Expose deep learning models on a Coral usb accelerator via a Flask app. To run the app and expose over a network:
$ python3 coral-app.py
Then use curl to query:
$curl -X POST -F image=@people_car.jpg 'http://localhost:5000/predict'
{'predictions': [{'bounding_box': {'x1': 838.29,
'x2': 918.53,
'y1': 135.01,
'y2': 407.59},
'confidence': '95.31',
'label': 'person'},
{'bounding_box': {'x1': 302.61, 'x2': 613.57, 'y1': 115.94, 'y2': 361.16},
'confidence': '91.02',
'label': 'car'},
{'bounding_box': {'x1': 226.54, 'x2': 350.93, 'y1': 143.46, 'y2': 374.72},
'confidence': '58.2',
'label': 'person'},
{'bounding_box': {'x1': 215.56, 'x2': 346.96, 'y1': 212.26, 'y2': 419.51},
'confidence': '26.95',
'label': 'bicycle'},
{'bounding_box': {'x1': 422.38, 'x2': 454.59, 'y1': 171.35, 'y2': 210.02},
'confidence': '21.09',
'label': 'person'},
{'bounding_box': {'x1': 359.8, 'x2': 389.19, 'y1': 161.48, 'y2': 204.62},
'confidence': '21.09',
'label': 'person'},
{'bounding_box': {'x1': 305.36, 'x2': 346.21, 'y1': 147.99, 'y2': 211.31},
'confidence': '21.09',
'label': 'person'},
{'bounding_box': {'x1': 5.55, 'x2': 21.94, 'y1': 9.36, 'y2': 45.3},
'confidence': '21.09',
'label': 'traffic light'},
{'bounding_box': {'x1': 239.76, 'x2': 324.57, 'y1': 260.07, 'y2': 406.78},
'confidence': '16.02',
'label': 'bicycle'},
{'bounding_box': {'x1': 299.75, 'x2': 358.61, 'y1': 154.64, 'y2': 298.7},
'confidence': '16.02',
'label': 'person'}],
'success': True}
See the Jupyter notebook for usage with python requests library.
NOTE: you need to update the MODEL and LABELS file paths in coral-app.py. For compatability with the way these paths are hard coded in this repo, you can on a pi cd ~, mkdir edgetpu, mkdir all_models, cd all_models, wget https://dl.google.com/coral/canned_models/all_models.tar.gz, tar xf all_models.tar.gz, rm all_models.tar.gz
I am running the server on a pi 3 with the raspi camera below. FYI the camera is mounted on a pan-tilt stage.
- The official pre-compiled models are at -> https://coral.withgoogle.com/models/
- It is also possible to train your own models -> try using Google Colaboratory as the free environment for training or -> https://cloud-annotations.github.io/training/object-detection/cli/index.html
- Simple front end for uploading images -> https://github.com/gxercavins/image-api
- Handle traffic -> https://www.pyimagesearch.com/2018/01/29/scalable-keras-deep-learning-rest-api/
I have published code for using this app with Home Assistant -> HASS-Google-Coral
There's a Dockerfile included that will construct a container to run this flask daemon.
By default, it will fetch
https://dl.google.com/coral/canned_models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite
to use as the model and
https://dl.google.com/coral/canned_models/coco_labels.txt for labels.
If you'd like to use alternative model and label files with an already
built container, you can put them into a directory that's mounted in the container and then
start the container using MODEL and LABELS environment variables to refer to the paths
inside the container.
Build the container like
$ docker build -t coral .
You can start the container with something like:
docker run --restart=always --detach --name coral \
-p 5000:5000 --device /dev/bus/usb:/dev/bus/usb coral:latest
it's important to use the --device option to pass in the USB bus device so that the Coral
edgetpu USB device can be found
Logging produced by the container is sort of broken. There seems to be a mixture of output to STDOUT and STDERR, each independently buffered.
- https://github.com/google-coral
- Using the official pi camera with Coral
- https://github.com/snmsung716/Following-camera-with-Coral-Accelerator-on-Raspberry-Pi
- https://github.com/lemariva/raspbian-EdgeTPU
- Pyimagesearch article on Coral
- Hands on with coral + docker
- Official Raspicam example
I forked the code in this excellent article -> https://blog.keras.io/building-a-simple-keras-deep-learning-rest-api.html code