Quaterion solves common issues with similarity learning. It is thought to be harder than traditional ML, but it doesn't have to. Quaterion is a framework for similarity learning, making you able to run several experiments while you have a cup of coffee.
Hello Everyone!
My name is Kacper, I’m a Developer Advocate at Qdrant. We’re a startup working on making the neural search easier.
Neural embeddings have been on everyone’s lips for a while now, as the models generating the encodings for unstructured data became publicly available, thanks to organizations like Hugging Face. These off-the-shelf networks typically require some additional fine-tuning to adapt them to a specific domain. And that’s where many people start to struggle.
We realized similarity learning doesn’t get adopted by the market as fast as it could. Up till now, this area was thought to be harder than classification or regression, even though the data requirements are much lower, and the results are decent, even in cases where those traditional approaches are not able to help. If you’re dealing with extreme classification or your classes may vary over time, similarity learning is a thing to try out!
Quaterion is a Python framework for similarity learning, based on PyTorch-Lightning. It is designed to combine the performance of pre-trained models with specialization for the custom task while avoiding slow and costly training. Our built-in unique cache mechanism enables you to train thousands of epochs with huge batch sizes even on laptop GPU.
We would love to hear your feedback! Stop by https://github.com/qdrant/quaterion
Quaterion