TODOs:

• Make inference more generic (command line/launch file configs)
• Add time benchmarks

Dependencies

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

Conversions

(outputNodeName)+ means one or more names

checkpoint to .pb

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

.pb to uff

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

.uff to .plan

You need to do this conversion on the device that is going to use the model

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

Inference

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

Example Model

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