This is a demo application that shows how to run a network using TorchScripts dumped from other platforms along with LibTorch C++ API.
The application has been tested using the NVIDIA Jetson Nano SDK. The performance is actually not that good as it takes about 4 seconds to run MobileNetV2 based SSD network with input resolution of 300x300.
This project is supposed to be used together with pytorch-ssd, which can dump the TorchScript.
- The NVIDIA Jetson Nano Developer Kit
- NVIDIA Jetpack SD card image
- Python wheels for Jetson Nano that provides LibTorch
- GCC, Cmake, OpenCV, Boost, all of which can be installed from the Ubuntu 18.04 official repository
mkdir build && cd build
cmake -DCMAKE_PREFIX_PATH=<your libtorch path> .../ts_ssd -s mb2-ssd-lite-mp-0_686.pt -l voc-model-labels.txt -p 0.5 -i messi.jpgYou should see message like the following:
–torch-script specified with value = mb2-ssd-lite-mp-0_686.pt
–labels specified with value = voc-model-labels.txt
–probability-threshold specified with value = 0.5
–input-file specified with value = messi.jpg
cuDNN: Yes
CUDA: Yes
Loaded TorchScript mb2-ssd-lite-mp-0_686.pt
Start inferencing ...
Original image size [width, height] = [1296, 729]
Inference done in 4305 ms
Tensor shape: {3000 21}
Tensor shape: {3000 4}
Class index [15]: person
Tensor shape: {2 5}