This project is an experiment to find the most effective model architecture for a 12-digit number recognition (OCR) task. We explore three different approaches, evolving from a simple linear model to a more complex Convolutional Recurrent Neural Network (CRNN), demonstrating why specialized architectures are critical for sequence-based problems. Kaggle Link

• Presentation

The experiment was conducted by building and training three distinct models on the same dataset.

This model serves as our baseline. It's a simple Multi-Layer Perceptron (MLP) that takes the flattened image as input.

• Architecture: havoc/registry/modelRegistry/MCLv1
• Result: This model failed to learn. By flattening the image, it destroys all spatial and sequential information. It cannot tell the 1st digit from the 12th. As seen in the logs, its accuracy plateaus at ~10%.

This model introduces Convolutional Neural Networks (CNNs) to extract meaningful spatial features before classification.

• Architecture: havoc/registry/modelRegistry/MCCv1
• Result: This model struggles but eventually learns. After being stuck at ~10% for over 20-30 epochs, it starts to "memorize" the scrambled feature patterns, peaking at ~78.08% accuracy.

This specialized architecture, commonly known as a CRNN (Convolutional Recurrent Neural Network).

• Architecture: havoc/registry/modelRegistry/MCCLv1
• Result: This model is highly effective. It starts learning almost immediately and achieves a peak validation accuracy of 99.40%.

This project uses uv for fast Python package management.

• You must have uv installed.

Follow these steps to create the virtual environment and reproduce the results.

1. Clone the repository:

   git clone https://github.com/SuriyaaMM/havoc
   cd havoc

2. Create and sync the virtual environment:

   uv venv
   uv sync

3. Activate the environment and run an experiment:

   source .venv/bin/activate
   uv run -m havoc.experiment.Ev3