Many thanks to Dima Rekesh, PhD who is the original author of these builds and write-up.

For initial Jetson hardware setup instructions, visit the wiki Home

Docker build instructions and files for deep learning container images

• Jetson TX2 (base)
• CUDA, CUDNN, OpenCV and supporting libs, full and lean variants
• Caffe deep learning framework
• Darknet deep learning framework with Yolo

These docker images are also available at the public bluehorizon docker hub repo as part of the Horizon project.

To build docker images (some successive images depend on previous. See below for prereqs)

1. Clone this repo locally
2. Build the base Jetson TX2 image. This base image is the prerequisite for later containers like Caffe, Darknet, and other deep learning frameworks

docker build -f Dockerfile.base -t openhorizon/jetson-tx2 .

This CUDNN container is the basis of builds for the containers below:

• BVLC's Caffe
• NVIDIA's Caffe
• Darknet (Yolo)

---

Caffe builds

To build Berkley's (BVLC) Caffe container:

1. First build the Jetson TX2 container with CUDNN: jetson-tx2 (see above)
2. Build the BVLC Caffe container:

cd caffe
docker build -f Dockerfile.caffe -t openhorizon/caffe-tx2 .

To build a container with NVIDIA's Caffe:

1. First build the Jetson TX2 container with CUDNN: 'jetson-tx2' (see above)
2. Build the NVIDIA Caffe container:

docker build -f Dockerfile.nvidia.caffe -t openhorizon/caffe-nvidia-tx2 .

NVIDIA DIGITS

To build NVIDIA DIGITS:

1. First build the NVIDIA Caffe container 'caffe-nvidia-tx2' (see above)
2. Build the NVIDIA DIGITS container:

docker build -f Dockerfile.digits -t openhorizon/digits-tx2 .

Darknet with Yolo

To build darknet with Yolo:

1. First build the 'jetson-tx2' container (see above)
2. Build darknet container:

cd darknet
docker build -f Dockerfile.darknet -t openhorizon/darknet-tx2 .

DustyNV's container:

To build Dusty NV's container:

1. First build the 'tensorrt-tx2' container TensorRT

cd tensorrt
docker build -f Dockerfile.tensorrt -t openhorizon/tensorrt-tx2 .
cd ../

2. Build Dusty NV's container:

cd dustyinference
docker build -f Dockerfile.dustyinference -t openhorizon/dustyinference-tx2 .

---

Optional base images:

Base drivers container (smallest): jetson-tx2-drivers

docker build -f Dockerfile.drivers -t openhorizon/jetson-tx2-drivers .

Base CUDA container: jetson-tx2-cudabase (for OpenCV4Tegra build)

docker build -f Dockerfile.cudabase -t openhorizon/jetson-tx2-cudabase .
cd caffe
docker build -f Dockerfile.opencv4tegra.caffe -t openhorizon/opencv4-tx2 .

---

Validating

To test the speed of AlexNet on the GPU:

docker run --privileged --rm openhorizon/caffe-tx2 /caffe/build/tools/caffe time --model=/caffe/models/bvlc_alexnet/deploy.prototxt --gpu=0

To do the same on the CPU:

docker run --privileged --rm openhorizon/caffe-tx2 /caffe/build/tools/caffe time --model=/caffe/models/bvlc_alexnet/deploy.prototxt

To run Digits:

docker run --privileged --name digits -p 5001:5001 -d openhorizon/digits-tx2
# Now you can connect to http://jetsonip:5001 using your browser 

To run Yolo:

# as root on host TX2 allow x, e.g. 'xhost +'
# then, assuming you have a usb webcam hooked up:
# tiny yolo model against the VOC data set is the fastest (only 20 classes though)
xhost + && docker run --privileged -e DISPLAY=$DISPLAY -v /tmp:/tmp --rm openhorizon/darknet-tx2 /darknet/darknet detector demo -c 1 cfg/voc.data cfg/tiny-yolo-voc.cfg tiny-yolo-voc.weights
# regular yolo model is much more accurate and against MS COCO 91 classes; however, 2x slower in my tests:
xhost + && docker run --privileged -e DISPLAY=$DISPLAY -v /tmp:/tmp --rm openhorizon/darknet-tx2 /darknet/darknet detector demo -c 1 cfg/coco.data cfg/yolo.cfg yolo.weights

To run DustyNV's container:

xhost + && docker run --privileged -v /dev:/dev -e DISPLAY=$DISPLAY -v /tmp:/tmp --net=host --ipc=host --rm -ti openhorizon/dustyinference-tx2 bash -c "cd /jetson-inference/build/aarch64/bin && ./imagenet-camera"