feat: add train classification download weights file endpoint: "TF_KERAS_MOBILENET_V2_ENDPOINT"

frigate/util/classification.py

@@ -22,6 +22,7 @@ from frigate.const import (
 from frigate.log import redirect_output_to_logger
 from frigate.models import Event, Recordings, ReviewSegment
 from frigate.types import ModelStatusTypesEnum
+from frigate.util.downloader import ModelDownloader
 from frigate.util.file import get_event_thumbnail_bytes
 from frigate.util.image import get_image_from_recording
 from frigate.util.process import FrigateProcess
@@ -121,6 +122,10 @@ def get_dataset_image_count(model_name: str) -> int:
 
 class ClassificationTrainingProcess(FrigateProcess):
     def __init__(self, model_name: str) -> None:
+        self.BASE_WEIGHT_PATH = os.environ.get(
+            "TF_KERAS_MOBILENET_V2_ENDPOINT",
+            "https://storage.googleapis.com/tensorflow/keras-applications/mobilenet_v2/",
+        )
         super().__init__(
             stop_event=None,
             priority=PROCESS_PRIORITY_LOW,
@@ -179,12 +184,23 @@ class ClassificationTrainingProcess(FrigateProcess):
                 )
                 return False
 
+            alpha = 0.35
+            # Download MobileNetV2 weights if not present
+            weights_filename = (
+                f"mobilenet_v2_weights_tf_dim_ordering_tf_kernels_{alpha}_224_no_top.h5"
+            )
+            weights_path = os.path.join(MODEL_CACHE_DIR, "MobileNet", weights_filename)
+            if not os.path.exists(weights_path):
+                weights_url = self.BASE_WEIGHT_PATH + weights_filename
+                logger.info(f"Downloading MobileNet V2 weights file: {weights_url}")
+                ModelDownloader.download_from_url(weights_url, weights_path)
+
             # Start with imagenet base model with 35% of channels in each layer
             base_model = MobileNetV2(
                 input_shape=(224, 224, 3),
                 include_top=False,
-                weights="imagenet",
-                alpha=0.35,
+                weights=weights_path,
+                alpha=alpha,
             )
             base_model.trainable = False  # Freeze pre-trained layers