This houses all things related to the RPi ML container for SONYC. The Dockfiles allow for the building of a Docker container with the required packages installed to run ML inference on an arm7l based sensor node.
It is recommended to use nohup to build the images on your test Raspberry Pi so you can leave the build process detached either when running locally or over SSH. The following can be run when within the directory of the Dockerfile:
nohup docker build --no-cache=true -t <IMAGE_NAME>:<IMAGE_VERSION> . &
There are currently two versions:
sonyc_ml/- only includes tflite runtime components (fully untested) - 985MBsonyc_ml_full_tf/- includes tflite runtime and full tensorflow - 1.3GB
To pull the docker image from Docker Hub to your test RPi use the following command:
docker pull cmydlarz/sonyc_ml:<IMAGE_VERSION>
To setup your test RPi to push updated images you need to add your credentials locally and login:
docker login --username=yourhubusername --email=youremail@company.com
Make sure you have been added as a collaborator to the cmydlarz/sonyc_ml repository: https://cloud.docker.com/repository/docker/cmydlarz/sonyc_ml
For docker push commands to complete from RPi, you need to set max-concurrent-uploads to 1 when pushing large images up:
nano /etc/docker/daemon.json
{
"max-concurrent-uploads": 1
}
systemctl restart docker
Then to push an updated image:
docker push cmydlarz/sonyc_ml:<IMAGE_VERSION>
SSH into the test Pi 2B and head to /home/pi/mnt/pi_ml_container. The username and password are the standard default Pi ones.
You can run the scripts using the following command (you must use full paths that exist inside the container):
docker run -it -v /home/pi/mnt:/mnt sonyc_ml_full:0.1 python /mnt/pi_ml_container/test_emb_gen_tflite.py <model file name> <audio file name in data folder> <target fs> <num mel bands> <mel hop len>
For example:
docker run -it -v /home/pi/mnt:/mnt sonyc_ml_full:0.1 python /mnt/pi_ml_container/test_emb_gen_tflite.py /mnt/pi_ml_container/tflite_models/quantized_model_8000_default.tflite /mnt/pi_ml_container/data/dog_1.wav 8000 64 160
INFO: Initialized TensorFlow Lite runtime.
== Input details ==
{'name': 'input_1', 'index': 33, 'shape': array([ 1, 64, 51, 1]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0)}
type: <class 'numpy.float32'>
== Output details ==
{'name': 'max_pooling2d_4/MaxPool', 'index': 34, 'shape': array([ 1, 1, 1, 256]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0)}
Processing time: 0.143