mitigate tensorflow atexit crash by pre-importing tflite/tensorflow on main thread

Pre-import Interpreter in embeddings maintainer and add defensive lazy imports in classification processors to avoid worker-thread tensorflow imports causing "can't register atexit after shutdown"

frigate/data_processing/real_time/bird.py

@@ -19,11 +19,6 @@ from frigate.util.object import calculate_region
 from ..types import DataProcessorMetrics
 from .api import RealTimeProcessorApi
 
-try:
-    from tflite_runtime.interpreter import Interpreter
-except ModuleNotFoundError:
-    from tensorflow.lite.python.interpreter import Interpreter
-
 logger = logging.getLogger(__name__)
 
 
@@ -35,7 +30,7 @@ class BirdRealTimeProcessor(RealTimeProcessorApi):
         metrics: DataProcessorMetrics,
     ):
         super().__init__(config, metrics)
-        self.interpreter: Interpreter = None
+        self.interpreter: Any | None = None
         self.sub_label_publisher = sub_label_publisher
         self.tensor_input_details: dict[str, Any] = None
         self.tensor_output_details: dict[str, Any] = None
@@ -82,6 +77,11 @@ class BirdRealTimeProcessor(RealTimeProcessorApi):
 
     @redirect_output_to_logger(logger, logging.DEBUG)
     def __build_detector(self) -> None:
+        try:
+            from tflite_runtime.interpreter import Interpreter
+        except ModuleNotFoundError:
+            from tensorflow.lite.python.interpreter import Interpreter
+
         self.interpreter = Interpreter(
             model_path=os.path.join(MODEL_CACHE_DIR, "bird/bird.tflite"),
             num_threads=2,

frigate/data_processing/real_time/custom_classification.py

@@ -29,11 +29,6 @@ from frigate.util.object import box_overlaps, calculate_region
 from ..types import DataProcessorMetrics
 from .api import RealTimeProcessorApi
 
-try:
-    from tflite_runtime.interpreter import Interpreter
-except ModuleNotFoundError:
-    from tensorflow.lite.python.interpreter import Interpreter
-
 logger = logging.getLogger(__name__)
 
 MAX_OBJECT_CLASSIFICATIONS = 16
@@ -52,7 +47,7 @@ class CustomStateClassificationProcessor(RealTimeProcessorApi):
         self.requestor = requestor
         self.model_dir = os.path.join(MODEL_CACHE_DIR, self.model_config.name)
         self.train_dir = os.path.join(CLIPS_DIR, self.model_config.name, "train")
-        self.interpreter: Interpreter | None = None
+        self.interpreter: Any | None = None
         self.tensor_input_details: dict[str, Any] | None = None
         self.tensor_output_details: dict[str, Any] | None = None
         self.labelmap: dict[int, str] = {}
@@ -74,6 +69,11 @@ class CustomStateClassificationProcessor(RealTimeProcessorApi):
 
     @redirect_output_to_logger(logger, logging.DEBUG)
     def __build_detector(self) -> None:
+        try:
+            from tflite_runtime.interpreter import Interpreter
+        except ModuleNotFoundError:
+            from tensorflow.lite.python.interpreter import Interpreter
+
         model_path = os.path.join(self.model_dir, "model.tflite")
         labelmap_path = os.path.join(self.model_dir, "labelmap.txt")
 
@@ -345,7 +345,7 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
         self.model_config = model_config
         self.model_dir = os.path.join(MODEL_CACHE_DIR, self.model_config.name)
         self.train_dir = os.path.join(CLIPS_DIR, self.model_config.name, "train")
-        self.interpreter: Interpreter | None = None
+        self.interpreter: Any | None = None
         self.sub_label_publisher = sub_label_publisher
         self.requestor = requestor
         self.tensor_input_details: dict[str, Any] | None = None
@@ -368,6 +368,11 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
 
     @redirect_output_to_logger(logger, logging.DEBUG)
     def __build_detector(self) -> None:
+        try:
+            from tflite_runtime.interpreter import Interpreter
+        except ModuleNotFoundError:
+            from tensorflow.lite.python.interpreter import Interpreter
+
         model_path = os.path.join(self.model_dir, "model.tflite")
         labelmap_path = os.path.join(self.model_dir, "labelmap.txt")
 

frigate/embeddings/maintainer.py

@@ -146,6 +146,29 @@ class EmbeddingMaintainer(threading.Thread):
         self.detected_license_plates: dict[str, dict[str, Any]] = {}
         self.genai_client = get_genai_client(config)
 
+        # Pre-import TensorFlow/tflite on main thread to avoid atexit registration issues
+        # when importing from worker threads later (e.g., during dynamic config updates)
+        if (
+            self.config.classification.bird.enabled
+            or len(self.config.classification.custom) > 0
+        ):
+            try:
+                from tflite_runtime.interpreter import Interpreter  # noqa: F401
+            except ModuleNotFoundError:
+                try:
+                    from tensorflow.lite.python.interpreter import (  # noqa: F401
+                        Interpreter,
+                    )
+
+                    logger.debug(
+                        "Pre-imported TensorFlow Interpreter on main thread for classification models"
+                    )
+                except Exception as e:
+                    logger.warning(
+                        f"Failed to pre-import TensorFlow Interpreter: {e}. "
+                        "Classification models may fail to load if added dynamically."
+                    )
+
         # model runners to share between realtime and post processors
         if self.config.lpr.enabled:
             lpr_model_runner = LicensePlateModelRunner(

web/src/components/classification/wizard/Step3ChooseExamples.tsx

@@ -141,7 +141,37 @@ export default function Step3ChooseExamples({
       );
       await Promise.all(categorizePromises);
 
-      // Step 2.5: Create empty folders for classes that don't have any images
+      // Step 2.5: Delete any unselected images from train folder
+      // For state models, all images must be classified, so unselected images should be removed
+      // For object models, unselected images are assigned to "none" so they're already categorized
+      if (step1Data.modelType === "state") {
+        try {
+          // Fetch current train images to see what's left after categorization
+          const trainImagesResponse = await axios.get<string[]>(
+            `/classification/${step1Data.modelName}/train`,
+          );
+          const remainingTrainImages = trainImagesResponse.data || [];
+
+          const categorizedImageNames = new Set(Object.keys(classifications));
+          const unselectedImages = remainingTrainImages.filter(
+            (imageName) => !categorizedImageNames.has(imageName),
+          );
+
+          if (unselectedImages.length > 0) {
+            await axios.post(
+              `/classification/${step1Data.modelName}/train/delete`,
+              {
+                ids: unselectedImages,
+              },
+            );
+          }
+        } catch (error) {
+          // Silently fail - unselected images will remain but won't cause issues
+          // since the frontend filters out images that don't match expected format
+        }
+      }
+
+      // Step 2.6: Create empty folders for classes that don't have any images
       // This ensures all classes are available in the dataset view later
       const classesWithImages = new Set(
         Object.values(classifications).filter((c) => c && c !== "none"),

web/src/views/classification/ModelTrainingView.tsx

@@ -866,6 +866,12 @@ function TrainGrid({
           };
         })
         .filter((data) => {
+          // Ignore images that don't match the expected format (event-camera-timestamp-state-score.webp)
+          // Expected format has 5 parts when split by "-", and score should be a valid number
+          if (data.score === undefined || isNaN(data.score) || !data.name) {
+            return false;
+          }
+
           if (!trainFilter) {
             return true;
           }