• Make inference more generic (command line/launch file configs)
• Change the way the input is passed to the gpu in InferenceEngine.cu
• Add time benchmarks

1. Install nvidia drivers/cuda
2. Install Tensorflow pip install tensorflow or pip install tensorflow-gpu
3. Install TensorRT sudo aptitude install libnvinfer5 libnvinfer-dev

(outputNodeName)+ means one or more names

cd conversions/scripts/
python3 convert_to_pb.py path/to/model.chk.meta dir/of/checkpoint/ path/to/output.pb (outputNodeName)+

cd conversions/scripts/
python3 convert_pb_to_uff.py path/to/model.pb path/to/output.uff (outputNodeName)+

cd conversions
mkdir build && cd build
cmake ..
make -j4
./uff_to_plan models/mymodel.uff models/mymodel.plan inputs/X 256 256 1 up23/BiasAdd 1 500000 float

cd inference
# fix paths and file names on src/run_plan.cu
mkdir build && cd build
cmake ..
make -j4
cd src # inside build
./run_plan

We provide a simple model trained on Kitti Road detection that can be used to test the entire pipeline here. All inputs were resized to (368x1200) for training.

Input

Output

This project uses scripts from NVIDIA trt_image_classification