feat: increase chunk size for face embedding processing to improve performance and increased the resolution for better accuracy

Author: Avneesh11905

inference_api/infrastructure/di_container.py

@@ -15,6 +15,7 @@ class Container(containers.DeclarativeContainer):
     face_detector = providers.Singleton(
         FaceDetector,
         model_path="models/buffalo_l/det_10g.onnx",
+        input_size=(1920, 1920),
         device="cuda",
         num_threads=1,
     )

inference_api/infrastructure/onnx/embedder.py

@@ -128,7 +128,7 @@ def embed_batch(self, aligned_faces: list[np.ndarray]) -> np.ndarray:
             return np.empty((0, 512), dtype=np.float32)
 
         # Process in chunks to avoid OOM
-        chunk_size = 32
+        chunk_size = 256
         embeddings_list = []
 
         for i in range(0, len(aligned_faces), chunk_size):