suppress tensorflow logging during classification training

docker/main/Dockerfile

@@ -237,8 +237,18 @@ ENV PYTHONWARNINGS="ignore:::numpy.core.getlimits"
 # Set HailoRT to disable logging
 ENV HAILORT_LOGGER_PATH=NONE
 
-# TensorFlow error only
+# TensorFlow C++ logging suppression (must be set before import)
+# TF_CPP_MIN_LOG_LEVEL: 0=all, 1=INFO+, 2=WARNING+, 3=ERROR+ (we use 3 for errors only)
 ENV TF_CPP_MIN_LOG_LEVEL=3
+# Suppress verbose logging from TensorFlow C++ code
+ENV TF_CPP_MIN_VLOG_LEVEL=3
+# Disable oneDNN optimization messages ("optimized with oneDNN...")
+ENV TF_ENABLE_ONEDNN_OPTS=0
+# Suppress AutoGraph verbosity during conversion
+ENV AUTOGRAPH_VERBOSITY=0
+# Google Logging (GLOG) suppression for TensorFlow components
+ENV GLOG_minloglevel=3
+ENV GLOG_logtostderr=0
 
 ENV PATH="/usr/local/go2rtc/bin:/usr/local/tempio/bin:/usr/local/nginx/sbin:${PATH}"
 

frigate/log.py

@@ -80,10 +80,15 @@ def apply_log_levels(default: str, log_levels: dict[str, LogLevel]) -> None:
     log_levels = {
         "absl": LogLevel.error,
         "httpx": LogLevel.error,
+        "h5py": LogLevel.error,
+        "keras": LogLevel.error,
         "matplotlib": LogLevel.error,
         "tensorflow": LogLevel.error,
+        "tensorflow.python": LogLevel.error,
         "werkzeug": LogLevel.error,
         "ws4py": LogLevel.error,
+        "PIL": LogLevel.warning,
+        "numba": LogLevel.warning,
         **log_levels,
     }
 
@@ -318,3 +323,31 @@ def suppress_os_output(func: Callable) -> Callable:
         return result
 
     return wrapper
+
+
+@contextmanager
+def suppress_stderr_during(operation_name: str) -> Generator[None, None, None]:
+    """
+    Context manager to suppress stderr output during a specific operation.
+
+    Useful for silencing LLVM debug output, CUDA messages, and other native
+    library logging that cannot be controlled via Python logging or environment
+    variables. Completely redirects file descriptor 2 (stderr) to /dev/null.
+
+    Usage:
+        with suppress_stderr_during("model_conversion"):
+            converter = tf.lite.TFLiteConverter.from_keras_model(model)
+            tflite_model = converter.convert()
+
+    Args:
+        operation_name: Name of the operation for debugging purposes
+    """
+    original_stderr_fd = os.dup(2)
+    devnull = os.open(os.devnull, os.O_WRONLY)
+    try:
+        os.dup2(devnull, 2)
+        yield
+    finally:
+        os.dup2(original_stderr_fd, 2)
+        os.close(devnull)
+        os.close(original_stderr_fd)

frigate/util/classification.py

@@ -19,7 +19,7 @@ from frigate.const import (
     PROCESS_PRIORITY_LOW,
     UPDATE_MODEL_STATE,
 )
-from frigate.log import redirect_output_to_logger
+from frigate.log import redirect_output_to_logger, suppress_stderr_during
 from frigate.models import Event, Recordings, ReviewSegment
 from frigate.types import ModelStatusTypesEnum
 from frigate.util.downloader import ModelDownloader
@@ -250,15 +250,20 @@ class ClassificationTrainingProcess(FrigateProcess):
             logger.debug(f"Converting {self.model_name} to TFLite...")
 
             # convert model to tflite
-            converter = tf.lite.TFLiteConverter.from_keras_model(model)
+            # Suppress stderr during conversion to avoid LLVM debug output
-            converter.optimizations = [tf.lite.Optimize.DEFAULT]
+            # (fully_quantize, inference_type, MLIR optimization messages, etc)
-            converter.representative_dataset = (
+            with suppress_stderr_during("tflite_conversion"):
-                self.__generate_representative_dataset_factory(dataset_dir)
+                converter = tf.lite.TFLiteConverter.from_keras_model(model)
-            )
+                converter.optimizations = [tf.lite.Optimize.DEFAULT]
-            converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+                converter.representative_dataset = (
-            converter.inference_input_type = tf.uint8
+                    self.__generate_representative_dataset_factory(dataset_dir)
-            converter.inference_output_type = tf.uint8
+                )
-            tflite_model = converter.convert()
+                converter.target_spec.supported_ops = [
+                    tf.lite.OpsSet.TFLITE_BUILTINS_INT8
+                ]
+                converter.inference_input_type = tf.uint8
+                converter.inference_output_type = tf.uint8
+                tflite_model = converter.convert()
 
             # write model
             model_path = os.path.join(model_dir, "model.tflite")