Detection Node

This is a node for the Learning Loop wich provides an RESTful API for edge devices to retrieve inferences. It is intended to run on NVidia Jetson (>= r32.4.4) by utilizing TKDNN.

Features

• Active Learning for the Zauberzeug Learning Loop (upload images & detections with bad predictions)
• RESTful interface to retrieve predictions

Usage

Runs only on NVidia Jetson (Tegra Architecture).

docker pull zauberzeug/tkdnn_detection_node:nano-r32.5.0 # to make sure we have the latest image
docker run -it --rm --runtime=nvidia -p 80:80 \
-v $HOME/data:/data \          # bind the model to make it persistent (should contain an model.rt file)
-e NVIDIA_VISIBLE_DEVICES=all \ # to enable hardware acceleration
-e ORGANIZATION=zauberzeug \    # define your organization
-e PROJECT=demo\                # define the project for which the detector should run
zauberzeug/tkdnn_detection_node:nano-r32.5.0

If the container is up and running you can get detections through the RESTful API:

curl --request POST -H 'mac: FF:FF:FF:FF:FF' -F 'file=@test.jpg' localhost/detect

Model Files

For startup the image expects a valid model.rt file, training.cfg and names.txt in the /data directory. These will automatically provided by converter nodes.

Development

Put a TensorRT model model.rt and a names.txt with the category names into the data folder. You can use the download_model_for_testing.sh helper.

Build the container with ./docker.sh build and run it with ./docker.sh run. Now you can connect to the container with vs code and modify the code.