Analyze_videos is not using GPU, but training is

When I ran the training step, I observed much faster than CPU-based training and a utilization of my GPU of about 65% processing and 95% VRAM, so I know the GPU is usable by the system. I did not provide any special instructions to the training command to utilize the GPU.

Running “analyze_videos( path_config_file, videofile_path, gputouse = 0 )” does not appear to use my GPU (0% utilization), and will apparently take at least 4 hours for a 15 minute video.

Meanwhile, my CPU utilization hovers around 1400% for ipython during analyze_videos. This was true both with and without the “gputouse = 0” argument.

What am I doing wrong?

Is the analyze_videos function able to use the GPU?

Hardware :

Nvidia GeForce RTX 2070 Super w/ 8 GB VRAM
AMD Ryzen 7 3700X (8 cores)

Software :

Ubuntu 20.04
Docker Instance of DLC


I performed some further investigation.

It appears that if the computer suspends itself to save electricity at night, the Docker instance loses track of the GPU. (Running the device_lib.list_local_devices() function reports only the CPU after a “suspend” operation.)

To fix this, I restarted the computer, restarted the Docker image, re-entered the Docker image instance and then saw that the GPU was again reported.

Now the GPU is being utilized to analyze the video and instead of 5+ hours it estimates about 25 minutes.

Maybe starting and stopping the Docker image would work without restarting, but I haven’t tried that.