fix: update input size for FaceDetector and enable TensorRT caching in ONNX providers

Author: Avneesh11905

inference_api/infrastructure/di_container.py

@@ -15,7 +15,7 @@ class Container(containers.DeclarativeContainer):
     face_detector = providers.Singleton(
         FaceDetector,
         model_path="models/buffalo_l/det_10g.onnx",
-        input_size=(1920, 1920),
+        input_size=(1024, 1024),
         device="cuda",
         num_threads=1,
     )

inference_api/infrastructure/onnx/utils.py

@@ -6,6 +6,17 @@ def get_providers(device: str) -> list:
     if device == "cuda":
         available = ort.get_available_providers()
         providers: list[str | tuple[str, dict[str, object]]] = []
+        if "TensorrtExecutionProvider" in available:
+            providers.append(
+                (
+                    "TensorrtExecutionProvider",
+                    {
+                        "trt_engine_cache_enable": True,
+                        "trt_engine_cache_path": "./trt_cache",
+                        "trt_fp16_enable": True,
+                    },
+                )
+            )
         if "CUDAExecutionProvider" in available:
             providers.append(
                 (

main_api/infrastructure/inference/hf_inference_service.py

@@ -24,7 +24,7 @@ async def get_face_encodings(
             },
         }
 
-        async with httpx.AsyncClient(timeout=120.0) as client:
+        async with httpx.AsyncClient(timeout=None) as client:
             response = await client.post(
                 f"{self.api_url}/", json=payload, headers=headers
             )