ocRnn

About this repo:

An OCR based on an end-to-end trainable Recurrent Neural Network

Content of the repo:

The project has been organized as follows:

• requirements.txt: a text file containing the needed packages to run the repo.
• train_model.py: a file with the training code.
• test_model.py: a file with the testing code.
• config/: the fodler containing the config files.
• data/: the folder that contains an example of the dataset used for training.
• model/: the folder that contains the model's architecture.
• saved_model/: the folder that contains the trained model, ready to be used.
• test_images/: the folder containing some images for test purpose.
• utils/: the folder that contains the utils files/methods.

Use the Repo:

N.B: use Python 3.8

1. Clone the repo:
on your terminal, run git clone https://github.com/maky-hnou/ocRnn.git
Then get into the project folder: ocRnn/
We need to install some dependencies:
sudo apt install python3-pip libpq-dev python3-dev

2. Install requirements:
Before running the app, we need to install some packages.

• Optional Create a virtual environment: To do things in a clean way, let's create a virtual environment to keep things isolated.
  Install the virtual environment wrapper: pip3 install virtualenvwrapper
  Add the following lines to ~/.bashrc:

export WORKON_HOME=$HOME/.virtualenvs
export PROJECT_HOME=$HOME
export VIRTUALENVWRAPPER_PYTHON=/usr/bin/python3
export VIRTUALENVWRAPPER_VIRTUALENV=~/.local/bin/virtualenv
source ~/.local/bin/virtualenvwrapper.sh

Run source ~/.bashrc
Run mkvirtualenv ocrnn
Activate the virtual environment: workon ocrnn (To deactivate the virtual environment, run deactivate)

• Install requirements: To install the packages needed to run the application, run pip3 install -r requirements.txt

N.B: If you don't have GPU, or don't have Cuda and Cudnn installed, replace tensorflow-gpu by tensorflow in requirements.txt.

3- Run the training:
The dataset: The dataset used to train the model is available via this link: a 10 GB dataset.
Once downloaded, extract it, then make the following changes to the config/config.yml file:
Line 15: put the path the training annotations.
Line 16: put the path the evaluation annotations.
Line 18: put the path the test annotations.

Once everything is set up, run the training command:

python3 train_model.py --config <config_file_path> --save_dir <path_where_to_save_the_model>

It is to notice that you need at least a 4GB GPU to be able to run the training. And the process will be so slow. So it is better to run the training on a server with a dedicated GPU (Colab, AWS ...).

4- Test the trained model:
There is a trained model, ready to use, included in saved_model/ folder.
To run the text recognition, use the following command:

python3 test_model.py --images <path_to_images> --config <config_file_path> --model <saved_model_path>

5- Demo:

Label: wonderful Prediction: [b'wonderful'] Confidence: [0.56338775]
Label: delighted Prediction: [b'delighted'] Confidence: [0.9994946]
Label: tiredness Prediction: [b'tiredness'] Confidence: [0.9997297]