RTMDet.md

RTMDet (MMYOLO) usage

• Convert model
• Compile the lib
• Edit the config_infer_primary_rtmdet file
• Edit the deepstream_app_config file
• Testing the model

Convert model

1. Download the RTMDet (MMYOLO) repo and install the requirements

git clone https://github.com/open-mmlab/mmyolo.git
cd mmyolo
pip3 install openmim
mim install "mmengine>=0.6.0"
mim install "mmcv>=2.0.0rc4,<2.1.0"
mim install "mmdet>=3.0.0,<4.0.0"
pip3 install -r requirements/albu.txt
mim install -v -e .
pip3 install onnx onnxslim onnxruntime

NOTE: It is recommended to use Python virtualenv.

2. Copy conversor

Copy the export_rtmdet.py file from DeepStream-Yolo/utils directory to the mmyolo folder.

3. Download the model

Download the pth file from RTMDet (MMYOLO) releases (example for RTMDet-s*)

wget https://download.openmmlab.com/mmrazor/v1/rtmdet_distillation/kd_s_rtmdet_m_neck_300e_coco/kd_s_rtmdet_m_neck_300e_coco_20230220_140647-446ff003.pth

NOTE: You can use your custom model.

4. Convert model

Generate the ONNX model file (example for RTMDet-s*)

python3 export_rtmdet.py -w kd_s_rtmdet_m_neck_300e_coco_20230220_140647-446ff003.pth -c configs/rtmdet/distillation/kd_s_rtmdet_m_neck_300e_coco.py --dynamic

NOTE: To change the inference size (defaut: 640)

-s SIZE
--size SIZE
-s HEIGHT WIDTH
--size HEIGHT WIDTH

Example for 1280

-s 1280

or

-s 1280 1280

NOTE: To simplify the ONNX model (DeepStream >= 6.0)

--simplify

NOTE: To use dynamic batch-size (DeepStream >= 6.1)

--dynamic

NOTE: To use static batch-size (example for batch-size = 4)

--batch 4

NOTE: If you are using the DeepStream 5.1, remove the --dynamic arg and use opset 12 or lower. The default opset is 17.

--opset 12

5. Copy generated files

Copy the generated ONNX model file and labels.txt file (if generated) to the DeepStream-Yolo folder.

Compile the lib

1. Open the DeepStream-Yolo folder and compile the lib

2. Set the CUDA_VER according to your DeepStream version

export CUDA_VER=XY.Z

• x86 platform

  DeepStream 7.1 = 12.6
  DeepStream 7.0 / 6.4 = 12.2
  DeepStream 6.3 = 12.1
  DeepStream 6.2 = 11.8
  DeepStream 6.1.1 = 11.7
  DeepStream 6.1 = 11.6
  DeepStream 6.0.1 / 6.0 = 11.4
  DeepStream 5.1 = 11.1
  

• Jetson platform

  DeepStream 7.1 = 12.6
  DeepStream 7.0 / 6.4 = 12.2
  DeepStream 6.3 / 6.2 / 6.1.1 / 6.1 = 11.4
  DeepStream 6.0.1 / 6.0 / 5.1 = 10.2
  

3. Make the lib

make -C nvdsinfer_custom_impl_Yolo clean && make -C nvdsinfer_custom_impl_Yolo

Edit the config_infer_primary_rtmdet file

Edit the config_infer_primary_rtmdet.txt file according to your model (example for RTMDet-s* with 80 classes)

[property]
...
onnx-file=kd_s_rtmdet_m_neck_300e_coco_20230220_140647-446ff003.pth.onnx
...
num-detected-classes=80
...
parse-bbox-func-name=NvDsInferParseYolo
...

NOTE: The RTMDet (MMYOLO) resizes the input with center padding. To get better accuracy, use

[property]
...
maintain-aspect-ratio=1
symmetric-padding=1
...

NOTE: The RTMDet (MMYOLO) uses BGR color format for the image input. It is important to change the model-color-format according to the trained values.

[property]
...
model-color-format=1
...

NOTE: The RTMDet (MMYOLO) uses normalization on the image preprocess. It is important to change the net-scale-factor and offsets according to the trained values.

Default: mean = 0.485, 0.456, 0.406 and std = 0.229, 0.224, 0.225

[property]
...
net-scale-factor=0.0173520735727919486
offsets=103.53;116.28;123.675
...

Edit the deepstream_app_config file

...
[primary-gie]
...
config-file=config_infer_primary_rtmdet.txt

Testing the model

deepstream-app -c deepstream_app_config.txt

NOTE: The TensorRT engine file may take a very long time to generate (sometimes more than 10 minutes).

NOTE: For more information about custom models configuration (batch-size, network-mode, etc), please check the docs/customModels.md file.