titans-unofficial 1.0.5

Unofficial PyTorch implementation of Titans: Learning to Memorize at Test Time

Unofficial Implementation of Titans: Learning to Memorize at Test Time

This is an unofficial PyTorch implementation of the paper "Titans: Learning to Memorize at Test Time" by Ali Behrouz, Peilin Zhong, and Vahab Mirrokni.

Overview

Titans is a novel neural architecture that combines attention-based short-term memory with a neural long-term memory module. The architecture addresses the limitations of both recurrent models (which compress data into fixed-size memory) and attention mechanisms (which have quadratic complexity).

Key Features

• Neural Long-term Memory: A module that learns to memorize historical context
• Persistent Memory: Learnable tokens that encode task-specific knowledge
• Three Architectural Variants:
  • MAC (Memory as Context): Uses memory as context for attention
  • MAG (Memory as Gate): Combines memory with core branch using gating
  • MAL (Memory as Layer): Integrates memory as a separate layer

Installation

From PyPI (Recommended)

pip install titans-unofficial

From Source (Development)

# Clone the repository
git clone https://github.com/Shehryar718/titans-unofficial.git
cd titans-unofficial

# Install in development mode with all extras
pip install -e ".[dev]"

Requirements

• Python 3.10+
• PyTorch 2.0+
• transformers (for tokenization)
• numpy
• pytest (for running tests)

Project Structure

titans-unofficial/
├── titans/
│   ├── __init__.py
│   ├── models/
│   │   ├── titans_base.py    # Base class for all variants
│   │   ├── titans_mac.py     # Memory as Context implementation
│   │   ├── titans_mag.py     # Memory as Gate implementation
│   │   └── titans_mal.py     # Memory as Layer implementation
│   └── utils/
│       ├── memory.py         # Neural Memory Module
│       ├── attention.py      # Attention mechanisms
│       └── persistent_memory.py  # Persistent Memory implementation
├── examples/
│   ├── text_classification.py  # Text classification example
│   ├── language_modeling.py    # Language modeling example
│   └── fine_tuning.py         # Fine-tuning example
├── pytests/
│   └── test_memory.py         # Tests for memory module
├── requirements.txt
├── LICENSE
└── README.md

Usage

Text Classification

from titans import TitansMAC, TitansMAG, TitansMAL
from examples.text_classification import TitansForClassification

# Initialize model
model = TitansForClassification(
    vocab_size=30000,
    d_model=128,
    n_layers=2,
    n_heads=4,
    num_classes=2,
    memory_depth=2,
    persistent_tokens=8,
    window_size=16,
    model_type="mal"  # Choose from: "mac", "mag", "mal"
)

# Train and evaluate
python examples/text_classification.py

Language Modeling

from titans import TitansMAC, TitansMAG, TitansMAL
from examples.language_modeling import TitansForLanguageModeling

# Initialize model
model = TitansForLanguageModeling(
    vocab_size=30000,
    d_model=128,
    n_layers=2,
    n_heads=4,
    memory_depth=2,
    persistent_tokens=16,
    window_size=128,
    model_type="mac"
)

# Train and generate text
python examples/fine_tuning.py

Architecture Details

Neural Memory Module

The neural memory module consists of:

• Key/Value/Query projections for memory access
• Multi-layer perceptron for memory processing
• Momentum-based update mechanism with configurable parameters
• Weight decay for forgetting mechanism
• Gradient scaling for numerical stability

Variants

1. MAC (Memory as Context)

   • Memory output serves as additional context
   • Efficient for tasks requiring long-range dependencies
   • Parallel processing with chunked computation
   • Configurable chunk size and parallel processing

2. MAG (Memory as Gate)

   • Gating mechanism to combine memory with core processing
   • Adaptive balance between short and long-term memory
   • Enhanced numerical stability
   • Improved gradient flow through gating

3. MAL (Memory as Layer)

   • Memory integrated as a separate layer
   • Direct memory access at each layer
   • Sliding window attention for efficiency
   • Layer-wise memory updates

Example Tasks

The repository includes implementations for:

• Text Classification (Binary and multi-class)
• Language Modeling with test-time adaptation
• Fine-tuning with early stopping

Each example demonstrates different aspects of the Titans architecture:

• Memory reset between epochs for fresh adaptation
• Efficient batch processing with dynamic batching
• Gradient scaling for numerical stability
• Early stopping and model checkpointing
• Proper memory state management

Testing

# Run all tests
pytest pytests/

# Run specific test file
pytest pytests/test_memory.py

Citation

This repository provides an unofficial implementation of the Titans architecture.
If you reference this work, please cite the original paper:

@article{behrouz2024titans,
  title={Titans: Learning to Memorize at Test Time},
  author={Behrouz, Ali and Zhong, Peilin and Mirrokni, Vahab},
  journal={arXiv preprint arXiv:2501.00663},
  year={2024}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.