django-ocr_translate 0.3.1

Django app for OCR and translation

OCR_translate

This is a Django app for creating back-end server aimed at performing OCR and translation of images received via a POST request.

The OCR and translation is performed using freely available machine learning models and packages (see below for what is currently implemented).

The server is designed to be used together with this browser extension, acting as a front-end providing the images and controlling the model languages and models being used.

Full Documentation

Running the server

If you plan to use a different settings (eg. database, or model location), you can either:

• manually edit the settings.py
• Use the provided Environment variables See below for a list of supported databases

You will also have to modify the ALLOWED_HOSTS in case you plan to access the server from somewhere other than localhost.

All the different way to run the server may provide different set of default values (each of them is targeted for a different level of usage).

From Release file

(Tested on Windows 11) From the github releases page you can download either:

• The CPU only version
• The GPU version split in file 1 and file 2 (The CUDA dependencies makes it take much more space), wich can be restored using tools like 7zip and NanaZip.

Usage: Unzip the file and from inside the folder, run the run_server-XXX.exe file (XXX=cpu/gpu)

The server will run with sensible defaults. Most notably the models files and database will be downloaded/created under %userprofile%/.ocr_translate. Also the gpu version will attempt to run on GPU by default, and fall-back to CPU if the former is not available.

For customization, you can set the environment variable yourself:

• Linux: use export ENV_VAR_NAME=XXX before launching the code from terminal.
• Windows: either via powershell or by searching for environment variable in the settings menu.

From Github installation

See the section on how to install from Github first.

The Github repo provides not only the Django app files, but also the already configured project files used to start the server.

Create/Initialize your database by running

• python manage.py migrate

inside your project folder.

Run the server using for example one of the following options:

• Django development server. This is more oriented for developing than deploying, but is fine for a self-hosted single-user server accepting connections only on localhost
  • From inside the project directory: python manage.py runserver PORT
  • The suggested PORT would be 4000 as it is the one set by default in the extension
• Nginx + Gunicorn:
  • Check the Dockerfile, as this is what the provided image makes use of.

At least for the first time, it is suggested to run the server with the Environment variables AUTOCREATE_LANGUAGES and AUTOCREATE_VALIDATED_MODELS set to "true" to automatically load the validated languages and models provided by the project.

Notes:

• Gunicorn workers will each spawn a separate instance of the loaded models, each taking its own space in the memory. This can quickly fill up the memory especially if running on GPU. Ideally set this to 1.
• Django development server will spawn new threads for handling incoming requests (if no currently existing thread is free), which share the same memory. Running more than one worker per loaded model concurrently might slow down the actual computation and in some case also block the execution.

From PyPI installation

See the section on how to install from PyPI first.

When installing the project from PyPI, only the app is available. This will need to be integrated in a Django project in order to be used. These are the minimal instruction for creating a project and start running the server:

• Run django-admin startproject mysite to create a django project
• Configure the server by replacing the automatically created files (strongly recommended):
  • settings.py with the one available on the repo.
  • urls.py with the one available on the repo.
• or by manually editing the files:
  • settings.py: Add the ocr_translated app to the INSTALLED_APPS
  • urls.py: Include the 'ocr_translate.urls' into your project urls.
• From here follow the same instructions as when starting from Github

From docker image

See the section on how to install from DockerHUB first.

This section assumes you have docker installed and the image of the project.

Run the command:

• docker run --name CONTAINER_NAME -v PATH_TO_YOUR_MODEL_DIRECTORY:/models -v PATH_TO_DIR_WITH_SQLITE_FILE:/data --env-file=PATH_TO_AND_ENV_VARIABLE_FILE -p SERVER_PORT:4000 -d ocr_translate

See the Environment variables section for configuring your environment variable file. Additionaly the docker image defines several other variables to automatically create an admin user for managing the database via the django-admin interface:

• UID: UID of the user owning the files in /models and /data
• GID: GID of the user owning the files in /models and /data
• NUM_WEB_WORKERS: Number of gunicorn workers for the server
• DJANGO_SUPERUSER_USERNAME: The username of the admin user to be created.
• DJANGO_SUPERUSER_PASSWORD: The password of the admin user to be created.

Installation

For both the Githyb and PyPI installation it is strongly suggested to install this project using a virtual environment.

From Github

• Clone or download the repository
  • git clone https://github.com/Crivella/ocr_translate.git
• Install the project dependencies (choose the appropriate files depending if you wanna run on GPU or CPU only):
  • pip install -r requirements-torch-[cpu/cuda].txt
  • pip install -r requirements.txt

From Docker

CPU and CUDA specific images are available on DockerHUB:

• CPU: docker pull crivella1/ocr_translate:latest-cpu
• GPU: docker pull crivella1/ocr_translate:latest-gpu

Manually create your image:

• Create a .pip-cache-[cpu/gpu] directory inside your project.
• Optional: re-install the project dependencies pointing this as the cache folder for pip (will make the build process much faster, by reusing the cached dependencies)
• Run docker build -t IMAGE_TAG -f Dockerfile-[cpu/gpu] .

From PyPI

Run the command

• pip install django-ocr_translate

By default torch 2.x will come in its CUDA enabled version. While this works also for CPU, it will also install ~1 GB of cuda dependencies. If you wish to run on CPU only, download the file requirements-torch-cpu.txt first and run

• pip install -r requirements-torch-cpu.txt

before installing the python package.

Supported Box OCR models

• EasyOCR

Supported text OCR models

• Hugging Face Transformers VisionEncoderDecoder models
• Tesseract (Requires tesseract to be installed on the machine)

Supported translation models

• Hugging Face Seq2Seq models

Writing plugins for the server

Since version 0.2.1 plugins can now be used to also add models to the database via the following entrypoints. See CHANGELOG for more details.

Since version 0.2.0 the server has been made pluginable. You can write a plugin for a model/web-service that has not yet been implemented, by subclassing the following models

• ocr_translate.models.OCRBoxModel: Must define the following methods (see the base models and the plugins under ocr_translate.plugins for example of function signature and expected input outputs):
  • load: function to load the model into memory. Can be defined to do nothing if not needed (e.g. another library that load the model on import or a plugin for a web-service)
  • unload: function to unload the model from memory.
  • _box_detection: Function that takes a PIL image as input and returns a list of bounding boxes.
• ocr_translate.models.OCRModel: Must define the following methods (see the base models and the plugins under ocr_translate.plugins for example of function signature and expected input outputs):
  • load: function to load the model into memory. Can be defined to do nothing if not needed (e.g. another library that load the model on import or a plugin for a web-service)
  • unload: function to unload the model from memory.
  • _ocr: Function that takes an image as input and returns the OCRed text (the image is the content of the bounding generated by a BOX model run)
• ocr_translate.models.TSLModel: Must define the following methods (see the base models and the plugins under ocr_translate.plugins for example of function signature and expected input outputs):
  • load: function to load the model into memory. Can be defined to do nothing if not needed (e.g. another library that load the model on import or a plugin for a web-service)
  • unload: function to unload the model from memory.
  • _translate: Function that takes a list of tokens or a list(list(tokens)) as inputs and returns the translated text as output either as a str or list[str] (this is needed to work efficiently with AI models that can perform multiple translations simultaneously)

NOTE: When subclassing the following has to be set inside the class (see django models doc)

class Meta:
    proxy = True

Until there is a registry service, do contact me if you write a plugin for the server so i can add a link to it in this README

Endpoints

This is not a REST API. As of now the communication between the server and a front-end is stateful and depend on the languages and models currently loaded on the server. In the future it would be interesting to separate the worker and database server, for an actual deployment, but might make the self-hosting more difficult to manage.

Endpoint	Method	Usage
/	GET	Handshake: the server will replay with a JSON response containing information about the available languages/models and the currently in use src/dst language and box/ocr/tsl models
/get_trans/	GET	Request to get all the available translations (e.g. using different models) of the text specified by the text GET parameter
/set_lang	POST	JSON request to switch the currently selected languages to the one specified by the keys: lang_src, lang_dst
/set_models	POST	JSON request to switch the currently loaded models to the one specified by the keys: box_model_id, ocr_model_id, tsl_model_id
/run_tsl	POST	JSON request to run the translation for the text specified by the key text
/run_ocrtsl	POST	JSON request to run the OCR and translation of an image (base64 as the contents key) or provide a previously obtained result (md5 of the base64 as the md5 key). md5 should be always specified, contents is optional

Environment variables

The first section of variable is defined at the APP level and will be available both for installation from Github or PyPI. The second section of variables is defined at the project level and is only available if using the settings.py provided in the repo.

App variables

Variable	Values	Usage
LOAD_ON_START	false[/true]	Will automatically load the most used source/destination languages and most used models for that language combination at server start
AUTOCREATE_LANGUAGES	false[/true]	Will automatically create the Language entries in the database as defined in languages.json
AUTOCREATE_VALIDATED_MODELS	false[/true]	Will automatically create the model entries defined in code and plugins entrypoints.
DEVICE	cpu[/cuda]	Which device to use with torch
EASYOCR_MODULE_PATH	$HOME/.EasyOCR	Directory where EasyOCR store its downloaded models
TRANSFORMERS_CACHE	$HOME/.cache/huggingface/hub/	Directory where Hugging Face models are being stored (either downloaded manually or downloaded by transformers)
TRANSFORMERS_OFFLINE	1[/0]	By default transformers will try to download missing models. Set this to 0 to only work in offline mode
TESSERACT_PREFIX	$TRANSFORMERS_CACHE/tesseract	Directory where tesseract will store and look for models
TESSERACT_ALLOW_DOWNLOAD	false[/true]	Control whether the app should download missing models (true) or work in offline mode only (false)
NUM_MAIN_WORKERS	4	Number of WorkerMessageQueue workers handling incoming OCR_TSL post requests
NUM_BOX_WORKERS	1	Number of WorkerMessageQueue workers handling box_ocr pipelines (Should be set as 1 until the pipeline is build to handle multiple concurrent request efficiently without slowdowns)
NUM_OCR_WORKERS	1	Number of WorkerMessageQueue workers handling ocr pipelines (Should be set as 1 until the pipeline is build to handle multiple concurrent request efficiently without slowdowns)
NUM_TSL_WORKERS	1	Number of WorkerMessageQueue workers handling translation pipelines (Should be set as 1 until the pipeline is build to handle multiple concurrent request efficiently without slowdowns)

Project/server variables

Variable	Values	Usage
DJANGO_DEBUG	false[/true]	Whether to run the server in debug (true) or production (false) mode
DJANGO_LOG_LEVEL	INFO	python logging level for
DATABASE_NAME	db.sqlite3	For sqlite3 this is the path to the database file. For other backend it should be the name of the database
DATABASE_ENGINE	django.db.backends.sqlite3	Change this to either a Django or 3rd party provided backend to use another Database type
DATABASE_HOST	optional	Required if using another db back-end
DATABASE_PORT	optional	Required if using another db back-end
DATABASE_USER	optional	Probably required if using another db back-end
DATABASE_PASSWORD	optional	Probably required if using another db back-end

Supported databases

• SQLite This is mostly fine for a self-hosted server accessed by a single or few users (and it's probably gonna be faster than any other database not running on the same network as the server because of latency).
• Postgresql
• MySQL/MariaDB