Merge pull request #55 from SnuffSocket/onnx-branch

Fix: Set ONNX Options: Disable telemetry, limit threads, lower CPU usage
This commit is contained in:
dfgHiatus authored and GitHub committed 2024-10-21 13:45:43 -07:00
commit 4cec4ff80d
2 files changed
+9 -2

No files matched your search

+4
View File
@@ -89,9 +89,13 @@ class BabbleProcessor:
self.calibrate_config = np.empty((1, 45))
self.min_max_array = np.empty((2, 45))
ort.disable_telemetry_events()
self.opts = ort.SessionOptions()
self.opts.inter_op_num_threads = 1
self.opts.intra_op_num_threads = settings.gui_inference_threads
self.opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
self.opts.add_session_config_entry("session.intra_op.allow_spinning", "0") # ~3% savings worth ~6ms avg latency. Not noticeable at 60fps?
self.opts.enable_mem_pattern = False
if self.runtime in ("ONNX", "Default (ONNX)"): # ONNX
if self.use_gpu:
provider = "DmlExecutionProvider"
+5 -2
View File
@@ -88,10 +88,13 @@ class LandmarkProcessor:
#self.calibrate_config = np.empty((1, 45))
#self.min_max_array = np.empty((2, 45))
ort.disable_telemetry_events()
self.opts = ort.SessionOptions()
self.opts.inter_op_num_threads = 1
self.opts.intra_op_num_threads = settings.gui_inference_threads
self.opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
self.opts.add_session_config_entry("session.intra_op.allow_spinning", "0") # ~3% savings worth ~6ms avg latency. Not noticeable at 60fps?
self.opts.enable_mem_pattern = False
if self.runtime in ("ONNX", "Default (ONNX)"): # ONNX
if self.use_gpu: provider = 'DmlExecutionProvider'
else: provider = "CPUExecutionProvider" # Build onnxruntime to get both DML and OpenVINO
@@ -219,7 +222,7 @@ class LandmarkProcessor:
self.current_image,
self.current_frame_number,
self.current_fps,
) = self.capture_queue_incoming.get(block=True, timeout=0.2)
) = self.capture_queue_incoming.get(block=True, timeout=0.1)
except queue.Empty:
# print("No image available")
continue