your123 - stock.adobe.com
Intending to build open source software system for time series forecasting to address challenges in forecasting, deep learning company Nixtla turned to PyTorch.
PyTorch is an open source framework for machine learning. Based on the Torch library, it uses computer vision and natural language processing.
In need of a framework that would speed up the implementation of new models in its library and the ability to test new time forecasting products quickly, Nixtla is turning to PyTorch Lightning, said CTO and co-founder Federico Garza Ramirez.
"Testing a new product is as fast as just developing it from our site," they said. "It's a lot easier to share the new product to some potential customers."
Unlike the original PyTorch, PyTorch Lightning lets users quickly scale their machine learning models and build more AI models.
On March 15, Lightning AI made PyTorch Lightning 2.0 generally available to customers.
PyTorch Lightning introduces a stable API and smaller footprint as well as makes it easier for customers to read and debug models.
The debugging feature is vital to Nixtla.
"It is really difficult to develop deep learning models because everything could break. But with these new features, it could be easy to debug them," Ramirez said.
Other than Lightning 2.0, Lightning introduced Lightning Fabric. Fabric gives users control of their training loop and lets them to use tools like callbacks and checkpoints. It also supports reinforcement learning, active learning and transformers while still providing complete control of the training code, according to Lightning AI.
"That's one of the features I'm more interested in," Ramirez said. "I think it would be a lot easier to build new deep learning models."
While Nixtla plans to use PyTorch Lightning 2.0 within the next few days, Ramirez said the company would first need to adapt previous codes to support the new version. Nixtla considers PyTorch Lightning a game changer, but the forecasting company sees room for improvement.
Federico Garza RamirezCTO, Nixtla
One is seeing more tutorials on implementing some of Lightning's features, Ramirez said.
"In particular, how to distribute tasks among the different clusters [such as] a GPU cluster or a TPU cluster," they said. He added that it would also be interesting to see tutorials on using Lightning to handle big datasets.
"In the field of time series forecasting, there is this tension between a statistical model, or deep learning model," Ramirez continued. "We have this intuition that the larger the datasets, the better the performance of the deep learning model. We really want to test that hypothesis."
"It simplifies a lot of the of the work required to run deep learning models," Ramirez said. "With a few lines of code -- compared with traditional PyTorch implementation -- you can start running deep learning models into production."
Esther Ajao is a news writer covering artificial intelligence software and systems.