Abdidechasa/Amharic-MMOCR

Custom Amharic OCR using a lightweight MMOCR framework. Highly optimized for CPU-only environments featuring mini-batch configurations, gradient accumulation settings, and complete Ethiopic script token support. # Amharic-MMOCR Custom Amharic OCR using a lightweight MMOCR framework. Highly optimized for CPU-only environments featuring mini-batch configurations, gradient accumulation settings, and complete Ethiopic script token support. # 🇪🇹 Mini MMOCR for Amharic A lightweight, CPU-optimized Optical Character Recognition (OCR) pipeline built on OpenMMLab's MMOCR framework, specifically tailored for the Amharic language (Ge'ez/Ethiopic script). This project downscales heavy architectures into a "mini" variant designed to run efficiently on commodity CPU hardware without causing system freezes. --- ## ✨ Features * **Complete Script Support:** Configured dictionary containing the 300+ primary Amharic characters (including punctuation and numbers). * **CPU-Optimized Pipelines:** Low memory footprint, optimized thread configurations, and micro-batch pipelines to prevent VRAM dependency. * **End-to-End OCR:** Modular structure featuring lightweight text detection (e.g., DBNet) paired with an efficient text recognizer. --- ## ⚙️ CPU Training & Performance Tweaks Training deep learning models on a CPU requires careful resource balance. This repository is configured by default with the following hardware protections: * **Micro-Batching:** Configured with a `batch_size` of `2` or `4` to prevent main memory caching bottlenecks. * **Gradient Accumulation:** Uses accumulated steps to simulate standard larger batch sizes (`effective_batch_size = 32`) without increasing processing overhead. * **Thread Capping:** Codebase automatically limits multi-threading workers to prevent 100% system thread locking. --- ## 🚀 Quick Start ### 1. Prerequisites & Installation Ensure you have Python 3.8+ installed. It is highly recommended to build this inside a clean virtual environment or Conda environment. ```bash # Clone the repository git clone github.com cd amharic-mini-mmocr # Install OpenMMLab base tools & dependencies pip install -U open …