Clay

Introduction

Clay is an open-source AI model designed to interpret geospatial and temporal data related to Earth. Developed as a self-supervised Masked Autoencoder (MAE), Clay uses an expanded visual transformer to enhance its understanding of Earth data. The model facilitates generating semantic embeddings, fine-tuning for various downstream tasks, and serving as a backbone for other models.

Architecture

Clay's architecture is based on an advanced visual transformer. This structure is optimized to comprehend complex geospatial and temporal relationships, making it suitable for tasks such as classification, regression, and generative processes.

Training

The Clay model is trained using a self-supervised learning approach, specifically as a Masked Autoencoder (MAE). This method allows the model to learn representations of Earth data effectively without requiring labeled datasets.

Guide: Running Locally

1. Clone the Repository: Obtain the model's code from GitHub.
2. Set Up Environment: Ensure you have the necessary dependencies installed, which are typically outlined in a requirements.txt file.
3. Download Model Weights: Access the pre-trained weights from Hugging Face.
4. Run the Model: Execute the model using your local machine or a cloud-based GPU for better performance. Recommended cloud GPU services include AWS, Google Cloud, and Azure.

License

The Clay model is distributed under the OpenRAIL-M license, which is detailed here. The code is available under the Apache license, and the documentation is licensed under CC-BY.