPU/CPU utilization graph plot to know how the GPU is loaded ?
  2. Do you recommend multi-threading for a larger batch size ? Or is there any other way I can increase the batchsize value ?
  3. Will I be able to define and use the GPU for certain tasks ?

Thanks.

Also, is there any way of know if the GPU is active and being used on the Jetson TX2 ?

Thanks

Hi,

1. You can check CPU/GPU utilization by tegrastats.

sudo ./tegrastats

2. Just set batch size > 1.

3. Sure. All the GPU tasks are required to implement with CUDA. Otherwise, it is a CPU job.

4. Check tegrastats for the information.

Thanks.

Thanks for the reply !

The batch size has to be set in the .prototxt file ?

Thanks.

YES.