API provides serialization / deserialization and inference of uff models using Tensorrt (c++) methods on JetsonTx2. Achievable fps 14-18!
If you are looking to optimize Cnn performance through Tensorrt (C ++ API) on JetsonTx2, then this tutorial may help you.
- JetPack 4.4
- CUDA V10.2.89
- CUDNN 8.0.0
- OpenCV 4.4.0
- Tensorrt 7.1.3
- Place the supplied folders in
/usr/src/tensorrt/samples
. - Replace items with
common
- Replace Makefile and Makefile.config respectively.
- Run
/inferUffModel$ make
. - You must first have a .uff network model. Use the converter
python3 /usr/lib/python3.6/dist-packages/uff/bin/convert_to_uff.py \
/frozen_inference_graph.pb -O NMS \
-p /config.py \
-o /frozen_inference_graph.uff
- Use jetsonStend similarly for caffe (ssd) models