search

robmarkcole / coral-pi-rest-server

Perform inferencing of tensorflow-lite models on an RPi with acceleration from Coral USB stick
https://coral.ai/products/accelerator
MIT License
67 stars 20 forks source link
deep-learning object-detection python raspberry-pi ssd tensorflow tensorflow-lite

readme

coral-pi-rest-server

Perform inference using tensorflow-lite deep learning models with hardware acceleration provided by a Coral usb accelerator running on a raspberry pi or linux/mac. The models are exposed via a REST API allowing inference over a network. To run the app with default model:

$ python3 coral-app.py --models-directory models

Then use curl to query:

curl -X POST -F image=@images/test-image3.jpg -F min_confidence=.5 'http://localhost:5000/v1/vision/detection'

{'predictions': [{'confidence': 0.953125,
   'label': 'person',
   'x_max': 918,
   'x_min': 838,
   'y_max': 407,
   'y_min': 135},
  {'confidence': 0.58203125,
   'label': 'person',
   'x_max': 350,
   'x_min': 226,
   'y_max': 374,
   'y_min': 143}]
 'success': True}

To see the help run:

$ python3 coral-app.py -h

See the coral-app-usage.ipynb Jupyter notebook for usage from python.

Warning: it has been quite frustrating working with this hardware, with many tedious setup problems, hardware timeouts. Therefore I created a fork of this project that does not require a Coral at https://github.com/robmarkcole/tensorflow-lite-rest-server On an RPI 4 inference times are very fast even without a coral.

Models

If you have installed the raspberry pi disk images from edgetpu-platforms then you already have all the models in home/pi/all_models. If you are using a mac/linux desktop you can download the models from here. 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

Pi setup on existing pi

Follow the instructions on -> https://coral.ai/docs/accelerator/get-started/#1-install-the-edge-tpu-runtime Note I had headaches with this approach, better to use the disk image.

Pi setup with disk image

Install one of the disk images from edgetpu-platforms. In the /home/pi directory git clone this repository. You wil now have a file structure like (pi3 &pi4 only, pi-zero differs):

$ ls

all_models  coral-pi-rest-server  edgetpu_api  examples-camera  project-posenet  project-teachable  simple-demo

Use the cd command to enter ~/simple-demo and test you installation:

pi@pi:~/simple-demo $ python3 classify_image.py --model /home/pi/all_models/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite --label  /home/pi/all_models/inat_bird_labels.txt --image  parrot.jpg

INFO: Initialized TensorFlow Lite runtime.
---------------------------
Ara macao (Scarlet Macaw)
Score :  0.76171875

Assuming you saw the result above, use the cd command to enter ~/coral-pi-rest-server and (system wide, no virtual environment) install the required dependencies:

~/coral-pi-rest-server $ pip3 install -r requirements.txt

Now run the app:

~/coral-pi-rest-server $ python3 coral-app.py

Service

You can run the app as a service. Edit coral.service file to fit your needs and copy into /etc/systemd/system as root using sudo cp coral.service /etc/systemd/system/coral.service. Once this has been copied, you can attempt to start the service using sudo systemctl start coral.service. See the status and logs with sudo systemctl status coral.service. Stop the service with sudo systemctl stop coral.service. You can have it start automatically on reboot by using sudo systemctl enable coral.service.

Pi 3

I am running a pi3 with the raspi camera below. FYI the camera is mounted on a pan-tilt stage.

Using the pi3 (which has only usb2) processing a single request from another machine via this server takes ~ 300 to 500 ms. Therefore this is certainly fast enough to process images at 1 FPS which in my experience is suitable for tasks such as counting people in a room.

Pi 4

The pi4 has USB3 so we would expect better speeds. However I get inference time consistently ~ 500 ms when querying from another machine. Therefore it appears the coral library does not yet take advantage of the USB3.

Pi-zero

I found that I get inference time in the range 2.5 to 5s when querying from another machine, so significantly slower than the pi3/pi4.

Power

I recommend using a power supply which can supply 3 Amps, I use this one. Note that the official RPI supply delivers only 2.5 Amp and I found this was causing an issue where the stick would 'go to sleep' after a day of continual use.

Mac

To use the coral with a mac follow the official install instructions here

No coral?

I created a fork of this project that does not require a Coral at https://github.com/robmarkcole/tensorflow-lite-rest-server

Deepstack & Home Assistant

The data returned by the app is as close as possible in format to that returned by Deepstack object detection endpoint, allowing you to use this app as the backend for HASS-Deepstack-object. There is also support for running as an addon via github.com/grinco/HASS-coral-rest-api

Troubleshooting

Q: I get the error: HandleQueuedBulkIn transfer in failed. Not found: USB transfer error 5 [LibUsbDataInCallback] A: I reflashed the SD card and tried again with success

Q: I get error ValueError: Failed to load delegate from libedgetpu.1.dylib A: libedgetpu is not in the expected location

References

Credit

I forked the code in this excellent article -> https://blog.keras.io/building-a-simple-keras-deep-learning-rest-api.html code

Development Mac

Note: important to use USB-C to USB-C cable. Had some connection issues initially documented in this issue