ROS inteface of Menoh library by @pfnet-research.

The MenohNodelet loads ONNX model and export input and output as ROS topics.

Thanks to the power of Menoh, it can run neural network models efficiently without GPGPUs.

$ mkdir -p catkin_ws/src
$ cd catkin_ws
$ catkin init
$ cd src
$ git clone https://github.com/akio/menoh_ros.git
$ catkin build

In preparation now.

Following diagram depicts the architecture of MenohNodelet pipeline.

        |
        | Any Input
        V
+-----------------+
| InputNode(let)  |
+-----------------+
        |
        | std_msgs/Float32MultiArray
        V
+-----------------+
| MenohNodelet    |<--- ONNX Model
+-----------------+
        |
        | std_msgs/Float32MultiArray
        V
+-----------------+
| OutputNode(let) |
+-----------------+
        |
        | Any Output
        V

MenohNodelet is a core nodelet of this package. This nodelet loads ONNX model file and export them as std_msgs/Float32MultiArray topics. When it receives a input message, it loads the message into the neural network model. After the neural network computes the output, the nodelet translate the output into a ROS message and publish it.

This nodelet subscribes sensor_msgs/Image and converts it to std_msgs/Float32MultiArray and publishes it to MenohNodelet.

This nodelet subscribes std_msgs/Float32MultiArray and lodas category label data from a textfile. When it receives a message, it compute softmax of the message and publish a corresponding label line as a std_msgs/String.

$ python scripts/retrieve_data.py
$ roslaunch launch/vgg16.launch

See launch/vgg16.launch as an example.

This package is available under terms of the MIT License.