The text recognition is done via a convolutional neural network (CNN). The architecture of the CNN is to a significant extant customizable by writing a .yaml specification.

Examples of such specification can be found in the model directory.

In order to train the neural network with the given architecture, a few things have to be done first.

1. Prepare data sets

The data is expected to be in Keras-like format. This means that the data directory consists of folders representing classes usually we want two to three such directories - train, validation and test.

Example of training data set directory structure could look like this:

.
+-- train_data
|   +-- class_A
|   |   +-- sample_A_1.png
|   |   +-- sample_A_2.png
        ...
|   +-- class_B
|   |   +-- sample_B_1.png
|   |   +-- sample_B_2.png
        ...
|   +-- class_C
|   |   +-- sample_B_1.png
|   |   +-- sample_B_2.png

By default, the estimator expects images of shape (32, 32, 1), but it should be possible to specify custom IMAGE_SHAPE.

NOTE: Take a look at the char-generator GitHub repository. The generator generates data suitable for this purpose. Feel free to experiment with the fonts used for data generation.

2. Deployment

As a minimum it is refered to a deployment of the application itself, that is without client. It is useful for training and evaluation. It is also possible to get a simple prediction, but for that we should deploy a dedicated server (more on that bellow).

The application is shipped in docker. Make sure you have it docker installed and properly set up.

• the minimal image

  • defined by Dockerfile.min

  From the main directory, run the following command to build the minimal image.

  docker build -t $IMAGE_NAME -f Dockerfile.min .

  NOTE: Replace the $IMAGE_NAME with your custom value.

• the complete image, which contains bundled TensorFlow Serving and sets up the client as well

  • defined by Dockerfile (default)

  From the main directory, run the following command to build the minimal image.

  docker build -t $IMAGE_NAME .

  or you can make use of predefined docker-compose.yml and run the container immediately by

  docker-compose up

In case of building the containers (not using docker-compose), you have to run them.

Then execute one of the following commands to run the container. (see docker help for more info about running a container)

docker run -dit -p 6006:6006 --name $CONTAINER_NAME $IMAGE_NAME

NOTE: Replace the $CONTANER_NAME with your custom value.

or you can bind-mount the working directory, for example to provide training data easily

docker run -dit -p 6006:6006 -v ${PWD}:/code --name $CONTAINER_NAME $IMAGE_NAME

3. Usage

docker attach $CONTAINER_NAME

The command above attaches the container development environment.

The setup creates two entry points (executable commands). Feel free to explore the following commands:

• intect

  The base cli for training, evaluation and simple prediction request

• intect-client (not usable in the minimal image)

  Client api for more comples prediction requests and serving pre-trained models

NOTE: The project is currently in development phase, and is still, unfortunately, meant for developers mainly.

After the installation, the command intact --help or intact-client --help is not functionnal. Use intact or intact-client only to display help.