Etlt model not detecting with deepsteam python apps

[zzww]

Please provide complete information as applicable to your setup.
• Hardware Platform: Jetson Xavier NX
• DeepStream Version: 6.0
• Language: Python

Hi,
I want to deploy .etlt models on deepstream with python bindings. I downloaded the models from this repo (https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps). The yolo models are working fine with “ds-tao-detection”, based on the steps in the documentation.

Now I want to check if those TAO models work in the deepstream with python bindings. The python sample code is downloaded from deepstream_python_apps. I used “deepstream-imagedata-multistream” as an example. In the configuration file, I replaced

#model-file=…/…/…/…/samples/models/Primary_Detector/resnet10.caffemodel
#proto-file=…/…/…/…/samples/models/Primary_Detector/resnet10.prototxt
#model-engine-file=…/…/…/…/samples/models/Primary_Detector/resnet10.caffemodel_b1_gpu0_int8.engine
#labelfile-path=…/…/…/…/samples/models/Primary_Detector/labels.txt
#int8-calib-file=…/…/…/…/samples/models/Primary_Detector/cal_trt.bin

with

tlt-encoded-model=/home/user/Projects/deepstream_tao_apps/models/yolov3/yolov3_resnet18.etlt
int8-calib-file=/home/user/Projects/deepstream_tao_apps/models/yolov3/yolov3nv.trt8.cal.bin
model-engine-file=/home/user/Projects/deepstream_tao_apps/models/yolov3/yolov3_resnet18.etlt_b1_gpu0_int8.engine
labelfile-path=/home/user/Projects/deepstream_tao_apps/configs/yolov3_tao/yolov3_labels.txt
parse-bbox-func-name=NvDsInferParseCustomBatchedNMSTLT
custom-lib-path = /home/user/Projects/deepstream_tao_apps/post_processor/libnvds_infercustomparser_tao.so

Here the engine file was generated automatically from running the “ds-tao-detection” in the deepstream_tao_apps. The parsing functions (parse-bbox-func-name, custom-lib-path) are set the same as in the deepstream_tao_apps.

Now the python apps run with no error, but it is not detecting any object.

Thanks.

[mchi]

Hi @zzww ,
Please take a look DeepStream SDK FAQ - #21 by mchi , you may miss some offset, scale settings.

[zzww]

Hi @mchi ,
Thank you for your reply. I added the scale, offset and a few parameters as in the configuration file in deepstram_tao_apps. It is now working in a more reasonable way.

I am wondering what is a good way to find out other parameters. The configs in deepstream_tao_apps may provide a hint on those parameters when I deploy them in deepstream_python_apps, but if I want to deploy my customized model trained with TAO, how should I set those parameters? Thanks.

[mchi]

Hi @zzww
I think the best way to go through the doc and understand the meanings of these parameters, if you know well about TLT train, it should be easy to map them.