This project implements an OCR system using ResNet34, Mamba, and CTC. It uses adapter-based fine-tuning to train efficiently on the vision backbone.

1. Preprocessing: Images are resized to 32px height, preserving aspect ratio, and padded.
2. Vision Backbone: Uses a pre-trained ResNet34 with injected adapters. freezes most of the network and only train the adapters, partial layer4, and projections.
3. Mamba Encoder: Models the sequence of features to understand character context.
4. Decoding: Uses CTC to convert the sequence into text.

• Mamba: Uses state space models instead of RNNs for better scaling.
• Adapters: Fine-tunes the backbone efficiently without breaking pre-trained weights.
• ResNet34: Strong baseline feature extractor.
• FP16: Supports mixed precision training.

• configs/: Configuration settings.
• data/: Dataset loading and processing.
• models/: Model architecture (ResNet + Mamba).
• train.py: Main training script.
• infer.py: Inference script.
• utils.py: Metrics and decoding helpers.
• gen_data.py: Synthetic data generator.

Clone the repo and install dependencies:

git clone <repo_url>
cd ocr_project
pip install -r requirements.txt

Note: mamba-ssm needs a GPU.

Check configs/config.py to set your data paths and batch size.

Run the training script (it handles freezing automatically):

python train.py

Test on images:

python infer.py

Use gen_data.py to create synthetic training data with random fonts and noise.