move cuda init

app.py

@@ -36,15 +36,6 @@ onnx_file = hf_hub_download(repo_id="dylanplummer/ropenet", filename="model.onnx
 # config = {"PERFORMANCE_HINT": "LATENCY"}
 # compiled_model_ir = ie.compile_model(model=model_ir, device_name="CPU", config=config)
 
-# check if GPU is available
-if torch.cuda.is_available():
-    providers = [("CUDAExecutionProvider", {"device_id": torch.cuda.current_device(),
-                                            "user_compute_stream": str(torch.cuda.current_stream().cuda_stream)})]
-    sess_options = ort.SessionOptions()
-    ort_sess = ort.InferenceSession(onnx_file, sess_options=sess_options, providers=providers)
-else:
-    ort_sess = ort.InferenceSession(onnx_file)
-
 
 class SquarePad:
     # https://discuss.pytorch.org/t/how-to-resize-and-pad-in-a-torchvision-transforms-compose/71850/9
@@ -63,6 +54,14 @@ def sigmoid(x):
 @spaces.GPU(duration=180)
 def inference(x, count_only_api, api_key, img_size=224, seq_len=64, stride_length=32, stride_pad=3, batch_size=4, miss_threshold=0.8, marks_threshold=0.6, median_pred_filter=True, center_crop=True, both_feet=True, api_call=False):
     print(x)
+    # check if GPU is available
+    if torch.cuda.is_available():
+        providers = [("CUDAExecutionProvider", {"device_id": torch.cuda.current_device(),
+                                                "user_compute_stream": str(torch.cuda.current_stream().cuda_stream)})]
+        sess_options = ort.SessionOptions()
+        ort_sess = ort.InferenceSession(onnx_file, sess_options=sess_options, providers=providers)
+    else:
+        ort_sess = ort.InferenceSession(onnx_file)
     #api = HfApi(token=os.environ['DATASET_SECRET'])
     #out_file = str(uuid.uuid1())
     has_access = False