gie.txt

# Example GIE (nvinfer) config — YOLO detection.
#
# Pass this to osprey.configure(gie_config="…/gie.txt").
#
# THREE paths below are environment-specific — set them before use:
#   onnx-file / model-engine-file  -> your model
#   labelfile-path                 -> your class labels (one per line)
#   custom-lib-path                -> the parser .so that SHIPS WITH ospreyai.
#
# Resolve the shipped parser path with:
#   python3 -c "import osprey,os;print(os.path.join(os.path.dirname(osprey.__file__),'server','deepstream','lib','nvdsinfer_yolo_det.so'))"
#
# Parsers shipped in the package (osprey/server/deepstream/lib/):
#   nvdsinfer_yolo_det.so  -> YOLO detection      (parse-bbox-func-name=NvDsInferParseCustomYoloDet)
#   nvdsinfer_yolo_seg.so  -> YOLO segmentation
#   nvdsinfer_rtdetr.so    -> RT-DETR detection
#
# Do NOT point custom-lib-path at /deepstream_app/… — that was the old dev
# layout; the parser now lives inside the installed ospreyai package.

[property]
gpu-id=0
net-scale-factor=0.0039215697906911373
model-color-format=0

# --- model (edit these) ---
onnx-file=/path/to/yolo26m-trt.onnx
model-engine-file=/path/to/yolo26m-trt.onnx.engine
labelfile-path=/path/to/labels.txt

batch-size=1
network-mode=2
num-detected-classes=80
infer-dims=3;640;640
interval=0
gie-unique-id=1
process-mode=1
network-type=0
cluster-mode=4
maintain-aspect-ratio=1
symmetric-padding=1

# --- parser (edit custom-lib-path to the shipped .so, see header) ---
parse-bbox-func-name=NvDsInferParseCustomYoloDet
custom-lib-path=/opt/venv/lib/python3.12/site-packages/osprey/server/deepstream/lib/nvdsinfer_yolo_det.so

[class-attrs-all]
threshold=0.25
topk=100