* Ignore numpy get limits warning
* Add function wrapper to redirect stdout and stderr to logpipe
* Save stderr too
* Add more to catch
* run logpipe
* Use other logging redirect class
* Use other logging redirect class
* add decorator for redirecting c/c++ level output to logger
* fix typing
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
Multiprocessing Manager queues don't have a close() or join_thread() method, and the Manager will clean it up appropriately after we empty it. This prevents an infinite loop when an AttributeError exception fires for Manager AutoProxy queue objects.
* Move log level initialization to log
* Use logger config
* Formatting
* Fix config order
* Set process names
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Set runtime
* Use count correctly
* Don't assume camera sizes
* Use separate zmq proxy for object detection
* Correct order
* Use forkserver
* Only store PID instead of entire process reference
* Cleanup
* Catch correct errors
* Fix typing
* Remove before_run from process util
The before_run never actually ran because:
You're right to suspect an issue with before_run not being called and a potential deadlock. The way you've implemented the run_wrapper using __getattribute__ for the run method of BaseProcess is a common pitfall in Python's multiprocessing, especially when combined with how multiprocessing.Process works internally.
Here's a breakdown of why before_run isn't being called and why you might be experiencing a deadlock:
The Problem: __getattribute__ and Process Serialization
When you create a multiprocessing.Process object and call start(), the multiprocessing module needs to serialize the process object (or at least enough of it to re-create the process in the new interpreter). It then pickles this serialized object and sends it to the newly spawned process.
The issue with your __getattribute__ implementation for run is that:
run is retrieved during serialization: When multiprocessing tries to pickle your Process object to send to the new process, it will likely access the run attribute. This triggers your __getattribute__ wrapper, which then tries to bind run_wrapper to self.
run_wrapper is bound to the parent process's self: The run_wrapper closure, when created in the parent process, captures the self (the Process instance) from the parent's memory space.
Deserialization creates a new object: In the child process, a new Process object is created by deserializing the pickled data. However, the run_wrapper method that was pickled still holds a reference to the self from the parent process. This is a subtle but critical distinction.
The child's run is not your wrapped run: When the child process starts, it internally calls its own run method. Because of the serialization and deserialization process, the run method that's ultimately executed in the child process is the original multiprocessing.Process.run or the Process.run if you had directly overridden it. Your __getattribute__ magic, which wraps run, isn't correctly applied to the Process object within the child's context.
* Cleanup
* Logging bugfix (#18465)
* use mp Manager to handle logging queues
A Python bug (https://github.com/python/cpython/issues/91555) was preventing logs from the embeddings maintainer process from printing. The bug is fixed in Python 3.14, but a viable workaround is to use the multiprocessing Manager, which better manages mp queues and causes the logging to work correctly.
* consolidate
* fix typing
* Fix typing
* Use global log queue
* Move to using process for logging
* Convert camera tracking to process
* Add more processes
* Finalize process
* Cleanup
* Cleanup typing
* Formatting
* Remove daemon
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* Add base class for global config updates
* Add or remove camera states
* Move camera process management to separate thread
* Move camera management fully to separate class
* Cleanup
* Stop camera processes when stop command is sent
* Start processes dynamically when needed
* Adjust
* Leave extra room in tracked object queue for two cameras
* Dynamically set extra config pieces
* Add some TODOs
* Fix type check
* Simplify config updates
* Improve typing
* Correctly handle indexed entries
* Cleanup
* Create out SHM
* Use ZMQ for signaling object detectoin is completed
* Get camera correctly created
* Cleanup for updating the cameras config
* Cleanup
* Don't enable audio if no cameras have audio transcription
* Use exact string so similar camera names don't interfere
* Add ability to update config via json body to config/set endpoint
Additionally, update the config in a single rather than multiple calls for each updated key
* fix autotracking calibration to support new config updater function
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* Combine base and arm trt detectors
* Remove unused deps for amd64 build
* Add missing packages and cleanup ldconfig
* Expand packages for tensorflow model training
* Cleanup
* Refactor training to not reserve memory
* Refactor common functions for tflite detector implementations
* Add detector using mesa teflon delegate
Non-EdgeTPU TFLite can use the standard .tflite format
* Add mesa-teflon-delegate from bookworm-backports to arm64 images
* Implement model training via ZMQ and add model states to represent training
* Get model updates working
* Improve toasts and model state
* Clean up logging
* Add back in
* Setup basic training structure
* Build out route
* Handle model configs
* Add image fetch APIs
* Implement model training screen with dataset selection
* Implement viewing of training images
* Adjust directories
* Implement viewing of images
* Add support for deleting images
* Implement full deletion
* Implement classification model training
* Improve naming
* More renaming
* Improve layout
* Reduce logging
* Cleanup
The PP_OCRv5 text detection models have greatly improved over v3. The v5 recognition model makes improvements to challenging handwriting and uncommon characters, which are not necessary for LPR, so using v4 seemed like a better choice to continue to keep inference time as low as possible. Also included is the full dictionary for Chinese character support.
* install new packages for transcription support
* add config options
* audio maintainer modifications to support transcription
* pass main config to audio process
* embeddings support
* api and transcription post processor
* embeddings maintainer support for post processor
* live audio transcription with sherpa and faster-whisper
* update dispatcher with live transcription topic
* frontend websocket
* frontend live transcription
* frontend changes for speech events
* i18n changes
* docs
* mqtt docs
* fix linter
* use float16 and small model on gpu for real-time
* fix return value and use requestor to embed description instead of passing embeddings
* run real-time transcription in its own thread
* tweaks
* publish live transcriptions on their own topic instead of tracked_object_update
* config validator and docs
* clarify docs
* Start Frigate in safe mode when config does not validate
* Add safe mode page that is just the config editor
* Adjust Frigate config editor when in safe mode
* Cleanup
* Improve log message
* Indicate no recordings on the history timeline with gray hash marks
This commit includes a new backend API endpoint and the frontend changes needed to support this functionality
* don't show slashes for now
* Add basic config for defining a teachable machine model
* Add model type
* Add basic config for teachable machine models
* Adjust config for state and object
* Use config to process
* Correctly check for objects
* Remove debug
* Rename to not be teachable machine specific
* Cleanup
* Include config publisher in api
* Call update topic for passed topics
* Update zones dynamically
* Update zones internally
* Support zone and mask reset
* Handle updating objects config
* Don't put status for needing to restart Frigate
* Cleanup http tests
* Fix tests
* Fetch all license plates outside of filter component
If the swr call took a long time, the entire select component may not display. This change moves the fetch to the parent component (like sub labels).
* add loading indicator
* improve query
* Remove torch install
* notification fixes
the pubkey was not being returned if notifications was not enabled at the global level
* Put back
* single condition check for fetching and disabling button
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* fix embeddings reindex
- always increment processed objects to prevent division by zero
- ensure description still gets processed even if there is no thumbnail
* clean up
* Add newer labels to default attribute map
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
When an IPv6 address that doesn't map to an IPv4 address was checked
against an IPv4 trusted proxy, we'd hit an exception because
ip.ipv4_mapped was None. Fix this by verifying ipv4_mapped is not None
Co-authored-by: me <me@me>
* Fix Norwegian language code for date-fns module import
* Fix incorrect apple compatibility change
* Add other and unused as separate storage items
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Don't support tensorrt detector for amd64 builds
* Add logs for directing users not to use tensorrt detector
* Rework docs
* Fix dockerfile index
* Don't undo jetson fix
* Fix showing review items that span over multiple days
* Simplify
* Fix tests
* Fix unchanged value
* Allow admin as default role and viewer as passed header for proxy auth
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* i18n translated label fixes
* Fix frame cache race bug
Objects that were marked as false positives (that would later become true positives) would sometimes have their saved frame prematurely removed from the frame cache.
* Object labels with spaces should use correct i18n keys
* Add Hungarian
* Ensure onvif move request has a valid speed before removing
When autotracking zooming is set to `disabled` (or is left out of the config), move_request["Speed"] may not exist, depending on the camera
* Add another frame cache debug log
* Don't allow editing of sub label until object lifecycle has ended
* Update sub labels in ended review segments
When manually editing a sub label for a tracked object from the UI, any review segments containing that tracked object did not have their sub_labels and objects values altered
* simplify
* Additional onvif debug logs in get_camera_status
* Ensure that best object is only set when the snapshot is actually updated.
* Don't hide downlaod button when there is no review item
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
Model can be adjusted ahead of time to NHWC to avoid transpose on CPU,
for example. All the model information is already presented in the
configuration, and the stringent checks implemented in openvino are not
present on other backends anyway. OpenVINO will properly report issues
with mismatched layouts anyhow.
* Ensure config editor recalculates layout on error
* ensure empty lists are returned when lpr recognition model fails
* Add docs section for session_length
* clarify
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* clarify
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* Catch missing file
* Improve graph axis colors
* Ensure playback rate controls are portaled to the video container in history view
On larger tablets in landscape view, the playback rate dropdown disappeared underneath the bottom bar. This small change ensures we use the correct container on the DropdownMenuContent so that the div is portaled correctly. The VideoControls are also used in motion review which does not pass in a container ref, so we can just fall back to the existing controlsContainer ref when it's undefined.
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* fix i18n keys
* hide disable from context menu for viewers
* Fix auto live check for default dashboard and camera groups
Disabling the Automatic Live View switch in Settings should prevent streaming from occurring. Overriding any settings in a camera group will override the global setting. The check here incorrectly always returned false instead of undefined.
* clarify hardware accelerated enrichments
* clarify
* add note about detect stream to face rec docs
* add note about low end Dahuas for autotracking
* Catch invalid face box / image
* Video tab tweaks
With the changes in https://github.com/blakeblackshear/frigate/pull/18220, the video tab in the Tracked Object Details pane now correctly trims the in-browser HLS video. Because of keyframes and record/detect stream differences, we can manually subtract a couple of seconds from the event start_time to ensure the first few frames aren't cut off from the video
* Clarify
* Don't use Migraphx by default
* Provide better support for running embeddings on GPU
* correctly join cameras
* Adjust blur confidence reduction
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* use mp Manager to handle logging queues
A Python bug (https://github.com/python/cpython/issues/91555) was preventing logs from the embeddings maintainer process from printing. The bug is fixed in Python 3.14, but a viable workaround is to use the multiprocessing Manager, which better manages mp queues and causes the logging to work correctly.
* consolidate
* fix typing
* Don't return weighted name if it has the same number of results
* Remove link to incorrect format yolov9 models
* Fix command list from appearing when other inputs are focused
the description box in the tracked object details pane was causing the command input list to show when focused.
* clarify face docs
* Add note about python yolov9 export
* Check if hailort thread is still alive when timeout error is run into
* Reduce inference timeout
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* improve spacing of face selection in mobile drawer
* fix spacing
* sort face names alphabetically
* Improve face selection dialog
* Use a state to track when face image loads
The naturalWidth and naturalHeight will always be 0 until the image loads. So we use onLoad and a state to track loading and then calculate the area after it has loaded
* Verify that a camera only tracks objects that are possible to track
* Fix test
* genai docs tweak
* Disable openvino model cache
* Clenaup
* Sanitize floats for estimated speed and angle
Users can configure speed zones in such a way that velocity estimates from Norfair cause a value of inf to be stored as an estimated speed. FastAPI doesn't serialize inf as a float, so trying to return this value would result in an API error. Sanitizing the value before storing should correct this.
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Add Thai (still need to merge weblate)
* Apply attribute logic to all label types
* Fix area check
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Don't use timezone in export dialog timestamps
Revert an unnecessary change made in https://github.com/blakeblackshear/frigate/pull/18257
* Ensure notifications register button is only disabled when both all cameras and every individual camera is disabled
* Send test notification if any cameras are enabled
* clarify docs about disabling cameras
* fix crash in autotracking zoom
* clean up
* masks and zones i18n fixes
* Check if camera is enabled in config
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Only check if an object is stationary to avoid mqtt snapshot
* docs heading tweak
* Add more API descriptions
* Add missing lib for new rocm onnxruntime whl
* Update inference times to reflect better rocm performance
* Cleanup resetting tracked object activity
* remove print
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
Not everyone is using /dev/dri/card0. I recommend switching to these more generic parameters for intel_gpu_top_command (line 307 in the 0.16.0-beta2 version of services.py):
intel_gpu_top_command += ["-d", "sriov"]
* Add option to not trim clip
* Improve API
* Update snapshot for new best objects
* Fix missing strings
* Convert to separate key
* Always include bounding box on snapshots
* improve autotracking relative zooming time calculation
* update proxy docs to note the need for comma separated header roles
* Add count translation
* tracked object lifecycle i18n fix
* update speed estimation docs
* clarity
* Re-initialize onvif information when toggling camera on live view
* Move time ago to card info and add face area
* Clarify face recognition docs
* Increase minimum face recognition area
* use clipFrom to in vod module endpoint to start at the correct time
* Cleanup media api
* Don't change duration
* Use search detail dialog for face library
* Move to segment based
* Cleanup
* Add back duration modification
* clean up docs
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* Fix the `Any` typing hint treewide
There has been confusion between the Any type[1] and the any function[2]
in typing hints.
[1] https://docs.python.org/3/library/typing.html#typing.Any
[2] https://docs.python.org/3/library/functions.html#any
* Fix typing for various frame_shape members
Frame shapes are most likely defined by height and width, so a single int
cannot express that.
* Wrap gpu stats functions in Optional[]
These can return `None`, so they need to be `Type | None`, which is what
`Optional` expresses very nicely.
* Fix return type in get_latest_segment_datetime
Returns a datetime object, not an integer.
* Make the return type of FrameManager.write optional
This is necessary since the SharedMemoryFrameManager.write function can
return None.
* Fix total_seconds() return type in get_tz_modifiers
The function returns a float, not an int.
https://docs.python.org/3/library/datetime.html#datetime.timedelta.total_seconds
* Account for floating point results in to_relative_box
Because the function uses division the return types may either be int or
float.
* Resolve ruff deprecation warning
The config has been split into formatter and linter, and the global
options are deprecated.
* Add camera name tooltip to previews in recording view
* Apply face area check to cv2 face detection
* Delete review thumbnails
* Don't import hailo until it is used
* Add comment
* Clean up camera name
* Filter out empty keys when updating yaml config
HA ingress seems to randomly add an equal sign to the PUT urls for updating the config from the UI. This fix prevents empty keys from being processed, but still allows empty values.
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Catch error and show toast when failing to delete review items
* i18n keys
* add link to speed estimation docs in zone edit pane
* Implement reset of tracked object update for each camera
* Cleanup
* register mqtt callbacks for toggling alerts and detections
* clarify snapshots docs
* clarify semantic search reindexing
* add ukrainian
* adjust date granularity for last recording time
The api endpoint only returns granularity down to the day
* Add amd hardware
* fix crash in face library on initial start after enabling
* Fix recordings view for mobile landscape
The events view incorrectly was displaying two columns on landscape view and it only took up 20% of the screen width. Additionally, in landscape view the timeline was too wide (especially on iPads of various screen sizes) and would overlap the main video
* face rec overfitting instructions
* Clarify
* face docs
* clarify
* clarify
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* face library i18n fixes
* face library i18n fixes
* add ability to use ctrl/cmd S to save in the config editor
* Use datetime as ID
* Update metrics inference speed to start with 0 ms
* fix android formatted thumbnail
* ensure role is comma separated and stripped correctly
* improve face library deletion
- add a confirmation dialog
- add ability to select all / delete faces in collections
* Implement lazy loading for video previews
* Force GPU for large embedding model
* GPU is required
* settings i18n fixes
* Don't delete train tab
* webpush debugging logs
* Fix incorrectly copying zones
* copy path data
* Ensure that cache dir exists for Frigate+
* face docs update
* Add description to upload image step to clarify the image
* Clean up
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* fix onvif reinitialization
* api docs: clarify usage of clip.mp4 endpoint
* Always show train tab
* Add description to API
* catch lpr model inference exceptions
* always apply motion mask when using yolov9 plate detection
* lpr faq
* fix incorrect focus when reopening search detail dialog on video tab
* only use keyboard listener in face library when train tab is active
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* use async/await instead of asyncio.run()
* fix autotracking
* create cameras in same event loop that will use them
* more debug
* try using existing event loop instead of creating a new one
* merge dev
* fixes
* run get_camera_info onvifcontroller calls in dedicated loop
* move coroutine call with loop to api
* use asyncio for autotracking move queues
* clean up
* fix calibration
* improve exception logging
* frigate+ pane i18n fix
* catch more exceptions
* explore search result tooltip i18n fix
* i18n fix
* remove comments about deprecated strftime_fmt
* Catch producers exists but is None
* Formatting
* fix live camera view i18n
* Add default role config for proxy users
This allows users to specify a default role for users when using a proxy for auth. This can be useful for users who can't/don't want to define a header mapping for the remote-role header.
* update reference config and auth docs
* clarify face rec camera level config
* clarify auth docs
* Fix onnx not working with openvino
* Update openvino to fix failed npu plugin check
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
On startup, the object processing pipeline would be delayed for 10ms every iteration through the loop while the end event subscriber blocked. For users with large numbers of cameras and a fair amount of detected objects, this would cause the detected objects queue length to rise and frames to be eventually dropped, leading to processing delays, incorrect frame times applied to thumbnail_data, and other seemingly unexplainable things.
* add zoom time to movement predictions
* config migrator
* add space to face rename regex
* more debug
* only calculate zoom time of relative move
* fix test
* make migrated movement weight a zero
* check for str and bool for movestatus support