Univer OCR

This project is the course work at ITMO University. Aim of the project is to develop OCR software using parallel programming and self-written Deep Learning Framework.

Project was developed and tested on Windows 10, however, it might work on Linux too.

If you have GPU, you may unleash it's power to significantly increase speed of neural net in this project. If you don't have one, or don't want to use it, skip steps 6 and 7 of Installation. Although this has not been tested either.

Installation

1. Install Python 3.7 from official site.

2. Install virtualenv via pip. This command must be run as administrator in Windows or using sudo in Linux.

   pip install virtualenv

3. Enter root folder of the project:

   cd /PATH/TO/PROJECT

4. Create virtual environment with Python 3.7:

   virtualenv .venv --python=python3.7

5. And activate it:

   In Windows:

   .venv\Scripts\activate.bat

   In Linux:

   source .venv/Scripts/activate

6. Download and install CUDA Toolkits. Refer to this table to find out which version you need. If you need to install CUDA Toolkits version different from 10.0, also install corresponding version of CuPy instead of specified one in the file requirements/base.txt.

7. Set environmental variable CUDA_HOME to directory of installed CUDA Toolkits:

   In Windows:

   set CUDA_HOME=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0

   In Linux:

   export CUDA_HOME=/PATH/TO/CUDA/TOOLKITS

   Also you may do this via your OS system settings or by adding latter command into your .bashrc file. In this case your terminal must be restarted to be able to use the environmental variable.

8. While being with activated virtualenv, install requirements. If needed (in step 6), change it. No administrator or sudo is needed here.

   pip install -r requirements/base.txt

Usage

To run any scripts in this project you should either activate virtualenv (see step 5 of Installation) or use Python executable from it (.venv\Scripts\python.exe or .venv/Scripts/python).

1. To start web application run this command:

   python start_web_app.py

   Now you can access it in your browser at http://127.0.0.1:80. If you need to start with different port, change it in file start_web_app.py.

   Though, it is not recommended to run training using web interface, because it noticeably slows down the process. It's better to train via command line. To do this see step 3 of Usage.

2. To generate train and validation data run this command:

   python run generate_data

   It will create directory generated_files/data.

3. To train the model run this command:

   python run train [use_gpu [console_mode [show_progress_bar [save_train_progress]]]]

   As you see, this command has a bunch of arguments:

   • use_gpu: may be True or False. If True, makes script to use your GPU, otherwise runs on CPU. Default is False.

   • console_mode: may be True or False. If False, makes your script to connect to web application (from step 1 of Usage), otherwise prints all output in the console. Default is True.

   • show_progress_bar: may be True or False. If True, displays progress bar for each epoch. Handy when running in console mode, but dramatically increases number of lines in log file, if you redirect output from console to file. Default is False.

   • save_train_progress: may be True or False. If True, saves all input and output pictures of each iteration of each epoch while training. This can help you visualize training process, but be very careful because it is extremely memory-consuming operation and may fill up your hard drive in no time. Saved pictures are located at generated_files/train_progress. Default is False.

   If you want to train the model from scratch, at first delete the file web_app/components/my_model/model_weights.json. It will initialize the model with random weights.