In the repository you will find an example of web app for the deployment of an object detection TF model. In the specific I implement a card detector for sicilian cards (https://it.wikipedia.org/wiki/Carte_da_gioco_italiane), and I implement an application for storage and tracking of sicilian card games (scopa, briscola and tivitti).
The user can login with given credentials and then start a game.
During the game the user can use its camera (selectable from the available ones from the device) to take pictures of the cards from the current round. Cards are then recognised and the output is stored in the page ready for submit. The user can also review the detected cards before submission to the database.
The app stays ready for the next steps. Once the game is finished, the "End Game" button will close the game from the database and redirect the user to start a new game.
Images are stored in order to perform a regular reinforcment learning for the card detector model, and in the future to train a recommendation engine for the next move to play on a game.
The whole application is implemented within the Flask framework, with the addition of gunicorn and an Nginx reverse proxying for improved security and use of SSL protocols. The Nginx is configured to work both with a local run of the docker instances, or to run on a K8s cluster. For a deployment on a K8s cluster please refer to https://github.com/paoloalba/deployer_k8s.
The application can also be used to perform some light-learning, in order to easily extend your training data set (see https://github.com/paoloalba/dev_tf_model/blob/master/app/helpers/prediction_handler.py) and then improve the model.
The web app is mobile compatible.
Bulma CSS is employed for html styling (https://bulma.io/),
To reproduce the steps to generate the card detector model, or to train your own, please refer to https://github.com/paoloalba/dev_tf_model
Modify the tf_build_run_server.bat file for your needs.
This will run two containers, respectively for the main flask app and for the Nginx reverse proxying.
On your local host you will be able to access the app directly on port 5000, or on port 80 through the Nginx proxy (this will result in better logins).
To generate the needed credentials see the "credential_creator" section.
Modify the tf_build_push_server.bat file for your needs.
This will first build the training and evaluation Docker images, and then push them to your specified Docker Registry.
The script is already set for the use of Azure Container Registry, but any other technology can be employed.
In analyse_logs I provide a convenience class to analyse the outcoming server logs, e.g. to inspect malicious connections attempts (there are more than one could expect!)
After I started to play sicilian card games with my wife, I started wondering if a machine learning approach would have been faster than her in learning the meaning of the cards (indeed the pattern on cards are quite peculiar and will surely confuse a first time player). In the end she was pretty much faster in learning the cards, but I am still happy to have had a chance to bring some of my roots to the machine learning world!