Replies: 33 comments 184 replies
|
Debug Replay I see I can select topics like Face Recognition and LPR but I am not seeing any messages for ether of those when using the debug reply. Even for events that had LPR or Faces in them originally. I would particularly like to debug LPR with this mode and see both the plate images and the debug messages as detection runs over and over so I could try and tune it. GenAI Chat The icon for GenAI chat should probably not appear on the sidebar if NO GenAI agent has been configured. Chapters for review items in exports Nice addition |
|
We'll ask for Debug Replay feedback at a later time, but it should be noted that the Messasges view is showing Frigate's internal communication topics and is not designed to be a replacement for the debug logs. If your plate or face is not detected in the original clip, you won't see messages for it in the Messages pane. Debug logs will continue to show in the regular Frigate logs for the Debug Replay camera. |
I guess it would be more correct to say no genai is enabled in my config, it was configured but has been set to false in all references in case I want to play with it again later. It is actually commented out in the global section of my config, and only hase enabled: false appearing on all my cameras as when I was testing it I had to add the config item to each camera.
Aww, it would then be nice to display the debug logs on this screen? |
Just open another browser window/tab and you can see them scrolling live in the System logs section in the UI. Debug Replay isn't intended to be a one-stop pane for all Frigate diagnostics or a comprehensive debugging environment for every Frigate feature. It merely makes it easier to spin up a dummy camera and perform some common adjustments in real-time. You'll still need to use the normal tools (logs, an MQTT client, etc) to debug your feature. The nice thing is that you can make most adjustments to your actual config now without restarting Frigate. |
I deleted all references to genai from my config and comments the chat icon still appears after restart. Unless I am not understanding what you mean by remove the role. 
It is also not configured in the GUI settings menu 
|
|
I already made changes for the next build for the Chat icon in #23092 |
|
Hello! Wonderful, now I don't have to make several discussions! 🤓 Feel free to ignore these. I am not sure which applies from your post:
Explore > Tracking Details for 1 Object
Export
Continuous history playback
Camera groups:
Intel iGPU Metrics page
Take care! ❤️ 0.18.0-ba4a6a5 |
|
The screenshots are of the classification menu Under explore it has snapshots & tracking details The people are obscured and come back a few times so there's 4 entries under classification and 4 explore entries. I can send you it unedited if you'd like. |
|
A clue about the earlier issue to help debug. 
|
|
Are you trying to use classification to identify a specific person? |
|
Yes, only 4 specific people. This was a false positive. I thought possibly due to the tree and grass it concluded it's someone else. Thank you all for being extremely helpful :) |
|
Frigate's object classification is for distinguishing sub-types of an object (like a delivery person vs. visitor by uniform, dog breeds, vehicle types). It uses MobileNetV2 and keys on visual attributes like clothing/pose, not identity. So unless the 4 people you are trying to identify are always wearing the same clothes or always have the same features, you'll want face recognition, which uses FaceNet/ArcFace embeddings. |
|
Hardware Support Just wanted to note that I am on OLDER intel hardware, particularly a i7-7700T which is in that strange i965 / iHD switch over area and it still seems to be working fine using hwaccel_args: preset-vaapi . Just calling it out since some previous driver and FFMPEG changes have caused issues for older devices, just wanted to say I have not seen any new issues so far. |
|
good to know. |
|
looking forward to the GenAI Chat. Will this be acessible via local API or for example, Voice Agent in Home Assistant, so I can ask it a question like how fast was that last car going or what time was the person detected. etc, via automation in HA? |
|
Yes, the latest version of the HA integration already supports |
|
Fantastic! (I haven't tried a dev version yet, but will try the RC ) |
|
ok, im testing 0.18.0-52a3301 On install, I had to change a couple of genai roles in my yaml (so it may be a breaking change for some). and a few things from this photo when I was using it: 
hope that helps, will continue to test more, but this is awesome! |
|
oh wait, I think I caught the chat error: from this chat: 
|
|
We might need to improve the debug logs, seems like a model error though. Thanks for the note on the speed I'll add a note on that. |
|
I tried to switch to a v18 dev build a while ago, but got a cuda/onnx error, which seems to fit into the "hardware support/nvidia" category. I postet about this a while ago and kind of forgot about it: #22577 (reply in thread) I just tried again with build 6a2b914-tensorrt, which gives me the exact same error. If I can do anything to help narrow the cause down, let me know. :) |
|
We've had a few reports of that being fixed, not sure why it would still be happening on latest |
|
Hm, understand. I'm running frigate via docker in an LXC on Proxmox, which I know is not officially recommended, but has been running solid so far. Last thing I did was updating all os versions to the latest (proxmox itself and debian13 in LXC). And installed the latest nvidia drivers 595.71.05. I also found another similar Issue here #22575 which doesn't seem to be resolved? |
|
I did some more testing and finally found something which might help: I always used fine tuned Frigate+ yolov9s 320x320 (no hailo) models so far.
In every case the error is identical to the one I posted in the other Thread I linked. Do you have any idea what could be different with the yolov9s 320 model? |
|
No, I also run the YOLOv9 S on 5060Ti without any issues, so I am not sure why it would work that way specifically unless it is a host / driver implementation issue that just happens to get hit based on the models size |
|
Yeah, this seems to be some weird issue of the model size combined with something on my system. Was really looking forward to use the new v18 features.. |
|
All is working good, not found bugs, only higher cpu, gpu and tpu usage |
|
It looks like we might have another option that doesn't use intel_gpu_top at all |
Yes Rc6 would be a pretty good measure for camera detection capacity. That works or possibly recording non-zero utilization of compute > throw out the bottom 1/3 of measurements then average the rest and this could be a better measure of maximum capacity. It'd probably have to be lowered a tad bit. Hmmmm would probably have to account for long periods of idle time as well into the heuristic because that is space to use for detection. It'd certainly take some thinking haha. You're right that--since the detector runs for milliseconds--any average or bar graph would look empty with room to spare. When in reality, just a few detections running at once can slow things down a lot. And thanks for all the hard work and community outreach! |
|
After build f448b25, GPU is back with stats like on 0.17.1, TPU is also down, only CPU is spiking to 46% and fan is loud, this was not the case on 0.17.1 |
|
You should use resources like top to figure out exactly which process is using CPU |
|
|
I am not really sure but it seems that snapshots are no longer provided to ..\clips\ directory? There seems to be no more any jpg files. I am using b6fd86a-rk built. 
|
|
Correct, this will be noted in the breaking changes in the release notes. As of 0.18, Frigate no longer saves annotated JPEG snapshots (.jpg) to disk. Only a clean, unannotated WebP snapshot ( Users who previously relied on the annotated .jpg files on disk should instead use the Existing configs containing clean_copy will be automatically migrated. See the documentation: https://docs-dev.frigate.video/configuration/snapshots/#rendering |
It seems the options in the UI settings section haven't been internationalized with i18n. 
I don't think the fields in the labelmap should be sorted by name. The default order of the labelmap is actually suitable for most users. I think it's enough to just move the enabled options to the top. Also, when using the labelmap with the Plus model, there is an unexpected |
|
I think sorting makes sense, if a user is looking for a certain label and does not know if it is in the labelmap or not it becomes a lot more effort to find if it is in the list vs just looking in an alphabetical list |
However, users of non-English languages are completely unable to use it properly. The current sorting is only based on the original text of the labelmap, which does not conform to the usage habits of other languages at all. Moreover, most people only commonly use objects such as person, car, and dog, and these are already placed relatively high up in the default labelmap order. |
|
We can look at improving sorting for other languages, but currently for example to enable |
Yes, I’ve shared a similar opinion before. We should manually adjust the sorting to put common labels such as person and car at the very top, and then pin all enabled items to the top as well, making configuration adjustments much more convenient. |
|
I don't think a shifting UI that changes the location of the label after you make changes and refresh the page makes sense, enabling / disabling shouldn't change the order. If anything, we can look at implementing a search filter in the config UI list items like this, but in alphabetical order it should be pretty easy to go through the list as a user and enable what you want. |
|
By the way, I also think the display of the current global settings is not very user-friendly. I suggest that when the global setting is disabled while the feature is enabled on individual cameras, the global settings page should display a reminder, informing users that this feature is turned on for certain specific cameras. |
|
Additionally, this badge changes its color and text when using the new Profiles feature to indicate that a camera setting is overridden for a specific profile. |
From the current perspective, this doesn't seem very noticeable, and I'm not the only one who has encountered this issue. Would it be possible to add an indicator or similar marker next to this option? |
|
I'm not sure what you mean. If you hover over the Overridden badge, a tooltip will display more information. |
I hope that, just like in my screenshot, within the Global Configuration settings, I can see which options have been overridden with different values at the camera level, and upon hovering over them, display the specific camera names. For example, I may have enabled this option for a single camera individually, but I forget which one it is. It would be very troublesome to look through all cameras to find it. Therefore, I hope to directly see in the global settings which cameras have custom, non-default values configured. |
|
The UI/UX will be increasingly complex when using profiles, so I am not sure a badge like that is the best option. I will think about what can be done to improve this. |
|
Hi after upgrading to 0.18 almost all my camera mainly wireless have this playback issue tapo tpc110_20260504_204752-20260504_204812_typmlt.mp4c110 |
|
Looks like there's a FEATURE FLAG in SAFARI TECH PREVIEW call "Media Containment" that is enabled by default. Did a flag by flag comparison between SAFARI-TP and SAFARI and turned this one off...voila started working. No issues in v0.17 with this flag enabled. Might be worth understanding how v0.18 changed to cause this issue. |
|
Interesting. From what I can see, the "Media Containment" flag tells Safari to do MSE container demuxing inside the web (renderer) process instead of farming it out to a separate sandboxed system daemon. There's very little that has changed with the MSE player in 0.18 beyond some websocket cleanup changes. Maybe it has something to do with the updated version of go2rtc. Nonetheless, glad it's working for you now. |
|
Me, too. ;) I did notice the ws process in network pane (dev console) continually going up size-wise. My concern would be Apple making that default in a mainline release and all your Mac users having issues. |
|
I suppose we can cross that bridge if we come to it, though I bet the technology preview still has a number of bugs. Are you seeing any console logs when you have the flag enabled? And I'd be curious to see if anything changes if you use the same go2rtc version as in 0.17 (1.9.10): https://docs.frigate.video/configuration/advanced#custom-go2rtc-version |
|
Doesn't work with 1.9.10 either. Console log don't really show anything when enabled. Lots when disabled. |
|
Intel Arc A380 running in Proxmox 9 with kernel 7.0. Interference speed ranges from 8ms to 11.15ms, averaging 9-10ms. If I remember corrected it was averaging 8ms with 0.17. I am assuming its not a concern due to new driver in 0.18? Other than that, no bug that I could find. Edit: I have genai review description disable on one of the 5 cameras under one of the profile. But it was still generating reviews description. It just started happening with the latest dev build. I will see if I can reproduce it. |
|
Using 0.18.0-4ff7ab9. |
|
thanks, fixed for the next version |
|
Hi |
|
To be clear, this is not a beta, it is early testing / feedback. It is not ready to be used in production. Builds can be found at https://github.com/users/blakeblackshear/packages/container/package/frigate |
|
I am fully aware of that |
|
The builds are created for every commit, so you would just choose the latest commit and the proper type for your hardware. |
|
You'll put this into docker compose: you get "704ee96" from here: a check mark and 11/11 or 10/10 means there is docker image ready for you (amd64 is last or close to it). Sometimes the builds fail at first. |
|
Can we get openai whisperr capable api for whisperr? |
|
That will likely come in the future, that will be a major re-write |
|
Not for 0.18 as we are feature-frozen, but |
|
Thanks for all your work on this! I haven’t had time to test everything yet, but so far it’s working great.
|
|
Thanks, we'll get those cleaned up |
|
Fellas, this is awesome. I'm trying out f448b25. I'm barely scratching the surface of what it can do, but I am trying to configure everything I can in the UI without touching the config, and I'm slowly but surely getting it. I did manage to get my llama.cpp model (Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf) configured 100% from the UI after a few tries-I kept putting the "/v1" at the end since I've been so used to using that for other stuff. While trying to get there, I clicked the link to the documentation on that settings page, and I guess it's dead https://docs.frigate.video/configuration/genai/config from here- 
I then added an additional model, and the field there shows "gpt-4o" which didn't get replaced once I pushed "save." It printed 
I didn't realize that putting When I got done with all that, I went to the chat tab and asked the Qwen model to keep an eye on the front door for me, which it said it would- I then walked outside and waved to see if it would let me know in the chatbox, but it didn't. I asked it if anyone had come by after I came inside, and it said no, so I asked it to review the most recent events, and it found me and showed snapshots of me. I'm not sure if that's expected behavior, but I was thinking I'd see it tell me it saw me in real time. Does asking it to "Watch a camera" just trigger it to turn on alerts for that camera or something? That's all I came up with so far. Pretty exciting stuff, highly anticipating this release! |
|
I'll add a filter for the /v1 To use embeddings you need to use an actual embeddings model. Where were you watching for the "watch me on the front door"? It gets sent as a built in notification (if you have that enabled) and as an MQTT message. |
|
Also worth noting that any docs links in the UI will not work correctly in the dev builds. You can add |
I was just watching for it the chat window itself, but that makes sense, I just thought it would show an answer straight in the chat window, which I know wouldn't be all that useful real-world. I'll hook up mqtt and try it out next time I spin it up.
Good to know! Some of the other links worked, so I just wanted to point that one out. |
|
I may have missed it but there doesn't seem to be a way of editing the settings of a camera once it's been created. I have two examples:
Obviously, both can be rectified by editing the config. Additionally, noticed that trying to remove Openvino detector (with Frigate+ yolonas9 model) and add USB Coral resulted in model incompatibility error. This is understandable but I wasn't able to remove or change the Frigate+ model from the UI - a catch 22 situation but again, resolvable by editing the config. Finally, seeing this error repeated twice in the log:
This is with either OpenVINO or Coral detectors. |
|
I've made changes to improve this here: #23216 |
|
Awesome! I've spun up a fresh install and will test this evening. Thanks guys, I appreciate this was no simple task :-) |
|
Started the test with OpenVINO configured and working. In Settings -> Detector and model, removed OpenVINO detector, added EdgeTPU detector and set to usb. Frigate prompted "The current Frigate+ model "yolov9s" requires the onnx, rocm, openvino, rknn, zmq detector. Pick a compatible model below or switch to Custom Model before saving." Used dropdown to select Edgetpu model from list (only edgetpu models listed as expected). Clicked save. Save spinner for about 2 minutes then error (this was too quick to note before it disappeared). Clicked "Undo" and tried same steps again. This time, Frigate's error message was "Error saving config: Model does not support detector type of openvino" when an edgetpu model is selected. Logs as below: |
|
Thanks, I pushed a change that should improve this here: #23232 |
|
Looks like the commit above fixes your case but actually introduces another issue. I'll need to do some more thinking on how to best fix this. |
|
Is it expected behavior that other enhancements (such as classification and LPR) do not work while rebuilding the semantic search index? |
|
Different models and code paths are used and there is nothing in Frigate's code itself that prevents enrichments from running when semantic search reindexing is in progress. It could be that the high CPU/GPU load and database writes during reindexing may cause enrichments to be significantly delayed, especially on underpowered systems. |
|
I encountered the following error when accessing the System Metrics page during the creation of the semantic search index: |
|
This is likely caused by CPU saturation and a failing subprocess that retrieves the metrics stats. The change here should prevent the crash. |
|
When I migrated from 0.17 to 0.18.0-1f154a0 I received this error in my config: "Input should be 'chat', 'descriptions' or 'embeddings'" with relation to my GenAI configuration. It autopopulated new lines on it's own that were: I just switched out "vision" and "tools" for "descriptions" and "chat and it worked fine. Apart from that, so far so good! EDIT: Unrelated, sort of, but it's entirely ignorning the "reasoning budget" provider option. I've got llama-server initiated with "reasoning-budget 1" (to essentially disable it, I use this llama-server instance for HA now and I don't want it thinking) and I figured I could just up the reasoning-budget with that provider option via API calls, but it still just thinks for 1 token, ignoring that param. |
|
Great to hear! Let us know how testing goes. We changed the role names a while ago and didn't update the config migrator. Though you're all good to go now, this will be fixed for the next build for new users. |
You can't, llama-server does not support arbitrary CLI args to be set via API, only specific ones and then things that are |
|
thinking is also generally just not recommended for Frigate's features, and any thinking logic won't be shown in the chat UI |
|
You guys are unbelievably responsive, it's crazy, thank you!! |
|
Just a heads up for anyone who finds this in the future, after some digging I discovered you can dynamically change the reasoning budget, it's just a different parameter. From here: ggml-org/llama.cpp#21445 (comment) "You're looking for thinking_budget_tokens. You can include this in the request body ({"thinking_budget_tokens": N}) as long as you haven't specified a budget on the command-line." I just tested it, and can confirm it works. I can now use the same model for HA without having it think/reason at all, and set my reasoning budget for Frigate. As to why I want Frigate to think at all, I'm using gemma-4-26B-A4B-it-GGUF (Q4_1) and it does much much better when it reasons for a little bit than it does when I've got it disabled entirely. Far better rule adherence (I've got my prompt setup mechanically for threat levels, versus letting it determine it on it's own...my prompt is massive, well over 400 lines). I've seen a really great improvement on speed by the way, down to around the 40 second with 1024 thinking tokens, and around 23 seconds without any reasoning budget set at all (referring to using Frigate 0.18). |
|
Working great (appreciate being able to use remotely running llama.cpp on GPU-accel HW for review summaries, embeddings, and chat! Looking forward to MCP and other enrichments being remote-capable) One thing I'm noticing is that on both iOS and Android, I'm getting two notifications for the MQTT topic |
|
that would be a question for the blueprint if that is what you are using, it should do that via the notification tag on Android and id on iOS. |
|
Apparently this is a known issue for the blueprint and the problem seems to be that:
While waiting for this PR to merge, you can load the Beta file for the PR branch and set Tag in the automation to "{{review_id}}" |
|
I seem to have encountered a memory leak issue while using GenAI Embeddings. Here is the memray report file: |
|
peak memory usage is 109MB so it seems to not be much data if it is, but generally if you can take another report after a few hours we can see which areas if any are growing |
Judging by the container's memory usage, the memory keeps increasing and has even reached 20GB. |
|
Seems it's a different process then |
|
I couldn't find the name of any other embedding process—only frigate.embeddings_manager can generate the memray file. |
|
You'll likely want to start by using top to see which process is using memory, then use that as the process marker |
|
Not sure if this was true in 0.17.1 as well, but doing some perf debugging and found this issue it seems would be good to handle correctly - submitted a PR here: Confirmed, model_size: large IS supposed to use the GPU On a single-GPU system, OpenVINO reports ['CPU', 'GPU'] and the check passes. On my dual-GPU system, it reports ['CPU', 'GPU.0', 'GPU.1'] - the exact string "GPU" isn't in the list, so it returns False and all enrichments fall back to CPU via ONNX Runtime. This means my face recognition and semantic search large models have been running entirely on CPU this whole time, which explains the 77% CPU / 2.4GB RAM. Confirmed. The fix is a one-line change - startswith instead of exact match. Patching it in the running container and testing: The results are dramatic. The detection inference halving from 40-60ms to 19ms is because the Arc A310's compute EUs are no longer contending with CPU-based enrichment overhead (memory bandwidth, cache pressure). With enrichments properly on GPU, everything runs more efficiently. |
|
I am running into a (probably harmless?) If I adjust the permissions of the folder recursively and restart, the log statement goes away. When further previews are created, they do not have the same permissions and the warning returns the folder: no warning: warning: if the files made it to there and are readable, it seems fine, so maybe the probe at startup is just a bit strict? This is an NFS mount on a synology if that is related (my whole frigate storage folder is on this mount) Probably related: #23082 |
|
we don't currently do anything with that, it is unclear to me why the permissions would be any different than any other file that is written, but we can look into it. |
|
LLM is pointing me to potentially os.access and NFS issues https://man7.org/linux/man-pages/man2/access.2.html and https://docs.python.org/3/library/os.html#os.access I don't use NFSv2, but 🤷 So most likely nothing to be concerned about, I am fine to ignore it and it makes sense why it started showing up after that commit (and addition of an access check) I might make some adjustments on my side but I will more likely just ignore it for now |
It looks like there’s an issue with the time tool used in the Chat function. When I asked who was in the garden earlier (I asked around 6:30 PM), it queried the time range from 6:00 AM to 7:00 AM instead. It doesn’t seem to be correctly distinguishing between AM and PM.
|
|
seems like a deviation from standard openai API standard to support a different model feature |
Can I add a new Provider similar to Azure OpenAI? Most generative AI service providers in China have slight differences in their standards compared to the standard OpenAI API. |
|
we can definitely look at some extensions of it, ideally it is a sub class so the entire code does not need to be re-implemented. I also don't think we want to have different providers for every service as this would not be maintainable. |
You're right — maintaining this going forward would be very troublesome.
Could this be made optional? Perhaps allowing users to choose whether to pass back the reasoning_content? After all, DeepSeek is indeed one of the top-tier generative AI API providers in China. |
|
No, the OpenAI implementation itself needs to stay true to the spec, the wider the surface area gets the more likely something will break as we make changes in subsequent versions, and also the more difficult maintenance / options will become as more changes happen in DeepSeek and related APIs in the future. Realistically if users want this then worst case they can just add a slightly tweaked provider locally as the plugin system makes this quite easy to do. |
|
Now that you mention it, why does the Audio filters section under Audio detection show settings for unselected audio types? What’s the purpose of doing this? The number of entries is far too overwhelming. |
|
I recently changed this behavior for object filters. I'll improve it for audio filters a future dev build. |
|
I'm currently using the Qwen VL Embedding 2B model for semantic search, but when searching in Chinese within Frigate, the accuracy drops significantly. It seems this is because sqlite-vec is currently set to 768 dimensions, causing a lot of detailed data to be lost. Are there any plans to optimize this in the future? |
|
at some point eventually, for now this is just a "nice to have" that hasn't seen much love on the llama.cpp side and isn't available in many other common providers. Qwen3-VL and others do put the most valuable tokens up front so it is generally okay to take a subset. |
|
With regard to the new GenAI chat function, I’ve just started playing around with. If I ask something like “when was the last time you saw (cats name) in kitchen?” it works great. However, other things that work when I manually search them don’t seem to work in the GenAI chat. Here’s a specific example (and my setup is only using GenAI for review summaries, not object descriptions). If within the search function of Frigate I search “riding mower” I get a ton of results from when they were here earlier this week. If I ask GenAI Chat “anyone using a riding mower this week?” it tells me no (I’ve attached screenshots). I tried in various other ways, like “any evidence anyone has cut my grass this week?” and again it says no. I’ve got no idea if this is because I’m expecting the chat feature to work differently than intended or if this is actually a bug. |
|
Let us know how it's working for you! |
|
WHHAAATTTTT?!!! This is CRAZY hahahaha. Pardon my lack of professionalism in my reply, but this is beyond exciting. At least this and HA match what my naive adolescent hopes & beliefs were for the future, this is just so damn cool. Check it out: |
|
By the way, in case you did not know, there is a |
|
It just gets better…it’s capturing ALL of the nuance that I have my GenAI review summary setup to include (I’ve got a massive prompt, but I’ve instructed it to include detailed descriptions of people: gender, age group, clothing, etc) and I can literally chat and ask things like “show me any man wearing a blue shirt in the past 48 hours” ….and it’s working!!! |
I just discovered that without exaggeration about 45 minutes ago. I’m going through and updating all my containers (zigbee, HA, etc, etc) and finally made the switch from “extended OpenAI conversation” to “skye harris hass local openai llm” (finally working at appropriate speeds, sub 1.5 seconds for my local voice assistant) and I noticed it had a built in “frigate services” function…ticked the box, and was able to ask the last time it saw anything of interest out front. The fact that I can ask it questions this detailed is, as previously stated, blowing my mind hahaha. Thank you guys!!! |
|
I am having issues with ... something, I switched out a quadro P400 for an intel Arc B50 due to not being able to process (ffmpeg processes) all of my cameras. I am running Frigate build b712e1f on Unraid 7.3.0, others on reddit have had success using a B50 gpu for other things with this version of Unraid. I have tried QSV and VAAPI presets as well as iHD and Xe for driver selection, though I'm unsure if Xe is a real option. I am not able to get the preview to load on the home page for the camera, but it will playback if I click into it using the config below. I also cannot get a preview of the footage on the Masks/Zones menu nor the camera debug display. I have tried adding cameras via the new gui method, as well as copy pasting from my old config file, and get similar results. I haven't attempted detect yet with the intel card, I have a hailo 8L I had been using for that task previously. I am unsure if re-bar is enabled on this motherboard, I have "Above 4G Decoding" enabled but did not find an option for re-bar in the BIOS. Below are the configuration files and log outputs of a simplified setup, which exhibits the same issue as when I have all of the cameras setup together. Docker setup in unraid: Frigate config.yaml mqtt:
enabled: true
host: 192.168.1.40
port: 1883
user:
password:
ffmpeg:
hwaccel_args: preset-vaapi
input_args: preset-rtsp-restream-low-latency
output_args:
record: preset-record-generic-audio-copy
go2rtc:
streams:
Front_Yard_West:
- ffmpeg:rtsp://user:pass@192.168.50.20/cam/realmonitor?channel=1&subtype=0#video=h264#hardware#audio=aac
Front_Yard_West_Sub:
- rtsp://user:pass@192.168.50.20/cam/realmonitor?channel=1&subtype=1
cameras:
Front_Yard_West:
enabled: true
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/Front_Yard_West
roles:
- record
- detects
version: 0.18-0
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
enabled: true
detect:
enabled: falsedocker log output GPU Stats in Unraid's gui: Other PC specs if needed: |
|
I didn't think not having detect setup would break it... but this config file appears to work. Here's the whole config file: One issue I suppose is that because I'm running detect on the main stream as the docs say not to do, I am maxing out the gpu and nearly maxing out the cpu, the reason I had it like this with the Quadro card and the hailo 8L is to better detect birds from further away. At least I think that was necessary to do that. 7 of the cameras are Amcrest 5MP turrets, so detect is running at 3072 x 2048 at 5 fps. The 8th camera is a pi camera running at 1080p. Here's the detector and gpu usage as reported by frigate in this configuration: |
|
There's no need to run that high, running at 1920x1080 should be plenty, you can downscale the main stream for that. |
|
I defined detect to be 1920x1280, due to aspect ratios, and now this is how the frigate metrics look for reference: I don't think there's an issue at this point with the intel stuff, just replying for closure. I'm happy to test more things if that is desired with regards to the intel card. |
|
I have an issue with the recordings it turns out. Frigate doesn't seem to be recording much at all to disk, system memory fills up instead. Below is the config file currently:mqtt:
enabled: true
host: 192.168.1.40
port: 1883
user:
password:
ffmpeg:
hwaccel_args: preset-intel-qsv-h264
input_args: preset-rtsp-restream
output_args:
record: preset-record-generic-audio-copy
apple_compatibility: true
go2rtc:
streams:
Front_Yard_West:
- ffmpeg:rtsp://user:pass@192.168.50.20/cam/realmonitor?channel=1&subtype=0 #video=h264#audio=copy#hardware
Front_Yard_West_Sub:
- rtsp://user:pass@192.168.50.20/cam/realmonitor?channel=1&subtype=1
Back_Yard_East:
- ffmpeg:rtsp://user:pass@192.168.50.21/cam/realmonitor?channel=1&subtype=0 #video=h264#audio=copy#hardware
Back_Yard_East_Sub:
- rtsp://user:pass@192.168.50.21/cam/realmonitor?channel=1&subtype=1
East_Side:
- ffmpeg:rtsp://user:pass@192.168.50.22/cam/realmonitor?channel=1&subtype=0#video=h264#audio=copy#hardware#rotate=90
East_Side_Sub:
- ffmpeg:rtsp://user:pass@192.168.50.22/cam/realmonitor?channel=1&subtype=1#video=h264#audio=copy#hardware#rotate=90
Back_Patio_and_Yard_West:
- ffmpeg:rtsp://user:pass@192.168.50.23/cam/realmonitor?channel=1&subtype=0 #video=h264#audio=copy#hardware
Back_Patio_and_Yard_West_Sub:
- rtsp://user:pass@192.168.50.23/cam/realmonitor?channel=1&subtype=1
Front_Porch:
- ffmpeg:rtsp://user:pass@192.168.50.24/cam/realmonitor?channel=1&subtype=0#video=h264#audio=copy#hardware#rotate=90
Front_Porch_Sub:
- ffmpeg:rtsp://user:pass@192.168.50.24/cam/realmonitor?channel=1&subtype=1#video=h264#audio=copy#hardware#rotate=90
West_side:
- ffmpeg:rtsp://user:pass@192.168.50.25/cam/realmonitor?channel=1&subtype=0 #video=h264#audio=copy#hardware
West_side_Sub:
- rtsp://user:pass@192.168.50.25/cam/realmonitor?channel=1&subtype=1
Driveway_East:
- ffmpeg:rtsp://user:pass@192.168.50.26/cam/realmonitor?channel=1&subtype=0 #video=h264#audio=copy#hardware
Driveway_East_Sub:
- rtsp://user:pass@192.168.50.26/cam/realmonitor?channel=1&subtype=1
cameras:
birdcam1:
ffmpeg:
inputs:
- path: rtsp://192.168.50.100:8554/birdcam1
input_args: preset-rtsp-generic
roles:
- detect
# - clips
- record
audio:
enabled: false
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
#motion:
# mask:
# - 0,0.001,0,0.118,0.232,0.112,0.489,0.081,0.595,0
# - 0.002,0.092,0,0.284,0.268,0.207,0.37,0.169,0.471,0.142,0.537,0.077,0.587,0.067,0.618,0.078,0.664,0.048,0.742,0.09,0.797,0.165,0.843,0.348,0.889,0.442,1,0.501,1,0.002,0.549,0.001
# threshold: 21
# contour_area: 11
# improve_contrast: true
#objects:
# track:
# - bird
# mask: 0,0,0.257,0,0.231,0.055,0,0.044
Front_Yard_West:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/Front_Yard_West
roles:
- record
- detect
- audio
live:
streams:
Main Stream: Front_Yard_West
Sub Stream: Front_Yard_West_Sub
detect:
width: 1920
height: 1280
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
zones:
pond:
coordinates:
0.423,0.787,0.379,0.712,0.378,0.626,0.397,0.535,0.362,0.491,0.33,0.487,0.333,0.441,0.292,0.392,0.236,0.377,0.197,0.416,0.196,0.474,0.139,0.554,0.104,0.601,0.186,0.815,0.266,0.929,0.321,0.945,0.364,0.918,0.406,0.89,0.44,0.842
loitering_time: 0
friendly_name: Pond
inertia: 3
front_walkway:
coordinates:
0.044,0.659,0.102,0.601,0.176,0.5,0.196,0.47,0.2,0.413,0.176,0.396,0.094,0.476,0.016,0.558
loitering_time: 0
friendly_name: Front Walkway
driveway:
coordinates:
0.095,0.475,0.302,0.303,0.406,0.227,0.439,0.207,0.328,0.191,0.293,0.217,0.082,0.388
loitering_time: 0
friendly_name: Driveway
inertia: 3
front_windows:
coordinates:
0.043,0.665,0.102,0.602,0.184,0.818,0.26,0.926,0.322,0.951,0.326,0.995,0.001,0.998,0,0.712
loitering_time: 0
friendly_name: Front Windows
front_yard:
coordinates:
0.448,0.997,0.993,1,0.971,0.371,0.962,0.285,0.485,0.173,0.224,0.364,0.242,0.375,0.288,0.387,0.319,0.416,0.335,0.435,0.336,0.48,0.364,0.489,0.397,0.53,0.38,0.624,0.382,0.712,0.423,0.783,0.441,0.84,0.441,0.881
loitering_time: 0
friendly_name: Front Yard
motion:
threshold: 30
contour_area: 16
improve_contrast: true
review:
alerts:
required_zones:
- pond
- front_walkway
- driveway
- front_windows
- front_yard
Back_Patio_and_Yard_West:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/Back_Patio_and_Yard_West
roles:
- record
- detect
- audio
live:
streams:
Main Stream: Back_Patio_and_Yard_West
Sub Stream: Back_Patio_and_Yard_West_Sub
detect:
width: 1920
height: 1280
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
threshold: 30
contour_area: 16
improve_contrast: true
zones:
back_patio:
coordinates:
0.182,0.522,0.262,1,0.882,1,0.959,0.751,0.599,0.405,0.533,0.345
loitering_time: 0
friendly_name: Back Patio
inertia: 3
back_yard:
coordinates:
0.183,0.516,0.157,0.345,0.327,0.26,0.329,0.233,0.005,0.394,0.003,0.995,0.258,0.996
loitering_time: 0
friendly_name: Back Yard
back_kitchen_window:
coordinates: 0.453,0.384,0.337,0.259,0.158,0.346,0.185,0.519
loitering_time: 0
friendly_name: Back Kitchen Window
review:
alerts:
required_zones:
- back_patio
- back_yard
- back_kitchen_window
Back_Yard_East:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/Back_Yard_East
roles:
- record
- detect
- audio
live:
streams:
Main Stream: Back_Yard_East
Sub Stream: Back_Yard_East_Sub
detect:
width: 1920
height: 1280
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
threshold: 30
contour_area: 12
improve_contrast: true
zones:
back_patio:
coordinates: 0.145,0.728,0.443,0.463,0.239,0.301,0.074,0.409
loitering_time: 0
friendly_name: Back Patio
back_kitchen_window:
coordinates: 0.139,0.734,0.213,1,0.517,1,0.732,0.732,0.445,0.463
loitering_time: 0
friendly_name: Back Kitchen Window
back_yard:
coordinates:
0.517,0.999,0.848,0.995,0.956,0.392,0.958,0.177,0.907,0.128,0.796,0.075,0.658,0.082,0.592,0.091,0.491,0.107,0.353,0.156,0.257,0.217,0.23,0.241,0.236,0.295,0.444,0.463,0.728,0.731
loitering_time: 0
friendly_name: Back Yard
inertia: 3
review:
alerts:
required_zones:
- back_yard
- back_kitchen_window
- back_patio
East_Side:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/East_Side
roles:
- record
- detect
- audio
live:
streams:
Main Stream: East_Side
Sub Stream: East_Side_Sub
detect:
width: 1280
height: 1920
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
threshold: 30
contour_area: 16
improve_contrast: true
zones:
back_yard:
coordinates: 0.139,0.174,0.408,0.136,0.17,0.075,0.104,0.083
loitering_time: 0
friendly_name: Back Yard
inertia: 3
east_side:
coordinates:
0.138,0.175,0.112,0.18,0.151,0.357,0.271,0.599,0.391,0.797,0.578,0.998,0.739,0.999,0.998,0.996,0.999,0.493,0.558,0.18,0.411,0.136
loitering_time: 0
friendly_name: East Side
review:
alerts:
required_zones:
- east_side
- back_yard
Front_Porch:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/Front_Porch
roles:
- record
- detect
- audio
live:
streams:
Main Stream: Front_Porch
Sub Stream: Front_Porch_Sub
detect:
width: 1280
height: 1920
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
threshold: 30
contour_area: 16
improve_contrast: true
zones:
front_door:
coordinates: 0.61,0.056,0.436,0.421,0.723,0.534,1,0.218,0.991,0.143
loitering_time: 0
friendly_name: Front Door
front_porch:
coordinates:
0.354,0.376,0.84,0.576,0.499,0.799,0.14,1,0.002,0.998,0.001,0.504
loitering_time: 0
friendly_name: Front Porch
review:
alerts:
required_zones:
- front_porch
- front_door
West_side:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/West_side
roles:
- record
- detect
- audio
live:
streams:
Main Stream: West_side
Sub Stream: West_side_Sub
detect:
width: 1920
height: 1280
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
threshold: 30
contour_area: 16
improve_contrast: true
zones:
west_yard:
coordinates: 0.25,1,1,1,0.998,0.714,0.293,0.201,0.135,0.184,0.163,0.595
loitering_time: 0
friendly_name: West Yard
inertia: 3
review:
alerts:
required_zones: west_yard
Driveway_East:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/Driveway_East
roles:
- record
- detect
- audio
live:
streams:
Main Stream: Driveway_East
Sub Stream: Driveway_East_Sub
detect:
width: 1920
height: 1280
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
threshold: 30
contour_area: 16
improve_contrast: true
zones:
driveway:
coordinates:
0.317,1,0.192,0.635,0.139,0.408,0.124,0.29,0.394,0.152,0.684,0.332,0.92,0.532,0.87,0.738,0.793,1
loitering_time: 0
friendly_name: Driveway
east_yard:
coordinates:
0.315,0.997,0.196,0.644,0.136,0.408,0.114,0.223,0.019,0.285,0.033,0.997
loitering_time: 0
friendly_name: East Yard
front_yard:
coordinates:
0.717,0.353,0.726,0.315,0.755,0.305,0.819,0.29,0.847,0.289,0.866,0.285,0.895,0.297,0.921,0.31,0.929,0.275,0.907,0.253,0.796,0.166,0.686,0.097,0.643,0.078,0.351,0.123,0.395,0.149,0.688,0.332
loitering_time: 0
friendly_name: Front Yard
front_walkway:
coordinates:
0.825,0.449,0.911,0.522,0.922,0.467,0.928,0.438,0.879,0.401,0.863,0.401,0.849,0.411
loitering_time: 0
friendly_name: Front Walkway
pond:
coordinates:
0.823,0.448,0.846,0.409,0.861,0.398,0.88,0.399,0.927,0.434,0.937,0.363,0.917,0.35,0.92,0.312,0.872,0.289,0.813,0.292,0.727,0.317,0.719,0.357
loitering_time: 0
friendly_name: Pond
review:
alerts:
required_zones:
- driveway
- east_yard
- front_yard
- front_walkway
- pond
birdseye:
enabled: false
detectors:
ov_0:
type: openvino
device: GPU
model:
path: plus://2c0ccc3ada5752fa1f1bbe416f6328ae
snapshots:
enabled: true
objects:
track:
- person
- face
- car
- bicycle
- squirrel
- dog
- cat
- bird
- rabbit
- usps
- ups
- fedex
- package
- waste_bin
filters:
bird:
min_score: 0.7
threshold: 0.8
min_area: 7000
person:
min_score: 0.85
threshold: 0.9
car:
min_score: 0.80
threshold: 0.9
min_area: 19000
dog:
min_score: 0.85
threshold: 0.90
cat:
min_score: 0.7
threshold: 0.85
squirrel:
min_score: 0.8
threshold: 0.85
rabbit:
threshold: 0.85
min_score: 0.8
package:
threshold: 0.9
min_score: 0.8
audio:
enabled: true
listen:
- bark
- fire_alarm
- scream
- speech
- yell
- knock
- cat
version: 0.18-0
semantic_search:
enabled: true
reindex: false
model_size: large
notifications:
enabled: 'true'
email:
#detect:
# enabled: true
face_recognition:
enabled: true
model_size: large
lpr:
enabled: false
classification:
bird:
enabled: true
custom:
Known cars:
enabled: true
name: Known cars
threshold: 0.8
object_config:
objects:
- car
classification_type: sub_label
record:
enabled: true
continuous:
days: 7
detections:
retain:
days: 31
mode: all
post_capture: 10
alerts:
post_capture: 10
retain:
days: 31
mode: all
motion:
days: 31
motion:
enabled: true
detect:
enabled: true
telemetry:
stats:
amd_gpu_stats: false
birdseye:
enabled: falseBelow are the frigate logs:and later when I turned on debug messageslogger:
logs:
frigate.record.maintainer: debugIt looks like only birdcam1 is being saved to disk, which I can confirm when looking back at the saved footage. The previews of the other streams' times are changing, but not when I click into one of them. frigate_help.mp4output of "iotop -o" on the host machineFrigate memory usage as reported by docker |
|
Your ffmpeg processes are crashing constantly, so that's the issue |
|
I've noticed 2 things with the new chat feature...
|
|
|
|
It's not at the bottom? That's not related to the model at all, it's just pulling from the result of the tool. |
|
Looks like there are a number of urls in some of the new 0.18 frontend code that aren't correctly respecting |
Correct. When I use gemma4-e4b the snapshots don’t show up. |
|
I previously encountered some memory leak issues that would eventually consume all the host machine’s memory, causing the host to crash. To prevent this, I set a maximum memory limit (--memory=20g) in Frigate’s Docker configuration. Below is the last line from
At this point, other Frigate-related processes inside the container are still running, and the container itself remains active. But once the main process is terminated, all functionality becomes unavailable, requiring a manual forced restart of the container. |
|
I'll add |
On another note, would it be possible to automatically restart the main process when it terminates unexpectedly? |
|
I think Frigate itself should fully restart, but we will see |
|
feedback continued here |
and others
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Frigate 0.18 includes many new features, some of which require additional feedback to ensure a smooth user experience. Frigate 0.18 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 the following features, all other feedback will be ignored for the time being. If you test any of the following, please leave a note even if it just says that things are working well.
If you experience any bugs with config migration, or bugs with existing 0.17 functionality then please give that feedback as well.
When providing feedback please do so in this discussion, if a new discussion must be created then use the
Beta Supportcategory notBugThe docs are: https://docs-dev.frigate.video/
Note that docs links in the UI in
devbuilds point to the stable docs hostname and may lead to an outdated page or 404s. You can manually add-devto the hostname for the correct link. For example:https://docs.frigate.video/configuration/genai/configbecomeshttps://docs-dev.frigate.video/configuration/genai/configHardware Support
Nvidiadependencies have been updated to support Blackwell and some fixes have been put in place for GPUs in general. Please give any feedback on issues using Nvidia GPUs.Inteldependencies have been updated to support Battlemage GPUs, this includes driver and firmware upgrades, any Intel users experiencing issues with iGPUs or GPUs please give that feedback. Positive feedback about Battlemage GPUs working is welcome as well. Would also be great to collect inference times on GPUs.Config UI
Frigate 0.18 introduces complete UI-based configuration management. The new Settings experience provides:
Please provide any feedback on the above aspects of the Settings UI. One suggested testing method would be to set up a fresh instance of
devto match your existing stable instance's configuration by exclusively using the Settings UI.GenAI Chat
Frigate 0.18 introduces a Chat interface for talking to your configured GenAI provider directly from the Frigate UI. Chat is implemented as a tool-calling LLM agent with access to Frigate-specific tools, so you can ask natural language questions about your cameras, tracked objects, and review activity.
Please provide any feedback on the chat feature, would be great to see cases where it works well and cases where it may have not behaved as expected. When giving feedback please show a clear screenshot as well as info about what model and backend provider was used.
All reactions