Introduction

Minimal example of how to generate an .exe executable using PyInstaller that can run Manga-Ocr. Due to Manga-Ocr's use of PyTorch, it is not trivial to generate a correctly functioning and lightweight .exe.

The instructions provide a way to generate an executable that launches Manga-Ocr using the CPU for image processing, as well as instructions for generating an executable that launches Manga-Ocr to use the GPU (Cuda).

We were able to successfully generate the executable by using Poetry to create a Python Environment, as well as a custom modification of PyInstaller's .spec file to optimize the compilation .

Instructions

0. Make sure to have Python 3.9 installed (other versions may cause conflicts).
1. Install Poetry if you don't have it.
2. Validate that it is correctly installed using: poetry -v
3. Create and activate the environment with: poetry shell
4. Install dependencies with: poetry install

---Instructions to compile Manga-Ocr to use CPU for image processing--- This compilation will result in a folder with an approximate weight of 650mb. If we compress the folder using WinRar, we can achieve a weight of 125mb.

5. Run the program with Python to validate its functionality: python main.py
6. Create the build using: pyinstaller MangaOcrPyInstaller.spec
7. The build can be found in directory /dist. To run the program, click on "MangaOcrPyInstaller.exe"

---Instructions to compile Manga-Ocr to use GPU (Cuda) for image processing--- Since Manga-Ocr uses PyTorch, GPU acceleration will only work if the computer has an Nvidia graphics card that supports Cuda. The graphics card drivers must be up to date. The build of Manga-Ocr with GPU usage is exponentially heavier. This compilation will result in a folder with an approximate weight of 4gb. If we compress the folder using WinRar, we can achieve a weight of 1.3gb.

5. Use poe force-cuda11 to install extra dependencies for GPU usage. This will install PyTorch 1.12, Torchvision 0.13, and Cuda 11.6 in our environment. This download is approximately 2.3gb
6. Run the program with Python to validate its functionality: python main.py. We should see a message indicating that the GPU (Cuda) is being used
7. Create the build using: pyinstaller MangaOcrPyInstaller.spec
8. The build can be found in directory /dist. To run the program, click on "MangaOcrPyInstaller.exe"

Considerations

• Manga-Ocr checks if we have the models stored in our cache. If they are not present, Manga-Ocr will download the models (approximately 400mb) before starting. Subsequent executions will be faster as the models will be cached
• If we use poe force-cuda11 and want to uninstall those dependencies to switch back to using only the CPU, we should run poe force-remove-cuda11 and then poetry install

Others

• The .spec file was generated using this solution as a base. Then it was merged with this implementation. Finally, it was manually validated which packages could be ignored.
• Achieving to install Cuda with Poetry was possible thanks to this approach.

Todo

• See if it's possible to reduce the weight of the GPU mode compilation by ignoring certain .dll and/or .lib files related to Cuda