On 9/5/2025 at 2:17 PM, KittenMittons said:

attached a copy of the log after restarting the container in unraid

frigate-logs.txt

12 hours ago, Professor Oak said:

Unable to get GPU detector working on v16.

I just upgrade to v16 and I can no longer get my NVIDIA GPU to work as a detector. I tried keeping the tensorrt settings, but that just throws an error stating it's no longer support and causes the docker container to endlessly restart. I also tried changing it to onnx, but that also causes it to endlessly loop, and I have no idea how to get any models to work with it. The only way for me to keep the container up and running is by using a CPU detector, so I know the rest of the config should be good.

GPU: Nvidia RTX 2060

Docker Image: ghcr.io/blakeblackshear/frigate:stable-tensorrt

YOLO_MODELS: yolov7-320,yolov7x-640

From what I understand after reading the documentation, the generation of YOLO models for TensorRT is now only available for Jetson Nano.

In your case, you have to switch to use ONNX models, as explained here.

The usage of the tensorRT version is still needed, but you have to manually download and place the model.

8 hours ago, yayitazale said:

From what I understand after reading the documentation, the generation of YOLO models for TensorRT is now only available for Jetson Nano.

In your case, you have to switch to use ONNX models, as explained here.

The usage of the tensorRT version is still needed, but you have to manually download and place the model.

That's what I was afraid of. I tried downloading/creating YOLO 3, 4, 7 and 9 models, but kept running into many issues. I'll see if I can take my time with it now since my deployment is currently stable.

When switching to ONNX, do I need to remove all the attributes that are listed in the unraid config that pertain to TensorRT?

2 hours ago, Professor Oak said:

That's what I was afraid of. I tried downloading/creating YOLO 3, 4, 7 and 9 models, but kept running into many issues. I'll see if I can take my time with it now since my deployment is currently stable.

When switching to ONNX, do I need to remove all the attributes that are listed in the unraid config that pertain to TensorRT?

Yes, they are already removed from the master template but will no make any difference to you until you reinstall the app, so the easiest way is just remove the entries.

18 hours ago, yayitazale said:

The usage of the tensorRT version is still needed, but you have to manually download and place the model.

I assume you're talking about this part?

Where am supposed to be running this commands? The console for the frigate container?

6 hours ago, KittenMittons said:

I assume you're talking about this part?

Where am supposed to be running this commands? The console for the frigate container?

No, in the host. Those are not yolo models -> https://docs.frigate.video/configuration/object_detectors/#downloading-yolo-models

Edited September 9, 2025Sep 9 by yayitazale

Definitely worth the swap to yolo9 from mobile coral edgetpu to intel igpu's
99% reduction in false detections. Also took time to use the visual debugger to reduce motion sensitivity saving cpu power.
Still catching everything as needed.

Object Detectors | Frigate

Frigate supports multiple different detectors that work on different types of hardware:

detectors:
	ov_0:
		type: openvino
		device: GPU

Edited September 10, 2025Sep 10 by dopeytree

On 9/8/2025 at 11:53 AM, yayitazale said:

Yes, they are already removed from the master template but will no make any difference to you until you reinstall the app, so the easiest way is just remove the entries.

I removed the entries from the unraid config, finally got the onnx file created and loading, switched to onnx for the dector, but it's still using CPU for detection.
From System Logs: "INFO: Created TensorFlow Lite XNNPACK delegate for CPU."

Unraid Docker Run Command:

Config YML settings:

^I tried the detector with, and without, the device and engine listed.

11 hours ago, Professor Oak said:

I removed the entries from the unraid config, finally got the onnx file created and loading, switched to onnx for the dector, but it's still using CPU for detection.
From System Logs: "INFO: Created TensorFlow Lite XNNPACK delegate for CPU."

Unraid Docker Run Command:

Config YML settings:

^I tried the detector with, and without, the device and engine listed.

I just set up yolov9 with OpenVino and I see also "INFO: Created TensorFlow Lite XNNPACK delegate for CPU." but the system is using the openvino GPU correctly.

Waht do you see in the /system#general

As you can see in the images below, in my case as I'm using openvino I see a OV as detector.

This are my stats with a T size 320 model:



With a T size 640 model:


With a M size 320 model:


With a M size 640 model:


If I switch back to Coral, I see coral as detector:

Edited September 10, 2025Sep 10 by yayitazale

Don't you need this instead of type:cuda

detectors:
  onnx_0:
    type: onnx
	device: gpu 

Edited September 10, 2025Sep 10 by dopeytree

Hi! I'm trying a fresh install with an rtx a4000 and the yolox_m.onnx but I'm getting this error:

onnxruntime.capi.onnxruntime_pybind11_state.InvalidArgument: [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Unexpected input data type. Actual: (tensor(uint8)) , expected: (tensor(float))

any ideas why?

thanks

9 hours ago, renk1 said:

Hi! I'm trying a fresh install with an rtx a4000 and the yolox_m.onnx but I'm getting this error:

onnxruntime.capi.onnxruntime_pybind11_state.InvalidArgument: [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Unexpected input data type. Actual: (tensor(uint8)) , expected: (tensor(float))

any ideas why?

thanks

You did not copy the required input_dtype config https://docs.frigate.video/configuration/object_detectors#yolo-v3-v4-v7-v9-1

detectors:

onnx:

type: onnx

gpu: true

input_tensor: nchw

input_dtype: float

num_threads: 2

model_path: /config/model_cache/yolox_m.onnx

Sorry I forgot to post my config, I did have it defined but still gives me the same error. Tried with float and float_denorm

11 minutes ago, renk1 said:

detectors:

onnx:

type: onnx

gpu: true

input_tensor: nchw

input_dtype: float

num_threads: 2

model_path: /config/model_cache/yolox_m.onnx

Sorry I forgot to post my config, I did have it defined but still gives me the same error. Tried with float and float_denorm

Width and height missing... Copy the proposed configuration

unraid 7.1.4

frigate 0.16.0-rc4

Hi, for many hours I've been struggling unsuccessfully to get face recognition to work in an efficient configuration. I've tried almost everything and would like to ask for help.

Goal: An efficient configuration: person detection on a 640x480 stream, face recognition on a 4K snapshot.

Tested Configurations and Results:

1. Two streams (directly from the camera):

   • Result: Detects person, but never attempts face recognition. The process of taking a 4K snapshot silently fails.

2. One stream (4K only) for detection and recording:

   • Result: Face recognition WORKS CORRECTLY.

   • Problem: Unacceptably high power consumption (~29W).

3. Two streams (via the built-in go2rtc):

   • Result: After rebooting the camera (to clear its IP block), the connection was established correctly.

   • Problem: Despite a working connection, the behavior is identical to Configuration 1 (no face recognition attempt).

4. One 4K stream with downscaling to 640x480 in the detect section:

   • Result: Low power consumption, person is detected.

   • Problem: Behavior identical to Configurations 1 and 3 (no face recognition attempt).

Probable Diagnosis (based on facts):

The problem likely lies within Frigate's internal mechanism. When the detection (detect) source and the snapshot (record) source are not the exact same stream (as in Configurations 1, 3, and 4), the process of fetching the 4K snapshot fails in a way that is not logged. This is likely a timing issue or a bug in switching context between streams.

The Only Working Solution (A Compromise):

Using a single 4K stream for everything (detect and record) with a drastically lowered fps in the detect section (e.g., to 2 FPS) to reduce power consumption. This bypasses the Frigate bug and is the only confirmed working method in this case.

Does anyone have any suggestions on how to solve this, to make it work with go2rtc and two separate streams while maintaining low power consumption?

25 minutes ago, Xwint5 said:

unraid 7.1.4

frigate 0.16.0-rc4

Hi, for many hours I've been struggling unsuccessfully to get face recognition to work in an efficient configuration. I've tried almost everything and would like to ask for help.

Goal: An efficient configuration: person detection on a 640x480 stream, face recognition on a 4K snapshot.

Tested Configurations and Results:

1. Two streams (directly from the camera):

   • Result: Detects person, but never attempts face recognition. The process of taking a 4K snapshot silently fails.

2. One stream (4K only) for detection and recording:

   • Result: Face recognition WORKS CORRECTLY.

   • Problem: Unacceptably high power consumption (~29W).

3. Two streams (via the built-in go2rtc):

   • Result: After rebooting the camera (to clear its IP block), the connection was established correctly.

   • Problem: Despite a working connection, the behavior is identical to Configuration 1 (no face recognition attempt).

4. One 4K stream with downscaling to 640x480 in the detect section:

   • Result: Low power consumption, person is detected.

   • Problem: Behavior identical to Configurations 1 and 3 (no face recognition attempt).

Probable Diagnosis (based on facts):

The problem likely lies within Frigate's internal mechanism. When the detection (detect) source and the snapshot (record) source are not the exact same stream (as in Configurations 1, 3, and 4), the process of fetching the 4K snapshot fails in a way that is not logged. This is likely a timing issue or a bug in switching context between streams.

The Only Working Solution (A Compromise):

Using a single 4K stream for everything (detect and record) with a drastically lowered fps in the detect section (e.g., to 2 FPS) to reduce power consumption. This bypasses the Frigate bug and is the only confirmed working method in this case.

Does anyone have any suggestions on how to solve this, to make it work with go2rtc and two separate streams while maintaining low power consumption?

Face Recognition | Frigate

Face recognition identifies known individuals by matching detected faces with previously learned facial data. When a known person is recognized, their name will be added as a sub_label. This informati

38 minutes ago, yayitazale said:

Face Recognition | Frigate

Face recognition identifies known individuals by matching detected faces with previously learned facial data. When a known person is recognized, their name will be added as a sub_label. This informati

It works for me too, it recognizes the faces of saved people, but only with the option I described, and that's the problem.

One stream (4K only) for detection and recording:

• Result: Face recognition WORKS CORRECTLY.

• Problem: Unacceptably high power consumption (~29W).

I just figure face detection is a growing extra.
It takes a while to build up a database.
I'm running the large model on an intel igpu works well.

5x streams 640x480 or similar sized.
High quality stream is 3k.
No extra power usage.

16 minutes ago, Xwint5 said:

It works for me too, it recognizes the faces of saved people, but only with the option I described, and that's the problem.

One stream (4K only) for detection and recording:

• Result: Face recognition WORKS CORRECTLY.

• Problem: Unacceptably high power consumption (~29W).

Of course 4K requires much more power to be processed...

7 minutes ago, yayitazale said:

Of course 4K requires much more power to be processed...

Unfortunately, with 4k and face recognition, the entire server consumes ~29W, but if go2rtc or a similar alternative worked, it would consume ~24W. Because then it does detection at 640x480, but when it detects a person, it does facial recognition from a 4k photo, but that's what I can't get to work. For me, detection is now running at 4k 2fps all the time with ~26,5W (server). As I described above, I tried various options.

Currently running on

model:
  path: /models/efficientdet_lite1_384_ptq_edgetpu.tflite
  labelmap_path: /models/coco_labels.txt
  width: 384
  height: 384

face_recognition:
  enabled: true
  model_size: large

In the past, I used OpenVINO and iGPU with i5-12500. I bought Coral for testing, but it didn't really give me anything except taking up an m2 slot. Perhaps with more cameras, it will be useful in terms of lower W. I am now using it on

detectors:
  coral:
    type: edgetpu
    device: pci

model:
  path: /models/efficientdet_lite1_384_ptq_edgetpu.tflite
  labelmap_path: /models/coco_labels.txt
  width: 384
  height: 384

When it comes to decoding, what works best for me is

ffmpeg:
  hwaccel_args: preset-vaapi
  input_args:
    - -rtsp_transport
    - tcp

It is better optimized than quick sync.

Edited October 3, 2025Oct 3 by Xwint5

1. what has 4k got to do with it? the model size is 384px

2. what size is your substream on the camera?

3. log into camera webgui or use vlc media info on stream url

4. the 4k video should just being recorded

5. detection should be on the substream

6. which is already a smaller resolution read for the models to work efficiently

Edited October 3, 2025Oct 3 by dopeytree

dopeytree

My main camera stream is 4k 20FPS (I reduced the fps due to W), the sub-stream is currently 640x480, but of course I can change it to 720p. Currently recording 4k and 20FPS video. Face recognition in Frigate runs at 4k 2FPS on the CPU. The 384px model only recognizes people and cars (in my case) and has nothing to do with face recognition, because that's always done by the CPU. Unless you're asking why I wrote about the model I use? It's about power consumption, it's a pretty good model, better than ssd_mobilenet_v2.

For me, face recognition has to be at 4k, because I have ~12m to cover to the gate from a height of 5m. 720p sometimes catches half of it, completely fails, and 640x480 is a total disaster.

Edited October 3, 2025Oct 3 by Xwint5

You likely need a zoom lens (either fixed telephoto or PTZ type or just take apart & change lens) on another camera rather than resolution to help solve Face Recognition from a distance

Thanks for the input. For my use case, a 4mm wide angle lens is optimal. A 6mm lens would make it easier to cover the gate, but I’d lose the garage view on the right. I don’t want another camera because I want both wide and long face capture. Nowadays 4K resolution is quite standard.

In my opinion, the problem lies in Frigate itself. In Frigate 0.16.0-rc4, snapshots for face recognition are always taken from the first input (detect). In practice, this means that with dual-stream (small stream for detection + 4K for recording), only the person is detected, but the face recognition mechanism does not receive the image from 4K. In my tests, it only worked on a single 4K stream.

You may well be correct but alas we can't control frigates coding updates - where-as changing the cameras is possible today.

Some options:

1. FCD600 - 6MP UHD Dual-Lens Outdoor PoE 180° Panoramic View https://www.annke.com/products/fcd600

2. CZ804 - 4K 4X Optical Zoom PoE 12 MM Motorized Varifocal Lens https://www.annke.com/products/cz804

3. NightChroma^TM NCT425 - 2-in-1 Dual Lens PTZ 25X Optical Zoom https://www.annke.com/products/nct425

Reaching out for help here as I've done research elsewhere and tried optimizing a handful of different things both in hardware and software but still having the same issue with the coral TPU going offline making Frigate stop working (even though the container doesn't crash). This has started happening more frequently recently, on the order of multiple times a week, escalating today where it happened post reboot within 30 minutes. How can i correctly troubleshoot this to figure out what to do to fix it?

System info:

• intel i7-9700K processor (w/ quicksync)

• ASRock Z390 Pro4 ATX LGA1151 Motherboard

  • 2x M.2 key M slot + 1x M.2 Key E slot

• 32 GB RAM (4x8GB)

• Google Coral TPU Accelerator, M.2 (A/E Key)

• 1 HDD for frigate storage in pool (no parity)

  • some dashcam footage is also synced to this drive

• 2 nvme for zfs cache (mirrored)

Cameras:

• 7x 1080p wyze v3 (thingino fw)

• 2x 2K tapo c120

• 1x 2K reolink wifi doorbell

• 1x 4K reolink POE

Here are some of the steps i've taken thus far:

• confirmed all streams are using sub stream for detect, and main stream only for record/audio

• moved 1 nvme ssd from M.2 key M slot to PCIe slot via M.2 to PCIe adapter

  • previously when coral tpu offline issue occurred, reboot would cause second nvme ssd to become unavailable - only solution was full shutdown. Making this change mitigated this, allowing normal reboot to bring coral tpu back online.

• optimized Frigate container Extra Parameters:

  • --shm-size=4G --mount type=tmpfs,target=/tmp/cache,tmpfs-size=1000000000 --restart unless-stopped --log-opt max-size=10m --log-opt max-file=3 --memory=16G --memory-swap=16G

• optimized Extra Parameters on one other high RAM use container:

  • --memory=12G --memory-swap=12G

• installed Netdata container

  • not sure what to look for yet

EDIT: given all i've read about the issues with coral A+E key TPUs causing issues on desktop motherboards' m.2 slot meant for wifi chips, I ended up replacing my coral M.2 A+E key TPU with a Hailo-8L M.2 B+M key TPU. It's only been a few weeks but haven't had any issues since.

Edited October 17, 2025Oct 17 by john_smith
update hw