From 40ed09b7cd9e062d2e7491ee595454d01bf7dcec Mon Sep 17 00:00:00 2001 From: Glenn Jocher Date: Mon, 26 Sep 2022 14:03:03 +0200 Subject: [PATCH] Apple MPS -> CPU NMS fallback strategy Until more ops are fully supported this update will allow for seamless MPS inference (but slower MPS to CPU transfer before NMS, so slower NMS times). Partially resolves https://github.com/ultralytics/yolov5/issues/9596 Signed-off-by: Glenn Jocher --- utils/general.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/utils/general.py b/utils/general.py index de7871cb23f9..a855691d3a1f 100644 --- a/utils/general.py +++ b/utils/general.py @@ -843,6 +843,8 @@ def non_max_suppression( if isinstance(prediction, (list, tuple)): # YOLOv5 model in validation model, output = (inference_out, loss_out) prediction = prediction[0] # select only inference output + if 'mps' in prediction.device.type: # MPS not fully supported yet, convert tensors to CPU before NMS + prediction = prediction.cpu() bs = prediction.shape[0] # batch size nc = prediction.shape[2] - nm - 5 # number of classes xc = prediction[..., 4] > conf_thres # candidates