Replies: 41 comments 192 replies
-
|
Also, 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! |
Beta Was this translation helpful? Give feedback.
-
|
First attempt to load latest dev build ended up with Config Editor (Safe Mode) due to this line: Commented out |
Beta Was this translation helpful? Give feedback.
-
|
Note that this seems to be the |
Beta Was this translation helpful? Give feedback.
-
|
yes, this has been fixed in the latest |
Beta Was this translation helpful? Give feedback.
-
|
Hm, I'm still seeing this issue on #85f7138 (ghcr.io/blakeblackshear/frigate:85f7138-rocm / 0.17.0-85f7138), which is currently the latest commit. |
Beta Was this translation helpful? Give feedback.
-
|
to be clear the config will be considered invalid. The issue that is fixed is that if the migration runs, it will not save that value there |
Beta Was this translation helpful? Give feedback.
-
|
Got it, thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Made a config change, hit save and restart, and got this series of errors I've not noted before. Restarted 5 minutes later, got a similar error on a different camera: |
Beta Was this translation helpful? Give feedback.
-
|
that is generally fine, just issues while shutting down which did not stop Frigate from stopping |
Beta Was this translation helpful? Give feedback.
-
|
i noticed when i try to create state classification i got this logs |
Beta Was this translation helpful? Give feedback.
-
|
might be an issue with the Nvidia GPU, it works fine for me, but again I have been looking for a reason to remove it so will just do that |
Beta Was this translation helpful? Give feedback.
-
|
sure ill wait for your updated version. everyday ill always wait for you guys update before ill go to sleep wait wait to try beta and stable version we are so close to beta its looks like |
Beta Was this translation helpful? Give feedback.
-
I would expect a few more posts like this, it definitely will be a little while |
Beta Was this translation helpful? Give feedback.
-
|
i just tried new version 338b681, its says "Don't use GPU for training" but i got error log like ModuleNotFoundError: No module named 'tensorflow' |
Beta Was this translation helpful? Give feedback.
-
|
interesting, I may have missed something |
Beta Was this translation helpful? Give feedback.
-
|
On "Create New Classification", when creating a "State" type classification, it's slightly confusing to me that the different states I add are labelled "Classes". This might be technically correct but is it clearer to just says "States"? The tooltip even says "Define the different states..." 
|
Beta Was this translation helpful? Give feedback.
-
|
yeah, it is generic because classes works for objects and states, probably fine to make that dynamic |
Beta Was this translation helpful? Give feedback.
-
|
looking for better homekit guide 
if i keep frigate network host mode my frigate will keep on restart. 2nd: some problem with audio transcribe with Telugu not tried with English yet because no one speak here 0% when i press transcribe ill get logs like 2025-10-31 19:18:11 but after waiting 10min nothing happing i see my GPU and CPU will go up but then few with later no logs and nothing just silent if i press again it start no log file failed or completed. or one time ill get like only ones its not even Telugu something else 
and trying lang is te |
Beta Was this translation helpful? Give feedback.
-
|
Thanks, however we are not looking for feedback on these things currently as it says in the OP |
Beta Was this translation helpful? Give feedback.
-
|
How do you delete classification models? I'm creating quite a few tests; looks like there are a few different locations I need to delete things in (and possibly the DB?) |
Beta Was this translation helpful? Give feedback.
-
|
Yeah we haven't added that yet. Nothing in the DB, just delete the files in /clips |
Beta Was this translation helpful? Give feedback.
-
|
I am running it since a few weeks and it is rock stable.
|
Beta Was this translation helpful? Give feedback.
-
|
In that case nothing in the logs regarding the down camera: It happens for go2rtc camera and direct camera: Some process are still there for that down camera: Up camera has: |
Beta Was this translation helpful? Give feedback.
-
|
@NickM-27 Here is some early NPU feedback FWIW. I changed my 2 Openvino detectors from GPU to NPU. Inference is maybe 1ms higher than GPU. This is on a Arrow Lake 225H running Proxmox with Frigate in a LXC container. 32GB Ram, 4vcpu's. My build is a day old (brand new setup). I have 4 cameras passing 8fps of 1920*1080 h265 CBR 1.5Mb/s Ill be playing around with this build for a while yet so happy to test/break anything you want to know..... Build 0.17.0-1e50d83 ov_1: model: Vainfo: |
Beta Was this translation helpful? Give feedback.
-
|
I seem to be getting face recoginition crashes. Only thing in the logs is. |
Beta Was this translation helpful? Give feedback.
-
|
Does the output of |
Beta Was this translation helpful? Give feedback.
-
|
Yep! |
Beta Was this translation helpful? Give feedback.
-
|
Yep, your OOM killer is killing the process. Are you running in an LXC? You probably should increase the available memory. |
Beta Was this translation helpful? Give feedback.
-
|
I am using an LXC, i'll try increasing the memory |
Beta Was this translation helpful? Give feedback.
-
|
This probably goes without saying as it's been said many times, but Frigate does run best on bare metal in Docker. We see similar issues with Proxmox users all the time. |
Beta Was this translation helpful? Give feedback.
-
|
I have a 185H cpu where I got NPU to work on 0.16 with some packages inside the frigate container. Now on 0.17 it doesnt say "NPU" in the system metrics. Im using frigate+ with yolo9 detectors: 
|
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks, should be fixed in the next build |
Beta Was this translation helpful? Give feedback.
-
No more errors, but it doesn't seem to ever start up. seems interesting? |
Beta Was this translation helpful? Give feedback.
-
|
NPU Intereference speed example with image: ghcr.io/blakeblackshear/frigate:256817d: Setup:
Detector interference speed:
Detector CPU usage: same at both: around 1,9% The power consumption of this setup is with NPU the same like with iGPU. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
I've started playing with classification (state) for a iron gate which is in plain sight of one of my camera. The "creation" UI is fine but can't seem to be accessed after initial setup (should be available as Edit in "..." to change Name or add additional states. Where is the actual "state" of the particular object supposed to be found ? |
Beta Was this translation helpful? Give feedback.
-
Yes, understood, I had a look at the code in the meantime : I've added an ugly try/except clause locally and it basically seems to work (and prevent crash of the whole embeddings thread). |
Beta Was this translation helpful? Give feedback.
-
|
I've also got the same kind of error when deleting previously classified clips, as after delete the UI is not reloaded and you're still able to select the same thumbnail. |
Beta Was this translation helpful? Give feedback.
-
|
Can you elaborate, I do not see any way in the UI to delete without the UI refreshing |
Beta Was this translation helpful? Give feedback.
-
|
If I open my object page, go to one of the object classified state pages, select one or several file and hit "Delete" I get a green pop-up but the page is not reloaded and the thumbnail are still there. Selecting them again and hitting "Delete" a 2nd time throw a red pop-up and the related error in the log. |
Beta Was this translation helpful? Give feedback.
-
I've slightly modified your fix locally to the following as "files" (at least thumbnails) where actually going over "100" which is the harcoded limit : |
Beta Was this translation helpful? Give feedback.
-
|
Having a search box to filter logs would be very helpful. As would having, on mobile, tapping the top of the screen scroll the log window frame to top instead of the page itself. Right now I don't see a way to jump to top of logs on mobile like there is on desktop Also what are people most commonly doing for genai ? I'm using the recordings setting for review summaries which I know mentions rate limits but I'm hitting Google's extremely quick. I really don't want to pay for a service to use this feature but at the same time I'm not sure I have a good enough system to run ollama. I'll need to look into it |
Beta Was this translation helpful? Give feedback.
-
|
I was using a 3050 via USB4 eGPU to a mini pc and that worked well, I have since upgraded to a 5060Ti which had problems with eGPU so I built a open bench to run it. Works very well for what it is, and doesn't use much power at all when idle |
Beta Was this translation helpful? Give feedback.
-
|
I just have a 1650 super so I think I'm probably SOL to do genai locally. I'll need to look more into Googles API and see if I can get more usage somehow and/or goto the recommended review summary settings which will cut down some |
Beta Was this translation helpful? Give feedback.
-
|
You might want to look into Ollama cloud |
Beta Was this translation helpful? Give feedback.
-
|
Not sure if this is me but I have moved to a bigger HHD, i moved all the files from the old to the new, my faces show in the faces section but people aren't being recognised i just get unknown. if i click on retain i get this: |
Beta Was this translation helpful? Give feedback.
-
|
I rolled back to commit 88a2f6c and its working again now? |
Beta Was this translation helpful? Give feedback.
-
|
If I had to guess something just didn't register correctly, there have been no changes made to faces, so it was probably just the restart that did it. |
Beta Was this translation helpful? Give feedback.
-
|
I'll switch back to 01452e4 then. |
Beta Was this translation helpful? Give feedback.
-
|
There haven't been any changes but the error you are seeing does make sense, so I have put up a fix for that |
Beta Was this translation helpful? Give feedback.
-
|
So I was trying to use the classifications, and while I was able to set it up, it never seems to actually classify anything. In the logs I found this, not sure if its related or not: 2025-11-09 00:06:51.457137862 Exception in thread embeddings_maintainer: |
Beta Was this translation helpful? Give feedback.
-
|
I'll see about adding some more debug logs to the classification running |
Beta Was this translation helpful? Give feedback.
-
|
I'm almost wondering whether I moved too fast and caused lots of issues. I've now had my first little bit of success with it. As I said, last night before I went to bed, I trained the 2 models, saw that error. Upgraded and left it overnight, and got a bunch of unknown classifications. I'm guessing that error happened during training? First thing I did this morning was train several of those unknown classifications, but still kept getting unknown. I finally decided to hit the Train button, and then after it gave me a notification that training was complete, I now have the first model producing classifications (I need to go over and open the door to see if it works still). I hit train on the second one, but its yet to produce a classification, but there hasn't been much motion over there yet. After waiting for the second one to train, I tried adding a third state classification, and I believe it worked (although its giving bad results, I think I might need to reposition my camera for this one, definitely wasn't as obvious as the first 2) I'd say as far as the UI, it'd be nice to have some real time feedback of whats going on. If I've classified images, I assume I need to hit the Train button for the model to use it (although the Successfully classified notification makes me feel like I don't, which is how it works if I'm in the facial recognition training). If I do it for 2 different models, do I need to hit the Train Model button on both models? Or does it train them all? Also, while I am getting the training successfully started and training successful notification, if I miss the notification, I seem to have no way to know if it was successful. Having a model status (Completed 11/9/2025 with 450 images, In Progress, Failed, Outdated (10 untrained photos) would be much appreciated. Edit: I guess it does spin while its training, not sure how I didn't notice that the first go around, but this time, I never got a successul training message in my notification, and it stopped, and I found this in my logs: unknown | 2025-11-09 10:05:19 | unknown | WARNING: All log messages before absl::InitializeLog() is called are written to STDERR |
Beta Was this translation helpful? Give feedback.
-
|
It would be quite complicated to have an active status. We can probably update the message to say "train the model to use the new images" or something, but I wouldn't want to keep track of how many images have been added since the the last training. Yes, you need to train the models separately. Face recognition was much simpler than this. |
Beta Was this translation helpful? Give feedback.
-
|
Even if it doesn't have a count, it probably is simple to have a flag if its currently up to date, that gets deleted as soon as an image gets classified after the training is requested? Perhaps spinning circle for currently training, checkmark for all images are currently trained, and active button for images need training? |
Beta Was this translation helpful? Give feedback.
-
|
the next version will have some improvements such as disabling the train button when no new images have been added, showing how many new images will be trained, and error handling in the logs. |
Beta Was this translation helpful? Give feedback.
-
|
Thought I would give 0.17 a try, but I think my old Core2Quad Q6600 is missing an opcode instruction causing it to not start (probably avx or something), this pc is dedicated to just frigate, it's old but works :X (stable works fine). frigate | 2025-11-09 08:13:11.050331985 [INFO] Service Frigate exited with code 256 (by signal 4) |
Beta Was this translation helpful? Give feedback.
-
|
That's the stable commit which was merged back into dev. You have to look at the master branch, that is what 0.16 is based on Your logs look to me like an ipv6 issue related to go2rtc |
Beta Was this translation helpful? Give feedback.
-
it's something related to this, every time it stops and restarts I get one of these in my dmesg logs, so what ever uses libctranslate2 is triggering the crash (perhaps the stable didn't try to start whatever is using libctranslate2) [1904910.909051] traps: python3[4166198] trap invalid opcode ip:7f38bba3954c sp:7ffc2cff2618 error:0 in libctranslate2-c75cca97.so.4.6.0[7f38bb8ae000+2dc5000] |
Beta Was this translation helpful? Give feedback.
-
|
Okay so it is indeed an Illegal Instruction crash. It is caused the |
Beta Was this translation helpful? Give feedback.
-
Perfect, I just tested build e4eac4a, and now it's working again, I haven't gone through all the features but so far the core program seems to be working again. I did see one error in the log on startup, not sure if it's important or not, and it didn't cause any crashes. frigate | 2025-11-11 06:30:02.241020132 Fatal Python error: Illegal instruction Edit: |
Beta Was this translation helpful? Give feedback.
-
btw is tensorflow required for the object creation / classification? I was thinking about building a custom non-avx / sse / non-gpu version of tensorflow and custom installing it into the frigate docker image, not sure if it will work but trying out different ideas :) |
Beta Was this translation helpful? Give feedback.
-
|
last build d41ee4f both face and classifications freezing too Much i have to restart every 5-10 min to set it work again,, |
Beta Was this translation helpful? Give feedback.
-
|
What CPU are you running? |
Beta Was this translation helpful? Give feedback.
-
|
I found the bug here |
Beta Was this translation helpful? Give feedback.
-
Intel I5 12gen its was ok in previous versions like still its a issue every 1-2hr ones but current version becomes too much its just freezing just 2-5min |
Beta Was this translation helpful? Give feedback.
-
|
Hi, I’ve tested it and it’s working fine — no freezes or crashes. (99a363c) However, I noticed a few issues: Training Button Behavior. I have a "front door open/close" state with 10 total metadata entries. After training, everything looks good and the "Train" button hides as expected. and one more i issue i found when creating new state class after selection all i can't process final complate button for some reason i need to go back to older bulid nor change config mannuly not happing by UI 
|
Beta Was this translation helpful? Give feedback.
-
You have to select all images for one state or another, you can't leave some images unselected |
Beta Was this translation helpful? Give feedback.
-
|
i really love the feature that progressively adds the tracked path points during review. questions/comments:
heres a screenshot for my item 3.2: |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
ref version: new roles seems to be a lot better. only issue is that cameras have to be defined individually since if you have something like 20 viewer groups, the config file can get excessively long. it would clean up and make it more manageable auth: camera_groups: instead of: auth: |
Beta Was this translation helpful? Give feedback.
-
|
Glad to hear it's working well for you. Frigate will be moving toward configuring more from the UI in future versions (YAML config will always be available). If config file management is difficult, the recommendation is to use the UI for configuring complex installations. |
Beta Was this translation helpful? Give feedback.
-
|
In the "Classification" section, it is not clear how to rename an already created object, just as after trying to delete an object, for some reason it is not deleted, but remains in the list. |
Beta Was this translation helpful? Give feedback.
-
|
You use the selector at the top to rename a class. For the deletion I think more steps are needed, I'm not seeing any issues deleting. |
Beta Was this translation helpful? Give feedback.
-
|
There is a delete button in the web interface, but nothing happens when you click on it. |
Beta Was this translation helpful? Give feedback.
-
|
Will be fixed in the next build |
Beta Was this translation helpful? Give feedback.
-
|
Hi guys. I've just managed to get d41ee4f installed and running. I've done a fresh install as making a copy of my "production" v16 environment is tricky given the volume of data I would need to duplicate (11TB) so... I've got 17 running in parallel using a cut-down configuration (2 cameras instead of 7). I've not gathered enough data to play with classifications yet but it did occur to me that I am probably well placed to do some comparative performance analysis on object detection and inference speeds. My question is if there is a way of extracting/saving performance data from v16 and the current dev build for offline comparison? |
Beta Was this translation helpful? Give feedback.
-
|
You can pull I don't know if full analysis is necessary though, at least for Nvidia the differences should be quite apparent. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks Nick. That’s exactly what I need. |
Beta Was this translation helpful? Give feedback.
-
|
So the rough cut is that the detection definitely is putting less strain on the CPU now. Here are the system metrics from 0.16 and 0.17 (d41ee4f). For reference, running on an Intel 10900K with an Nvidia RTX 2070 gpu.
|
Beta Was this translation helpful? Give feedback.
-
|
Great to see. This is due to the CUDA improvements in 0.17. |
Beta Was this translation helpful? Give feedback.
-
|
Has the camera_context configuration been removed? Just updated to 99a363c and now getting invalid config errors: "review -> genai -> camera_context - Extra inputs are not permitted" |
Beta Was this translation helpful? Give feedback.
-
|
yes, it was removed. In my thorough testing it only made the GenAI more confused about direction of travel as the context confused it. After a lot of variations I don't think it makes sense to recommend users use that approach. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the speedy response. I'd had some issues with the direction of travel being incorrect, but hadn't spent any time refining the prompts yet. Finding the Review Item descriptions really useful - helped simplify my notification configurations & reduces the number of GenAI calls required. No issues identified so far! |
Beta Was this translation helpful? Give feedback.
-
|
Thanks, glad to hear it is working well for you |
Beta Was this translation helpful? Give feedback.
-
|
Does Review Item Descriptions pass structured data for zones and camera names? Currently, GenAI keeps misidentifying my yard as a balcony and also struggles to properly distinguish public areas. |
Beta Was this translation helpful? Give feedback.
-
|
It does, though currently it does not use the friendly names so that might be the issue |
Beta Was this translation helpful? Give feedback.
-
I believe the camera_context is still necessary, but I think we could approach it differently—by adding a "description" field to the zones. This field could serve both as frontend display content and as context for GenAI. |
Beta Was this translation helpful? Give feedback.
-
|
Like I said, I spent a lot of time with this working with many different approaches with this, and it did not help, it actively made things worse. You are free to adjust the activity_context_prompt on a per camera basis if you want to change what is sent in the prompt for each camera. |
Beta Was this translation helpful? Give feedback.
-
|
The next build will use the friendly names in the prompt. For anyone looking closer at genai review feature I would highly suggest enabling |
Beta Was this translation helpful? Give feedback.
-
|
There is definately something not right with face after build 88a2f6c. as soon as I move on from that commit faces are comming up as unknown. if i try and reprocess them i get the following. If i roll back to 88a2f6c they reprocess. 
|
Beta Was this translation helpful? Give feedback.
-
|
We've found the issue, looking into a fix |
Beta Was this translation helpful? Give feedback.
-
|
Should be fixed in the latest build. |
Beta Was this translation helpful? Give feedback.
-
|
Lastest build is working, thanks. |
Beta Was this translation helpful? Give feedback.
-
|
There is something weird still going on with the reprocess function. if you add a face and then try to reprocess the image next to it it says no faces were found until there is a new image added to the face page then i reprocesses ok. |
Beta Was this translation helpful? Give feedback.
-
|
The reason is because now the reprocess function returns correctly. Training the face model takes time so if you try to reprocess when it isn't up to date it will give that error. I will update the error to be more clear |
Beta Was this translation helpful? Give feedback.
-
|
So as a quick update on my usage of classification models: its working fairly well. I'm randomly getting errors when trying to train a new model that its failing to find images, but if I hit try again, it works. Its done in multiple times now. One other issue I'm running into now is one of the states that I thought would be nice to have (position of a manual gate) required repositioning the camera (you could see it before, but it was really tough, even for a human, and I couldn't get a classification model trained on it, I moved the camera a slightly to make it much easier), but now when I'm trying to train the model, its obviously pulling tons of images from before the repositioning. Is there a recommended way to handle camera repositioning? |
Beta Was this translation helpful? Give feedback.
-
very odd, I have never seen that, will have to see what might be going on
I would delete the images from the previous angle |
Beta Was this translation helpful? Give feedback.
-
|
Feedback on the new Review interface In short, love it. The Event/Detail views are super useful but the killer feature for me is the Object Path tracking. I think that's going to help a lot of users understand what's happening under the hood when it comes to object detection and debug configuration issues. |
Beta Was this translation helpful? Give feedback.
-
|
That's great. Glad it's working well for you. |
Beta Was this translation helpful? Give feedback.
-
|
I'm testing the tab with the "Classification" of objects and for some reason I can't train the model, when I click on the button, an endless download appears, and various errors appear in the logs from version to version. There is a version where learning takes place without problems (01452e4), but for some reason this functionality does not work on the last two versions. 
b"[concat @ 0x55d3557cbd80] Impossible to open '/tmp/cache/preview_frames/preview_Entry-1762930288.004772.webp'\n[in#0 @ 0x55d3557cba40] Error opening input: Invalid data found when processing input\nError opening input file /dev/stdin.\nError opening input files: Invalid data found when processing input\n"
|
Beta Was this translation helpful? Give feedback.
-
|
While viewing specific tags on the Explorer page and attempting to submit some snapshots to Frigate+, I encountered an issue: after submitting to Frigate+, if I close the Tracked Object Details page and scroll down, it enters an infinite loop, continuously requesting 
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks, this will be fixed in the next build. |
Beta Was this translation helpful? Give feedback.
and others
Uh oh!
There was an error while loading. Please reload this page.
-
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 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 functionality then please give that feedback as well.
The docs are: https://deploy-preview-19787--frigate-docs.netlify.app/
Object Detectors
Nvidia: CUDA Graphs are now utilized to greatly reduce the involvement of the CPU when running inferences. This is done automatically with no config changes required. This should lead to lower minimum inference speeds and dramatically lower maximum inference speeds for Nvidia GPUs. This primarily affects object detection, but will help some other models such as face recognition. Do note that complex models such as YOLO-NAS and Semantic Search will not be supported / accelerated by CUDA Graphs.OpenVINO GPU: More subtle than Nvidia, but optimizations have been made to all OpenVINO inferences that should reduce CPU usage, memory contention (fewer number of memory transfers), and potentially lower memory usage as well.OpenVINO NPU: OpenVINO now supports NPU. Simply setNPUas the device on a supported platform and make sure/dev/accelis mapped into the container. Would be great to see inference speed examples, all model types that GPU supports should be supported by the NPU as well.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 docsClassification Models
Frigate now has end-to-end support, including training, for classification models. These come in two types,
stateclassification (ex: gate open / close) andobjectclassification (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
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/
Authentication / Roles
main-user) to one of Frigate's supported roles (admin) see the docsBeta Was this translation helpful? Give feedback.
All reactions