Dependency Management Strategy

Relevant source files

The following files were used as context for generating this wiki page:

• requirements.txt
• train_model/train_requirements.txt
• setup.py
• README.md
• mb_capcha_ocr/predict.py
• train_model/train.py

The dependency management strategy for this project is bifurcated into two distinct environments: the Inference Environment and the Training Environment. This separation ensures that the lightweight production package remains minimal by utilizing ONNX Runtime, while the resource-intensive training utilities are isolated within a sub-module.

The project uses standard Python packaging tools, specifically setuptools for distribution and requirements.txt files for environment replication.

Environment Separation

The project explicitly distinguishes between the dependencies required to run the OCR model (Inference) and those required to develop or retrain it (Training). This is reflected in the file structure and the instructions provided in the documentation.

Inference Environment

This environment is designed for end-users of the mb_capcha_ocr library. It focuses on a minimal footprint to perform predictions using a pre-trained model.onnx file.

Dependency	Purpose
onnxruntime	Executing the optimized ONNX model
pillow	Image loading and preprocessing (grayscale, resizing)
numpy	Numerical operations on image arrays
setuptools/wheel	Package installation and distribution

Sources: requirements.txt:1-5, setup.py:19-27, README.md:52-56

Training Environment

The training environment requires a more robust stack, including deep learning frameworks and visualization tools. These dependencies are housed separately within the train_model/ directory.

Dependency	Purpose
torch	Core deep learning framework for model architecture
torchvision	Computer vision utilities and ResNet18 base model
matplotlib	Generating training loss plots (loss.png)
onnx / onnxscript	Exporting the PyTorch model to ONNX format

Sources: train_model/train_requirements.txt:1-7, README.md:58-64, train_model/train.py:1-12

Installation Flow

The following diagram illustrates the workflow for setting up the environment based on the user's intent (using the library vs. training the model).

Automated Distribution Configuration

The setup.py file automates the dependency resolution for the distributable package. It dynamically reads the requirements.txt file and includes necessary non-code assets.

• Dynamic Requirements: The script reads requirements.txt into the install_requires parameter.
• Version Tracking: It extracts the package version from mb_capcha_ocr/__init__.py using regular expressions.
• Asset Inclusion: It ensures the pre-trained model.onnx is bundled with the package using package_data.

with open("requirements.txt") as f:
    req = f.read().splitlines()

setup(
    name='mb_capcha_ocr',
    version=version,
    packages=["mb_capcha_ocr"],
    package_data={
        'mb_capcha_ocr': ['model.onnx'],
    },
    install_requires=req,
    include_package_data=True
)

Sources: setup.py:4-5, setup.py:20-33

Summary of Dependency Logic

The project's architecture ensures that the high-level API (OcrModel in predict.py) remains decoupled from the training framework. While train_model/train.py relies on torch, the mb_capcha_ocr/predict.py module only requires onnxruntime, allowing for a lighter deployment in production environments.

Sources: mb_capcha_ocr/predict.py:1-10, train_model/train.py:1-12