Running Docker on Mac M1 with USB Coral Accelerator

[jimshaggy]

Good day all.

Relatively new to Frigate and Home Assistant, really really new to Docker. I am attempting to move my Frigate off of my ODroid-N2+ and onto a separate dedicated machine.

While Frigate and the USB Coral Accelerator works on my ODroid, it's pretty heavy and keeps my CPU pinned. I've decided to try to move it, which has necessitated learning about Docker and getting that setup on my Mac Mini M1. That has been relatively straight forward and I've been successful getting Frigate working, but only with CPU detectors. I haven't been successful in getting the USB Accelerator connected. Any hints or success stories?

Thanks Jim

[t0ny-peng]

You need to find a way to mount the usb device into the virtual machine that runs docker(Hyperkit?), or create another virtual machine(e.g., using Virtualbox) and mount the USB.

[bmclapham1973]

Were you able to use your USB Coral with you M1 Mac? I am running 3 1920x1080 cameras with, CPU only, at the moment and I am seeing inference speed consistently 9-10ms.

The M1 seems to do a really good job running frigate. I am not sure I would need to use a Coral with my M1 Mac mini.

[skynet01]

Another confirmation with M1 Mac mini. Have 8 cameras and get around 11ms when there is motion on a few cameras. Also, I am running my docker inside Ubuntu 20 on Parallels with 16 other containers :) They should update the doc saying that running on CPU is ok if you have an M1 chip

[skynet01]

Another update, after running this setup for a month I do have times when inferences spike all the way from 10 to 50 and even 200 sometimes. Here is the inference history graph

so I figured I’d try the coral but I can’t pass usb to docker and when I pass it through parallels it crashes every 30 seconds or so. Has anyone got usb coral to run successfully on Mac m1?

[NickM-27]

Mac doesn't support usb passthrough to docker, you'd need to run a Linux VM and pass it through that way

[skynet01]

Running it inside the Linux vm makes coral crash every 15 seconds. Would be nice if there was a way to install Frigate on Mac mini directly

[NickM-27]

Running it inside the Linux vm makes coral crash every 15 seconds. Would be nice if there was a way to install Frigate on Mac mini directly

The VM should work might just be something wrong with the way it was setup.

In any case, at this time we're not looking to support / maintain a bare metal install

[knaaptime]

thanks for this thread, everyone! I was getting really poor performance before realizing i was using the wrong container (amd64 instead of aarch64). One thing I was curious about is, if I manage to get a different version of ffmpeg into the mix, compiled with h264_videotoolbox (e.g. the current version [5.0] from homebrew), is that likely to provide a separate boost in the encoding as well? thanks for any tips

[digaus]

Anyone made a direct comparison between the Google Coral and Mac Mini M1 CPU? Would be really interesting as I am planning to use my Mac Mini M1 (used normally for iOS development) for Frigate and I do not really want to buy a Coral if it is not necessary.

Also any info about hwaccel_args for Mac Mini M1?

[NickM-27]

Hmm, the ffmpeg in 0.11-rc doesn't seem to have video toolbox support enabled:

What did you do to pass the GPU in to docker? Running ffmpeg -h encoder=hevc_videotoolbox on the host works just fine, odds are it would work fine as is except that docker needs access to the mac GPU which I don't believe is currently supported by docker

[NickM-27]

Trying to run with GPU capabilities in docker compose returns: Error response from daemon: could not select device driver "" with capabilities: [[gpu]]

[ireuben]

Yeah, I didn't do anything to expose the M1 GPU to Docker container, and from some googling it doesn't appear possible currently (looks like Docker has some NVDIA-specific stuff to expose NVIDA GPUs). Doh...

[digaus]

I am now running HA OS aarch64 as a VM in UTM with virtualized GPU. Currently no access to my Mac but yesterday I was able to start frigate with h264_v4l2m2m (no startup error) but did not test with rtsp stream yet or if performance is really improved :) will do it later today and report back

[bluSCALE4]

I am now running HA OS aarch64 as a VM in UTM with virtualized GPU. Currently no access to my Mac but yesterday I was able to start frigate with h264_v4l2m2m (no startup error) but did not test with rtsp stream yet or if performance is really improved :) will do it later today and report back

I have this setup but I can't seem to get my instance to run. I do have duckdns running and I wonder if that's having an affect.

[rmarinho]

is there a way to pass -c:v h264_videotoolbox on the M1 ?

[NickM-27]

No, docker on Mac doesn't support hardware mapping

[marcelklopper]

Has anyone managed to access VideoToolBox inside docker? I found two new leads. The Jellyfin community is using an SSH method to access the host ffmpeg. ( https://oliverbley.github.io/posts/2022-12-27-jellyfin-in-docker-hardware-acceleration-on-macos/ )

Then I also found this project that seems to be utilising VideoToolBox inside docker somehow. https://github.com/photoprism/photoprism

I am to much of a noob with these things to get it to work. but will be great if someone is willing to help.

[gniting]

apologies for reviving this old thread as I could not find relevant info anywhere else...

@jimshaggy @bmclapham1973 would you mind sharing your docker compose/run command and your frigate config's ffmpeg, detector and model sections? I've got frigate running on a M1 Studio, but it feels "sluggish." I am hoping its due to lack of M1 related optimization params, hence the request.

Here's what I have so far:

docker run command that I've used

docker run -d \
    --name frigate \
    --privileged \
    -e PLUS_API_KEY='my_api_key' \
    --restart unless-stopped \
    --shm-size=125m \
    -v /etc/localtime:/etc/localtime:ro \
    -v /Users/xxxx/Frigate:/config \
    -v /Users/xxxx/Frigate/nvr:/media/frigate \
    --mount type=tmpfs,target=/tmp/cache,tmpfs-size=1000000000 \
    -p 5000:5000 \
    -p 8554:8554 \
    -p 8555:8555/tcp \
    -p 8555:8555/udp \
    -e FRIGATE_RTSP_PASSWORD='xxxx' \
    ghcr.io/blakeblackshear/frigate:0.14.0-beta2

relevant sections from my config.yaml file

ffmpeg:
  global_args: -hide_banner -loglevel warning -threads 2
  hwaccel_args: auto
  input_args: preset-rtsp-generic
  retry_interval: 10

detectors:
  cpu1:
    type: cpu
    num_threads: 4
  cpu2:
    type: cpu
    num_threads: 4

model:
  path: /cpu_model.tflite
  labelmap_path: /labelmap.txt
  width: 320
  height: 320
  input_pixel_format: rgb
  input_tensor: nhwc

[marcelklopper]

Which image are you using? You need to use the arm64 image. I am using standard config and docker config also no hardware acceleration. I am not home at the moment, but will post my config later on

[gniting]

When I run uname -m inside the frigate docker container, I get the response as: aarch64, which I believe is the correct docker image for the M1 Studio Mac.

[newness0]

Running on M4 Mac mini with 2 CPU detectors. Inference speeds mostly stay under 7 and spike up when lots of activity. CPU user usage ranges between 20-40% overall (also has a Debian vm in utm running Homebridge)

1. Would be great to have the ability to use coral instead
2. Use GPU for inference/ffmpeg

[NickM-27]

Both of those require docker on macos to support it