Frigate 0.17: Looking for feedback from users on dev builds

[NickM-27]

Frigate 0.17 includes many new features, some of which require additional feedback to ensure a smooth user experience. Frigate 0.17 is still under active development and is not recommended for use in production systems (there may be large / breaking bugs).

As we get closer to a beta cycle we are looking to get additional feedback from users in some specific areas so that we may get a better understanding of how things are working and fix any bugs.

We are currently looking for feedback on everything in the release, but specifically looking for feedback on the below items.
If you experience any bugs with config migration, or bugs with existing functionality then please give that feedback as well.
Please leave all feedback in this discussion, do not create separate bug reports or support discussions.

The docs are: https://deploy-preview-19787--frigate-docs.netlify.app/

Object Detectors

• RockChip NPU: RockChip now supports automatic model conversion which means Frigate+ and enrichments models are automatically converted to RKNN format and run on the NPU. Looking for general feedback such as inference speed and any issues.
• Apple Silicon: Apple silicon is now supported. This requires setting up a plugin on the host so Frigate can communicate with the GPU/NPU. see the docs

Classification Models

Frigate now has end-to-end support, including training, for classification models. These come in two types, state classification (ex: gate open / close) and object classification (ex: dog recognition). You can use the UI to create and start training the model.

We are specifically looking to see examples of accuracy (good and bad) with images, any feedback on the setup wizard, bugs, etc.

https://deploy-preview-19787--frigate-docs.netlify.app/category/custom-classification

Triggers

Triggers utilize Semantic Search to automate actions when a tracked object matches a specified image or description. Triggers can be configured so that Frigate executes a specific actions when a tracked object's image or description matches a predefined image or text, based on a similarity threshold. Triggers are managed per camera and can be configured via the Frigate UI in the Settings page under the Triggers tab.

https://deploy-preview-19787--frigate-docs.netlify.app/configuration/semantic_search#triggers

Review Item Descriptions

Frigate now supports using GenAI to summarize review items. Unlike object descriptions which are searchable, review item descriptions are structured metadata that include suspicious activity categorization. This allows Frigate UI and notifications to show the title / description along with indicators that something may require review based on the GenAI's understanding of the activity.

https://deploy-preview-19787--frigate-docs.netlify.app/configuration/genai/genai_review/

UI Changes

Frigate 0.17 has many UI changes and additional features, some that would be good to get additional feedback on for layout or behavior bugs:

• Review Detail Stream
• Explore Dialog
• Camera Creation Wizard
• A new HLS player is used in some cases, let us know if there are any recording playback issues

[hawkeye217]

As we've mentioned in the past when asking for feedback, please don't open bug reports or support requests for development builds. At this time, what would be most helpful to us would be to keep all of your feedback here on this post. Thanks for helping test the new features!

[deanpm70]

Minor UI bug in build "[fbf4388-tensorrt]"

From Review screen, select an object type. Click on an item to review. Click on "Yes" in response to "Is this object a []?" and "Submitted" appears. Now click on Tracking Details tab then to the Snapshot tab. The Submit to Frigate+ yes/no options reappear. The behaviour seems consistent across all object types.

[NickM-27]

Thanks, looks like the submit to frigate+ image dialog is missing the buttons too

[hawkeye217]

This will be fixed in the next build.

[xconverge]

My questions still stand related to the tracked object panel when I asked in #20748 (comment) This screen is pretty awesome, but there is still some usability I am wondering about

I can press pause, and then click the selector dots along the line-of-ants, but the bounding box is not shown at each timepoint, only the first and last detection? Let me know if you want a more well-written explanation, I am being pretty vague I think

[NickM-27]

We don't have the bounding box for every point. And submitting to Frigate+ can be done using the button on the video overlay

[xconverge]

I see! I really thought we did in 0.16, I guess I misremembered, all good then! Thanks for the update

[hawkeye217]

The old Object Lifecycle pane made it less obvious that bounding boxes were not available for every point on the path line, but @NickM-27 is right, we have only ever saved bounding boxes for the significant ones, which are now shown as the "line-of-ants" (great analogy!).

I'll also add that submitting to Frigate+ can also take place directly from the 3-dots menu next to each significant moment in the Tracking Details pane.

[raders]

The Apple Silicon detector doesn't reconnect if the client process is stopped and restarted. Currently you have to start the client first or the detector won't work. It would be nice if the connection was more resilient since there is no indication in frigate that the client is not working.

[NickM-27]

I think we would need to see logs, it definitely has the logic to reconnect

[gigabyte-1000]

213a1fb
I ran into a strange bug today with one of my cameras. When I viewed the recording history, I could scrub through the timeline normally and the preview images loaded just fine. But the moment I stopped at any point and tried to play the actual video, it only showed a loading spinner. This happened anywhere between about 5 AM and 8 PM, and the issue was the same across multiple browsers and devices.

Recording itself was definitely working, since the review snapshots and timeline previews were generated correctly, Frigate just refused to play the video segments from that timeframe.

I restarted the Docker container, and everything recorded after the restart works perfectly (scrubbing, playback, reviews, etc.). However, the older recordings from before the restart still won’t play at all.

There’s no reliable way to reproduce it right now, and it only affected this single camera. No errors showed up in the logs for that camera either, so it may have just been a one-off glitch.

[NickM-27]

Recording itself was definitely working, since the review snapshots and timeline previews were generated correctly

neither of those things are related to recording, so that does not mean recordings were working.

[ZhaiSoul]

Sometimes the generative AI doesn't work in the language I specified, and the output ends up being in English.

So, how should I properly configure the activity_context_prompt?When I use the auto-completed default configuration in the editor, it causes an error:

Currently, I'm using the following configuration— is this correct?

review:
  genai:
    alerts: true
    activity_context_prompt: "### 正常活动指标（0级） - 任何时段出现在任意区域内的已知/已验证人员 - 居民区内携带宠物的人员
      - 白天/晚间（早6点至晚10点）的配送或服务行为：如将包裹送至门口/门廊、放置物品后离开 - 白天可见携带工具、身着制服或驾驶服务车辆的服务/维修人员
      - 活动始终仅限于公共区域（人行道、街道、花园外），未进入私人领地 ### 可疑活动指标（1级） - 测试或试图打开门窗/车辆/建筑物的把手​ —— 无论时间或持续时间，始终列为1级​
      - 深夜/凌晨（晚11点至早5点）出现在私人区域（车道、车辆/建筑物附近）的未识别人员​ —— 无论行为或持续时间，始终列为1级​ - 拿取不属于自己的物品（包裹、门廊/车道上的物件）
      - 攀爬或翻越围栏/障碍物进入私人领地 - 试图隐藏行为或物品 - 长时间徘徊：在大部分监控时段内停留于同一区域且无明确目的 ### 重大威胁指标（2级）
      - 持有破门工具（撬棍、拔钉器、钢丝钳） - 可见的武器（枪支、刀具、攻击性使用的球棒） - 正在进行的强行闯入 - 肢体攻击或暴力行为 - 持续进行的财产破坏或盗窃行为"
    preferred_language: 简体中文

[NickM-27]

You need to use proper yaml format for multi line values

Regarding the wrong language, what model are you using? Perhaps the model isn't following instructions correctly. Otherwise, I'll do some testing and see if I can reproduce.

[InstigatorX]

I have a classification setup for dogs and get the following when training. Happens on both of the classifications (dog and car) I have setup.

2025-11-20 15:07:53.566967306  [2025-11-20 15:07:53] absl                           INFO    : Function `function` contains input name(s) resource with unsupported characters which will be renamed to sequential_1_dense_1_2_biasadd_readvariableop_resource in the SavedModel.
2025-11-20 15:07:53.566967306  [LOGGING] Last message repeated 1 times
2025-11-20 15:07:55.792675582  [2025-11-20 15:07:55] tensorflow                     INFO    : Assets written to: /tmp/tmp2psjh3iy/assets
2025-11-20 15:07:56.629149846  WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
2025-11-20 15:07:56.629152841  W0000 00:00:1763669276.629089 3590852 tf_tfl_flatbuffer_helpers.cc:365] Ignored output_format.
2025-11-20 15:07:56.629153994  W0000 00:00:1763669276.629107 3590852 tf_tfl_flatbuffer_helpers.cc:368] Ignored drop_control_dependency.
2025-11-20 15:07:56.717614250  I0000 00:00:1763669276.717500 3590852 mlir_graph_optimization_pass.cc:425] MLIR V1 optimization pass is not enabled
2025-11-20 15:07:59.434866373  fully_quantize: 0, inference_type: 6, input_inference_type: UINT8, output_inference_type: UINT8
2025-11-20 15:07:59.748432666  [2025-11-20 15:07:59] frigate.util.classification    INFO    : Wrote training metadata for dog: 48 images
2025-11-20 15:07:59.748567596  [2025-11-20 15:07:59] frigate.util.classification    INFO    : Finished training dog

[NickM-27]

Is there something specific you are pointing to? I don't see any issues here. We unfortunately don't have a good way to block these hard coded prints that tensorflow has

[InstigatorX]

No idea if this is just cosmetic or not. Sounds like it is. Figured I'd err on the side of caution and let you know.

[gu3stZA]

Testing roles:

1. I've assigned a user to a role that can only see 1 camera. The user is still able to see all camera groups. If they click on a camera group that contains no cameras that they have access to, they see a screen that says "no cameras configured. Click to add a camera". Clicking this button just takes them to settings, so they can't actually add a camera. I don't believe these users should see the option to add a camera at all (but regardless, the button isn't functional anyway).

2. Follow same steps as above, log in with a restricted user who has access to only one camera. In Safari Mobile (iOS 26.1), if you now click on the group icon where you have no access to cameras, as above, you get the add camera screen. However, the menu bar at the top of the screen (where you can change to other groups) disappears. So you are now stuck on this screen and can't view any cameras. Even logging out and in doesn't fix the issue - so this user is now stuck viewing no cameras. This only happens in Safari on iOS, it does not happen on desktop (Safari MacOS).

3. When logging in with a user who has restricted cameras on Safari Mobile (iOS 26.1), the first page I see says "Access Denied. You don't have permission to view this page". I can then click on cameras to view cameras then it shows fine. Not sure what the first page is directing me to, as once I leave the screen, I can't navigate back to any screen that shows Access Denied again. Again, only an issue on mobile - does not occur on desktop.

[hawkeye217]

Thanks for testing. I've pushed some changes and fixes for the next build.

[gu3stZA]

Thanks! This is one of my most awaited features so excited to get it working

[bracklanna]

So more-refined Roles are coming in 0.17? I don't see that listed in the features described at the top of this post, but it sounds like they are.

I appreciate & respect all the hard working going into the "fancy stuff" classification/AI/etc. features, but actually I've been waiting for improved Roles before setting up a new site location w/ Frigate. Really need the ability to use administrative policies to group and restrict camera views. (E.e. the residents of Unit A should only be able to view cameras 1 and 2, while residents of Unit B can only see cameras 3 & 4.)

I've looked into / read about using nginx reverse proxy as a work-around for this but if it's coming as a native feature that would be preferable.

[NickM-27]

yes, the ability to define roles and which cameras are available to each role is available in 0.17. The above list is just a list of things we want extra testing for. It is nowhere near a changelog

[gu3stZA]

Minor issue, but when using NPU has one of my detectors, the size of the NPU frame in the statistics page is slightly off, and looks weird. Not sure if this is intentional or not.

[NickM-27]

Thanks, will be fixed in the next build

[Korenchkin]

in 213a1fb (2 days ago) seems to be missing motion debug (that window with ~8 switches what to show),or i can't find it

[NickM-27]

Are you referring to motion tuner? We haven't made any changes to the menus

[hawkeye217]

Do you mean Debug view? That is no longer accessible from Settings, but can be viewed from single camera Live view using the settings cog icon and enabling the Debug switch. You can also long press/right click on a camera in the dashboard and choose Debug View from the context menu.

This changed with the builds all the way back in early October.

[Korenchkin]

this one

[Korenchkin]

ah,you were faster,thanks! i was fighting downgrade to take a picture :)

[gu3stZA]

In the add camera wizard, the little help "tooltips" (clicking on the question marks that give you extra info) just flashes the tooltip so fast you can't actually read it.

Using Safari on MacOS 26

[hawkeye217]

Ah yes, we've seen this on Safari before. I've pushed fixes for that as well.

[gu3stZA]

Some feedback on inference times.

I run an Intel Ultra 7 265K. Testing the NPU, I get inference times of between 8-10ms, and very low CPU usage (less than 1%) per detector. This compares well to the iGPU which has a similar inference time, but higher CPU usage (around 1-2% with low motion, but when there is a lot of motion this goes up to 8%).

Overall, it seems the NPU has similar inference times, but it seems to be more scalable. If I increase the number of GPU detectors, my inference times slow down significantly. I have added up to 4 detectors and then see inference times around 25-30ms each.

With the NPU, I have added 4 detectors and inference times remain at 8-10ms (possibly due to the lower CPU usage? Not sure what the mechanisms are here, to be honest). I will continue testing to see how the inference times hold up when the wind picks up tomorrow again and movement ramps up significantly.

One big caveat is that movement seems to be a big factor in using CPU and therefore the speed of inference. So some of the variation I am seeing may be due to different levels of wind/movement.

But overall, it seems to me that this CPU is great for quite large installs with lots of movements. I have tried up to 4 iGPU and 4 NPU detectors together (for a total of 8), with good inference times (around 10-20ms), so this should be good enough for quite a lot of cameras with a lot of movement (400-800 detections per second by my rough maths).

[kirsch33]

any ideas why i cant see the tracking paths when im viewing the details page on iOS 26.1? works fine on laptop.

[hawkeye217]

The object path is only valid for the active item in the list. Tap the icon/label or use the dots menu and tap "Show object path".

[kirsch33]

ahh- i didnt know you had to specifically click "Show Object Path"

its working fine now.

that said, heres some feedback:

1. testing on my phone seems to indicate that the "Show Object Paths" setting is persistent between opening/closing the PWA but only once enabled on a specific item. it doesnt carryover the setting between review items. I'd like to see the persistence expanded OR a config setting to have it always enabled
2. track paths are by far my favorite new Frigate feature in a long time--but--i hate that it also enables the zone overlays too. the track paths are very discrete and dont pollute the screen but the zone display does and IMO it doesnt look good. just my personal opinion, id like a config item maybe to show paths only? or if you guys dont want to add more config items maybe consider increasing transparency on the zones or tweaking them to make them not so overpowering in the view?

[hawkeye217]

Great, glad it's working now!

Object paths were implemented primarily for debugging in Review, and especially in Explore’s new Tracking Details view.

In Review, selecting “Show Object Path” seeks the video to the start of the selected object and displays its path - including any zones entered - until the end of that specific review item. The path is only “persistent” while the footage within that review item is being shown in the video player.

Because a single review item can contain multiple tracked objects that overlap in time and space, displaying more than one path at a time would quickly become confusing. For example, in a scene with many people walking through the frame, you’d likely end up with a mess of overlapping “spaghetti lines,” which isn’t helpful for debugging. In addition, an object can begin in one review item, remain stationary, and then appear again in a later item. Displaying paths across multiple items would make it increasingly difficult to understand why certain paths appeared or disappeared.

While it can certainly be fun and visually interesting to watch object paths animate over the footage, the primary goal of this feature is debugging. That’s why it’s designed the way it is today.

That said, we’re absolutely open to feedback. If there are important debugging use cases that the current behavior doesn’t support well, we’d love to hear them.

[Williamh695]

I changed my media location /mnt/storage/frigate:/media/frigate and its saving there but its not reflecting that on the storage page, its still showing the old location. Also Camera Storage shows old location and 0% on every camera unless I do a manual recording then it shows just that.

Ive had the new one (System 0.17.0-213a1fb) running for almost 24 hours now and the general states page is VERY slow to load or do anything. All other pages for everything have not done this. Quick restart fixes it. The memory bug seems to be fixed if not its filling up very slowly now.

I did the auto detect for my Amcrest cameras and it worked great! Would be nice if it asked to limit the FPS of the cameras on the setup. I see the options for detect and record. Maybe after they are setup have an advance config page with more options. Makes an easier setup for the less techy types.

After adding some objects to classification and it finds them, I can right click to select multiple on the first page/batchs it finds but my only options are to delete. I wanted to mark entire folders for training but I had to click into it and do them 1 by 1. Even after clicking in being able to select more then 1 at a time would be great. I just added me, dogs, chickens after a day of recording so we will see how it goes.

For states I have all outdoor cameras so I have it set to see when my gate opens and closes. I wanted to set one for my dog bed when a dog was in it or not but it would never find it. In fact nothing ever came up. I just tried to do it again. I get No sample images generated. No matter how big I make the square. Also does it have to be that shape? Maybe allow rectangle? idk if that's possible.

I added some triggers. Lets see if they work tomorrow.

Can we move the "Mark these items as reviewed" to the top? I have cameras facing the road and I get 100's if not 1k a day. I know I know I don't have mine setup right, but I like it this way. It works for me for reasons no one on the internet cares about.

Thats all I have for now. Thanks a lot for everything!

[NickM-27]

I changed my media location /mnt/storage/frigate:/media/frigate and its saving there but its not reflecting that on the storage page, its still showing the old location. Also Camera Storage shows old location and 0% on every camera unless I do a manual recording then it shows just that.

Did you change the DB? this is usually because the same DB is still being used

I did the auto detect for my Amcrest cameras and it worked great! Would be nice if it asked to limit the FPS of the cameras on the setup. I see the options for detect and record. Maybe after they are setup have an advance config page with more options. Makes an easier setup for the less techy types.

Not sure what you mean by limit fps. Recording always copies the cameras fps, and detect is automatically set to 5 fps which is the recommended for all setups

I wanted to mark entire folders for training but I had to click into it and do them 1 by 1. Even after clicking in being able to select more then 1 at a time would be great. I just added me, dogs, chickens after a day of recording so we will see how it goes.

similar to face it is not recommended to train large batches at a time, you usually just want to train a handful of different images which scored low, if you include a lot of already correct classifications that can do more harm than good.

Also does it have to be that shape? Maybe allow rectangle? idk if that's possible.

yes, it has to be a square

[gu3stZA]

I'm trying to track down an issue I'm seeing whereby if I change the motion threshold to 5 (I was trying to stress test the detector by ramping up the movement detection), then cameras stop sending frames to the detector. I tried to enable the debugger logging for the detection, but not seeing anything in the logs. Is this a known issue with the logger? I am running 0.17, hence me posting here, but I'm not sure if the issue is present in 0.16 too.

The logging part of my config is like this:
logger:
default: info
logs:
frigate.object_detection: debug

With logger settings like that, there is no debug level logging happening, despite the camera stats page showing a large number of frames being sent to the detector.

Should I be logging something else?

EDIT: To be clear, this screenshot is after I set the motion threshold back up to 30 and I get detections on all cameras again

[hawkeye217]

That logging name is invalid, there is no frigate.object_detection module.

When you turn down the motion threshold, you might be causing the lightning_threshold to trigger as more than a certain percentage of the frame is detected as motion. This is expected behavior.

[gu3stZA]

Ah thanks. I got it from the https://docs.frigate.video/configuration/advanced page in the documentation. So that page is not accurate?

I will play around with the lightning_threshold number just to test if that was really the case, thanks!

[hawkeye217]

Some modules have changed for 0.17, and those docs are for stable. Nonetheless, the dev docs still have that key so I'll make sure that's updated.

[gu3stZA]

In the classification state models, I have set it to do inference every 1 second (extreme, I know, but just trying to test). However, I don't see any inferences - either in the metrics tab, or in the recent classification list (I deleted all the old classifications to clean it up).

Am I misconfiguring something? Or is the metrics tab and the recent classifications list not a reliable way to determine if classification is actually happening or not?

[NickM-27]

I'll take a look and see

[deanpm70]

Super weird edge case for classification

tldr; I've moved my camera and now the suggested images when building a new classification are mostly from the cameras old position.

I created a classification a few days ago. I then tweaked the camera angle today to better see the subject and deleted the classification with the intention of re-creating it with better data. Unfortunately, the images I've been presented with are mostly from the "old" camera position and are therefore useless. I'm guessing if I try again tomorrow, there will be fewer images from the old position...

Question
Does the process limit how far back in time it will select reference images from?

Suggestion
Allow the user to specify the maximum age of the stream from which reference images are taken.

[NickM-27]

You'd be better off just leaving the model with no image and train from there

The auto selection is meant to be fit convenience to get started quicker not a fully inclusive solution

[mike2307]

Hi,

I just tried the latest dev build (1b57fb1-rocm) on a system with an AMD Ryzen 9 6900HX and AMD Radeon 680M iGPU.
This has the update to ROCm 7.1 inside.
Unfortunately that does not work any more.

My previously used dev build (224cbdc-rocm) on that system worked without any issues.
I have set HSA_OVERRIDE_GFX_VERSION: "10.3.0" in case this is of importance.

I'm seeing the following logs:

2025-11-27 20:47:23.080888352  MIOpen(HIP): Error [Compile] 'hiprtcCompileProgram(prog.get(), c_options.size(), c_options.data())' static_kernel_gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.cpp: HIPRTC_ERROR_COMPILATION (6)
2025-11-27 20:47:23.081205424  MIOpen(HIP): Error [BuildHip] HIPRTC status = HIPRTC_ERROR_COMPILATION (6), source file: static_kernel_gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.cpp
2025-11-27 20:47:23.081208130  MIOpen(HIP): Warning [BuildHip] In file included from /tmp/comgr-2cf73d/input/static_kernel_gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.cpp:1:
2025-11-27 20:47:23.081213330  In file included from /tmp/comgr-2cf73d/include/static_kernel_common_header.hpp:9:
2025-11-27 20:47:23.081242336  In file included from /tmp/comgr-2cf73d/include/static_kernel_ck_utils_type.hpp:5:
2025-11-27 20:47:23.081244620  /tmp/comgr-2cf73d/include/miopen_type_traits.hpp:151:10: fatal error: 'type_traits' file not found
2025-11-27 20:47:23.081245973    151 | #include <type_traits>
2025-11-27 20:47:23.081246995        |          ^~~~~~~~~~~~~
2025-11-27 20:47:23.081248327  1 error generated when compiling for gfx1030.
2025-11-27 20:47:23.083314218  MIOpen Error: /longer_pathname_so_that_rpms_can_support_packaging_the_debug_info_for_all_os_profiles/src/rocm-libraries/projects/miopen/src/hipoc/hipoc_program.cpp:299: Code object build failed. Source: static_kernel_gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.cpp
2025-11-27 20:47:23.083534504  migraphx_program_compile: Error: /longer_pathname_so_that_rpms_can_support_packaging_the_debug_info_for_all_os_profiles/src/AMDMIGraphX/src/targets/gpu/include/migraphx/gpu/miopen.hpp:114: find_solution: MIOpen: miopenFindSolutions failed
2025-11-27 20:47:23.089148629  2025-11-27 20:47:23.089101990 [E:onnxruntime:, inference_session.cc:2544 operator()] Exception during initialization: Failed to call function
2025-11-27 20:47:23.093719198  Process frigate.detector:onnx:
2025-11-27 20:47:23.093723366  Traceback (most recent call last):
2025-11-27 20:47:23.093725580    File "/usr/lib/python3.11/multiprocessing/process.py", line 314, in _bootstrap
2025-11-27 20:47:23.093726743      self.run()
2025-11-27 20:47:23.093728416    File "/opt/frigate/frigate/object_detection/base.py", line 143, in run
2025-11-27 20:47:23.093730239      object_detector = LocalObjectDetector(detector_config=self.detector_config)
2025-11-27 20:47:23.093732043                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-27 20:47:23.093733836    File "/opt/frigate/frigate/object_detection/base.py", line 62, in __init__
2025-11-27 20:47:23.093735399      self.detect_api = create_detector(detector_config)
2025-11-27 20:47:23.093736902                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-27 20:47:23.093738696    File "/opt/frigate/frigate/detectors/__init__.py", line 18, in create_detector
2025-11-27 20:47:23.093784544      return api(detector_config)
2025-11-27 20:47:23.093786478             ^^^^^^^^^^^^^^^^^^^^
2025-11-27 20:47:23.093788883    File "/opt/frigate/frigate/detectors/plugins/onnx.py", line 39, in __init__
2025-11-27 20:47:23.093815895      self.runner = get_optimized_runner(
2025-11-27 20:47:23.093817869                    ^^^^^^^^^^^^^^^^^^^^^
2025-11-27 20:47:23.093820514    File "/opt/frigate/frigate/detectors/detection_runners.py", line 570, in get_optimized_runner
2025-11-27 20:47:23.093822307      ort.InferenceSession(
2025-11-27 20:47:23.093825313    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 485, in __init__
2025-11-27 20:47:23.093827868      self._create_inference_session(providers, provider_options, disabled_optimizers)
2025-11-27 20:47:23.093831154    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 584, in _create_inference_session
2025-11-27 20:47:23.093854920      sess.initialize_session(providers, provider_options, disabled_optimizers)
2025-11-27 20:47:23.093856924  onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Exception during initialization: Failed to call function
2025-11-27 20:47:29.095342089  [2025-11-27 20:47:29] frigate.watchdog               INFO    : Detection appears to have stopped. Exiting Frigate...

[mike2307]

Am I the only one with this issue?

[NickM-27]

We aren't aware of many users running ROCm. AMD released 7.1.1 so we'll likely update to that and see if this is fixed.

[gu3stZA]

I updated to the latest build (1b57fb1) to fix the issue with viewer and custom roles, but issue still persists. To clarify, the issue is that when logging in with a custom role or a viewer role user, no cameras display in the view where all cameras or grouped cameras should show. You see a black box and spinning loading icon. When clicking on the camera (or where the camera should be), you can live view the camera fine - so it's the grouped views that don't work.

[hawkeye217]

Thanks, that helps, I'll look into it further.

[hawkeye217]

This should be improved in #21094

[gu3stZA]

@hawkeye217 definitely improved in the latest build - thanks. Viewing cameras works as expected as far as I can tell. I have an issue on Safari though... logging in gives ACCESS DENIED screen. Navigating off this screen returns everything to normal, and it is not possible (as far as I can tell) to navigate back to this access denied screen.

It only seems to happen in Safari. I can reproduce it reliably on my Mac with Safari, but switching to Chrome I can't reproduce.

Steps to reproduce:

• Be in Safari (iOS 26 or MacOS 26)
• Set up two camera groups, A and B
• Have two cameras, camera 1 in group A and camera 2 in group B
• GROUP A IS LISTED FIRST IN THE CONFIG AND GROUP B IS LISTED SECOND (the ORDER parameter is irrelevant, only the order of the group in the config file
• create a user that only has access to camera 2
• log in with user, and you will see the below screen:

Not sure if the network tab is useful or not, but here it is:

I played around with various settings, and the weird thing is:

• it doesn't matter what the "ORDER" parameter is for the camera groups. It only matters which is listed first in the config. If your user has no access to any cameras in the first group, logging in will give this error.
• only an issue in Safari - not reproducable in Chrome (on the same Macbook)

[hawkeye217]

I'll see if I can reproduce that. I bet it has something to do with the last camera group you have selected before logging out and then logging back in with another user. Frigate saves the last selected camera group in local storage.

Are you using a private browsing window for testing, and closing/reopening it after you change users?

[gu3stZA]

Yes, using private browsing with a new private browsing session for each login attempt

I think you're right - it has something to do with that. I just tested the following:

• open private window and login with user - get access denied
• click on the camera group to see the appropriate cameras
• close private browsing window
• open new private browsing window
• log in with user - I can now see the cameras on login
• close private browsing window
• open new private browsing window
• log in with user - get access denied again

So it does seem to remember the last opened group and use that, but it only lasts across one session. Not sure if that makes sense or not...

[Psix-anp]

Hi. Button Continue not working. Version 0.17.0-1b57fb1

[deanpm70]

You need to select all of the images before you can continue.

[NickM-27]

All images need to be classified

[bagobones]

Some thoughts questions on the /classification screen..

1. Edit: Never mind I realized that "Recent Classification (xxx) is a filter... However A full screen of 100% closed isn't that useful by default.
2. Edit: (These might be but the default view is not obvious?)Could the samples be "capped" by state? IE right now it appears the last 100.. They are all closed.. It might be better to cap 50 per state then I could see the last 50 open and last 50 closed. Might be more important for more complex sets of states?
3. Cap streaks / time periods of 100% for a state to force more training examples? While initially 100% could be a false positive to train against it seems to quickly turn into a screen of 100% / 99% Close for the day?

[NickM-27]

Yeah, I've pushed a change for the next build that will limit saving images for redundant updates that are the same state as before with confidence

[bagobones]

Sounds good

Yeah, I've pushed a change for the next build that will limit saving images for redundant updates that are the same state as before with confidence

Sounds good, I went an checked the "cameras" motion only history and it has a fair bit however object detection is spot on and and at night it probably needs to be this sensitve to pickup people at a distance

Is the Classification model only triggering on motion within its sample box or for the whole camera?

[NickM-27]

It only triggers when motion overlaps with the clarification area

[bagobones]

It only triggers when motion overlaps with the clarification area

Without time stamps I am not sure why it is triggering on all those clear closed cases.. The obvious ones when I scrub the timeline are when the lighting changes / motion spot light triggers but it doesn't explain all the clear daytime ones.

[gu3stZA]

if it's only for the last few minutes then your motion detection is likely too sensitive, assuming no activity is happening

It is real movement - but even 5-6 seconds of movement basically fills up the recent classifications list. But I think your change described above should help so will wait for that!

[lawul]

I use Home Assistant to create custom events in Frigate in some cases (e.g. when door sensors are triggered).
Using state classifications for a while now, I started to do the same with one of the classification sensors as the trigger, so I can easily check the corresponding recording if needed.

Would it make sense to include this functionality directly into Frigate, since the classification data originates from Frigate in the first place?
This would eliminate the need for home assistant or similar software to achieve this. And I think it's likely, that users will use some of their state classifications in a "something happened on a camera, which might needs to be checked" way, without an object triggering an event.

[NickM-27]

Eventually, not in this release though. We will want to understand more about how the feature is used and what desirable behavior would be.

[bagobones]

I currently use custom events for door sensors to ensure there are recordings and event tags on those recordings.

As per lawul it would be nice to setup a classification for doors, gates, maibox state etc and get those tagged as events as while there will also be people there it is also a significant event when they alter those objects.

Hell could use it to track who ate the last slice of pie by classifying the fridge door opening and closing.

[bagobones]

I have previously reported this but since 0.16.1 and still in the current dev, cold loading of the frigate UI is still VERY VERY slow.. 10s ish before the spinning wheels stop and show something. It takes almost twice as long as it used to in 0.16.0 and earlier releases.

Using something like the Advanced camera card in home assistant with all cameras configured takes about 4s for the first camera to load in then they all swap fast.

Would be nice to get to the bottom of this, I am considering upgrading my host but I think bases on past discussion it might be something in go2rtc querying the cameras or something EVERY time frigate is opened.

[bagobones]

Based on the speed of the camera groups vs the ALL I wonder if there isn't a concurrency issue and my backend host is just on the slow side and gets hung up with too many camera requests at once.

[NickM-27]

A browser can only use 6 connections at once, all others will wait. I don't know why your api calls to go2rtc are taking so long. I am working on some changes that should improve this, but will see if they are viable.

[bagobones]

time wget http://192.168.102.30:5000/api/go2rtc/streams/front_yard_cam
--2025-11-29 14:31:17--  http://192.168.102.30:5000/api/go2rtc/streams/front_yard_cam
Connecting to 192.168.102.30:5000... connected.
HTTP request sent, awaiting response... 200 OK
Length: 4603 (4.5K) [application/json]
Saving to: ‘front_yard_cam.1’

front_yard_cam.1              100%[=================================================>]   4.50K  --.-KB/s    in 0s

2025-11-29 14:31:35 (292 MB/s) - ‘front_yard_cam.1’ saved [4603/4603]


real    0m18.101s
user    0m0.000s
sys     0m0.006s

Hmm maybe I can run some tests on this / see if there is a fix in go2rtc for how slow it is.

Going back to past discussions I still don't know why this info needs to be called every time, for every camera AND is set to no-store for cache-control.. The odds of it changing are very low most of the time and it probably could be held valid by E-tags with revalidate or something.

[NickM-27]

It's already cached internally and is only called for the streams that are shown (as opposed to all streams in 0.16)

The next build will consolidate these calls and run them after the images load.

[bagobones]

It's already cached internally and is only called for the streams that are shown (as opposed to all streams in 0.16)

The next build will consolidate these calls and run them after the images load.

The interesting thing is it is not consistent even with WGET however if I call several cameras one after the other at least one will be 10s+ slow

The slowness inconstancy is also present in the go2rtc web ui if selecting prob on a camera so I opened an issue to better understand what might be the cause.

AlexxIT/go2rtc#1960

[gigabyte-1000]

Sometimes I notice frigate shows the camera offline, even though it's online and mainview/subview work fine (which connects directly to go2rtc, birdseyeview will just show black for the camera which I think uses the detection stream).

Looking at "ps aux", I noticed some "ffmpeg [defunct]" processes, I believe frigate is unable to spawn a new ffmpeg process when this happens, it's either stuck trying to kill the existing process which is unkillable for whatever reason, and fails to spawn a new ffmpeg, so the camera cannot resume recording / detection until the docker container is restarted.

Can anything be done to prevent this? Like if there are defunct processes auto restart frigate, or have a timeout in the watchdog and spawn a new ffmpeg process if the old process fails to exit?

[gigabyte-1000]

that's not really how it works, if the main watchdog process gets stuck and unable to terminate the process from a terminate command then there is nothing that can be done because that process is stuck

hmm tricky, there must be some way to circumvent this since restarting frigate either through webiui or docker clears these stuck processes, which usually means ffmpeg is waiting on another process to exit or something is not letting it close proper, maybe when terminating ffmpeg perhaps it should also restart the watchdog + new ffmpeg process? This way it kind of starts new.

I am thinking possibly it tries to close ffmpeg gracefully at first but for whatever reason stalls and fails to exit, then the watchdog tries to force kill it, but the system already sees it as an exiting process and blocks the kill -9 command.

[NickM-27]

There's no use in speculating what might be happening. Closing the watchdog would only extend the problem to whichever process is managing the watchdog (which would effectively just be nested watchdogs, which is not a good pattern)

We've been unable to reproduce the described issue in any way including cpu throttling or randomly killing various processes, so until someone can provide better details / information then we won't be able to make improvements here. We haven't seen a high number of reports especially since the fix in 0.16 which many users reported fixed similar issues

[gigabyte-1000]

There's no use in speculating what might be happening. Closing the watchdog would only extend the problem to whichever process is managing the watchdog (which would effectively just be nested watchdogs, which is not a good pattern)

We've been unable to reproduce the described issue in any way including cpu throttling or randomly killing various processes, so until someone can provide better details / information then we won't be able to make improvements here. We haven't seen a high number of reports especially since the fix in 0.16 which many users reported fixed similar issues

Ok ill keep an eye out next time it happens, I believe it may have to have something to do when the object detector process is running at the same time the process quits (like from a camera reboot), since I noticed the last event in home assistant says “Detected” for motion,

[gigabyte-1000]

Updated to custom ffmpeg 8.0 and go2rtc 1.9.12, and it seems like most of the errors went away, a few normal camera disconnects but no issues re-connecting the cameras. Everything else also working as it should detections, motion & snapshots / tracking.

[we-sell-bags]

it seems to be something related to detected.
normal camera looks like this.

same camera when this happens

log.txt

[adamburgoyne]

If I could train a model to detect if my car on the driveway has it's electric mirrors folded in or out to determine lock state that would be great! I've tried with passing camera grabs to an LLM in homeassistant but they've been extremely unreliable. Fingers crossed!

[Calimerorulez]

I just built my 0.17 setup. Everything seems to be working fine, except when I want to 'view' an audio detection, I only see a snapshot and no clip, so I can't hear the sound, and certainly I can't see it ;-)

[hawkeye217]

This will be fixed in the next build.

[bagobones]

There are a few issues with resizing the tracking details screen if you reduce the size of your browser window too much.

1. The timelines on the right hand side under tracking detail will blend together / overlap. They probably should have a minimum constraint on them.
2. The playback bar gets pushed outside of the video window area and becomes unusable, you can click on it if it is still overlapping the video but if the window is constrained enough the playback bar goes below and stops being clickable.

Unrelated yes my black cat still gets mistaken for a bear.

[bagobones]

I have an ultrawide monitor (2560x1080) so generally ALL windows I use the OS snap function to snap one window left or right making the browser approximately 1280x1080.

This works most of the time just fine but it is something I have just noticed caused issues with this new version of the dialog.

[bagobones]

The shape and aspect ratio of the video also seems to matter. This only seems to be an issue on my doorbell camera that has a more square aspect ratio for the video. My other 16:9 and DUO camera feeds do not get the crushed layout, it is only the more square doorbell.

[bagobones]

[hawkeye217]

What are the stream dimensions of your doorbell with the more square aspect ratio?

[bagobones]

Reolink WiFi Doorbell 2560*1920 so it is a 4:3 camera

[gu3stZA]

In the latest build, I see the following rules for classification models:
Determine if we should save the image for training. Save when: - State is changing or being verified (regardless of score) - Score is less than 100% (even if state matches, useful for training) Don't save when: - State is stable (matches current_state) AND score is 100% """

May I be so bold as to suggest you reduce the 100% requirement to something like 95%? My experience with the classification model is that if you have a scene that isn't completely binary (the states are binary, but the POV has subtle changes), the recent classifications is filled up by a lot of 95% and up classifications. But I don't want to retrain those. I only want lower percentages.

I do realise it is a slippery slope, and there is no "right" answer as to what is low enough to warrant submitting for additional trraining, but 100% feels too lofty a goal.

[NickM-27]

yes, we can't store that many images, so filtering only applies to the images that are still saved. I had cases like that too but I just trained them and now the model is able to cover those cases too and most of them are 100%

The problem is if we try to focus too hard on perfectly matching a specific user then you risk things being lost / unhelpful in general for other users. Realistically, if your sensor is working correctly there is no reason to train any more images

[gu3stZA]

With normal object detection we have a risk of overfitting the model. I guess then what you are implying is that we don't really have that risk here, because the states are binary - so training the 95+ images up to 100 is not dangerous (dangerous in the sense that we are making the model less accurate)?

I'll play around with it and test that hypothesis. My scene is such that I am not able to perfectly crop only the area of interest, so I'm always going to have subtle changes in the scene - but we'll see if it's possible to train most of those to 100%.

Thanks!

[NickM-27]

There is certainly still a risk of overfitting, but that is exactly what training lower scoring images which are different than what have been trained helps. If you often trained 100% scoring images it would be problematic.

To give a more concrete example, see this image for my gate. In this case the 99% images are likely scoring that instead of 100% because we have gotten our first snow of the year and there are no images in the training set for that. Adding a few of these would be the opposite of overfitting.

We can see a few other images where my dog is near the bins or a bug flew by at night. These scoring below 100% due to this and would be fine to train as they are different from the default state. You just want to avoid having images that look very similar in high quantities. But also again, there is no reason to try and train images when the model is working well. Focus on the cases where the sensor is entirely incorrect or perhaps flips between open / closed when it should not have.

[gu3stZA]

Gotcha - thanks!

On a slightly related note - do state classification models take the colour into account? i.e. are two otherwise identical images where one is in IR mode and the other in colour mode the same image from a model point of view? Or are they very different?

I'm purposefully testing this functionality with quite a challenging use-case, so it's probably not likely to come up too often. But I'm trying to train the model that the presence of a person when the camera is in IR mode means state_1, whereas the presence of a person when camera is in colour means state_2.

[NickM-27]

yes, training is done with all three channels, so color will have some affect.

[bagobones]

I know it is off topic for this discussion (if there is a separate thread for this let me know) but what is the point of the "Audio Transcription" as implemented? What benefit is it supposed to provided? It can display temporary? Real time Transcripts, or transcripts can be requested for events manually?

Nether of the use cases I would have wanted it for seem to be possible with the implementation?

1. I would like to have labels / events for key words / phrases that could be used to mark important parts of video or maybe used as triggers in home assistant.. Like setting up a text filter for the word "dad dad" IE twice to maybe trigger an alert.. Or maybe a specific name said.
2. Being able to quickly seek through video and highlight times when specific words where said to find a specific conversation.

[NickM-27]

It will likely make more sense in a future version to open up transcription options once we support external whisper docker containers as well as running the model locally. That way a single service can be used for users that also use whisper for home assistant voice or other features, and it can work on larger machines that have better resources if needed.

[bagobones]

ya I am just going to ignore it in 0.17 as I see no use for it, I guess it provides a silent way to review video or search a LONG event that for some reason you have already identified and manually marked to transcribe?

Was hoping for triggers and labels.

[hawkeye217]

The implementation for 0.17 is basic, but functional. We will see what use cases come up in beta and after the release. The feature certainly has room for improvement.

[bagobones]

For me finding the event is the hard part, a bunch of person events in the same room to go through, I don't want to listen to all of them to find the one where something specific was discuss, an argument started, a specific topic was discuss etc.

I know the notes indicated it would be a complex to batch these but even a low priority serial process that searched for the latest un transcribed event, in the last X hours. Then transcribed it, should recursively catch up over night when few if any new speech events are being created leaving all past days transcribed and searchable assuming the host is not a potato and not too many cameras have been selected to transcribe.

IN most homes people sleep over night, and in most businesses they are a closed at night creating many hours to process files without new events.

[hawkeye217]

We can't assume capable hardware or the same use schedule for all (or even the majority of) Frigate users.

For now, what has been implemented covers what was desired in the feature request, which was to add the ability to transcribe speech events. We'll consider other use cases and more ways to improve this feature in future versions.

[kgomez6]

I think auth with proxy is broken in 0.17 97b29d1-tensorrt builds because what ever I do it force me to directly to login, because my setup with 0.16 tensorrt works before with this nginx proxy manager config

location / {
proxy_pass http://frigate_ip:8971;

# Frigate official secret
proxy_set_header X-Proxy-Secret "secret";

# WebSocket support
proxy_http_version 1.1;
proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection "upgrade";

# Forward client info
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;

proxy_buffering off;
proxy_read_timeout 3600s;
proxy_send_timeout 3600s;

}

And my logs

2025-11-30 23:50:10.466137325 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /api/config HTTP/1.1" 401 581 "http://nvr.home/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.020" upstream_response_time="-" 2025-11-30 23:50:10.466150545 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /api/profile HTTP/1.1" 401 581 "http://nvr.home/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.023" upstream_response_time="-" 2025-11-30 23:50:10.466226291 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /ws HTTP/1.1" 401 581 "-" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.021" upstream_response_time="-" 2025-11-30 23:50:10.534113805 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /api/profile HTTP/1.1" 401 581 "http://nvr.home/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.005" upstream_response_time="-" 2025-11-30 23:50:10.546792466 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /login HTTP/1.1" 200 2668 "http://nvr.home/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.000" upstream_response_time="-" 2025-11-30 23:50:10.624439946 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /login HTTP/1.1" 200 2668 "http://nvr.home/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.000" upstream_response_time="-" 2025-11-30 23:50:10.956285960 192.168.25.5 - - [30/Nov/2025:23:50:10 -0700] "GET /api/auth/first_time_login HTTP/1.1" 200 32 "http://nvr.home/login" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/142.0.0.0 Safari/537.36" "192.168.50.3" request_time="0.003" upstream_response_time="0.002"

[hawkeye217]

Does your frigate config have the necessary requirements?

auth:
  enabled: False

tls:
  enabled: False

proxy:
  auth_secret: "secret"  # Must match your X-Proxy-Secret header
  trusted_proxies:
    - your_nginx_subnet_here