ashrot/astra-clr-base

---

ASTRA: Model Checkpoints & ONNX Exports

This repository hosts the pre-trained and fine-tuned model weights for the ASTRA (Attention-based Self-supervised Time-series Representation Architecture) framework. It includes the original TensorFlow/Keras model weights as well as optimized ONNX formats designed for lightweight production inference.

🏆 Support & Recognition

This research is supported by the NVIDIA Academic Grant Program (2025–2026). The project utilizes NVIDIA's high-performance compute resources to architect foundational-scale SSL frameworks for the next generation of time-domain astronomy.

---

🌐 The ASTRA Ecosystem

The ASTRA project is split across four distinct repositories:

Repository	Description	Link
Development Code	Core framework, data pipeline, and model training.	GitHub: astra
Inference Code	Lightweight, production-ready inference scripts and utilities.	GitHub: astra-infer
Astronomy Dataset	Light curve datasets formatted for ASTRA.	Hugging Face: Dataset
Model Weights & ONNX	Pre-trained & finetuned model checkpoints along with ONNX exports (This Repo).	[Current Repository]

---

📂 Repository File Structure

The model weights in this repository are organized as follows:

├── tf_weights.h5          # Original TensorFlow/Keras HDF5 model weights
└── onnx/                  # Optimized ONNX model files for inference
    └── astra-clr.onnx
    └── astra-clr_ft_v1.onnx    

The files in the onnx/ are designed to be consumed directly by the astra_infer package.

• tf_weights.h5: The raw model weights exported from the TensorFlow-based ASTRA training pipeline. These weights are ideal for resuming pre-training or performing fine-tuning.
• onnx/astra-clr.onnx: The pre-trained onnx model
• onnx/astra-clr_ft_v1.onnx: The fine-tuned onnx model

---