deepstream如何使用自己训练的darknet模型

---

1 编译

• (1) 准备自己的模型文件
  • .weight
  • .cfg
  • .label.txt
• (2) 根据模型，修改对应的配置文件/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_Yolo/config_infer_primary_yoloV3_tiny.txt

[property]
gpu-id=0
net-scale-factor=1
#0=RGB, 1=BGR
model-color-format=0
custom-network-config=/home/nvidia/darknet/yolov3-tiny.cfg
model-file=/home/nvidia/darknet/yolov3-tiny.weights
#model-engine-file=model_b1_fp32.engine
labelfile-path=/home/nvidia/darknet/voc.names
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=2
num-detected-classes=20 # 类别数
gie-unique-id=1
is-classifier=0
maintain-aspect-ratio=1
parse-bbox-func-name=NvDsInferParseCustomYoloV3Tiny
custom-lib-path=nvdsinfer_custom_impl_Yolo/libnvdsinfer_custom_impl_Yolo.so

• (3) 如果有INT8精度校准模型,可以设置选项int8-calib-file
• (4) 查询CUDA版本 nvcc -V，设置环境变量CUDA_VER

nvcc -V
export CUDA_VER=10.1

• (5) 修改nvdsinfer_custom_impl_Yolo/nvdsparsebbox_Yolo.cpp文件
  • 修改NUM_CLASSES_YOLO=
  • 修改NvDsInferParseCustomYoloV3Tiny(根据实际模型的参数)
• (6) 编译生成链接库.so

make -C nvdsinfer_custom_impl_Yolo

2 运行

• （1） 设置PGIE属性(可以读配置文件)

[PGIE]
config-file-path=/home/nvidia/darknet-car/config_infer_primary_yoloV3_tiny.txt
batch-size=1
interval=1
unique-id=1
gpu-id=0
model-engine-file=/home/nvidia/darknet-car/model_b1_fp16.engine

• (2) 设置config-file-path

[property]
gpu-id=0
net-scale-factor=1
#0=RGB, 1=BGR
model-color-format=0
custom-network-config=/home/nvidia/darknet-car/yolov3-tiny.cfg
model-file=/home/nvidia/darknet-car/yolov3-tiny.weights
labelfile-path=/home/nvidia/darknet-car/labels.txt
## 0=FP32, 1=INT8, 2=FP16 mode
batch-size=1
network-mode=2
num-detected-classes=20
gie-unique-id=1
is-classifier=0
maintain-aspect-ratio=1
parse-bbox-func-name=NvDsInferParseCustomYoloV3Tiny
custom-lib-path=/home/nvidia/darknet-car/nvdsinfer_custom_impl_Yolo/libnvdsinfer_custom_impl_Yolo.so
cluster-mode=2

[class-attrs-all]
nms-iou-threshold=0.5
pre-cluster-threshold=0.25

3 注意事项

• 不用运行每次重新编译engine文件:

设置PGIE属性，model-engine-file

• 如何设置`darknet’运行时置信度:

cluster-mode=2

[class-attrs-all]
nms-iou-threshold=0.5
pre-cluster-threshold=0.25

4 参考链接

《How to Use the Custom YOLO Model》