Facebook Releases Another Deep Learning Tool
Facebook has released to open-source developers a deep learning framework that does double duty for computer vision and language tasks.
Built on its PyTorch framework, Facebook is positioning Pythia as supporting “multitasking” for vision and language “multimodal AI models.” Among its capabilities are identifying visual data and automatically generating image captions, company developers said this week.
Pythia also is designed to allow developers to build, reproduce and benchmark AI models.
Facebook said Pythia incorporates elements of winning entries in recent AI competitions, including reference implementations showing how previous models achieved their benchmark results. Another reference allows developers to gauge the performance of new models.
The new framework also supports distributed training and a range of data sets.
“Pythia smooths the process of entering the growing subfield of vision and language and frees researchers to focus on faster prototyping and experimentation,” Facebook said. “Our goal is to accelerate progress by increasing the reproducibility of these models and results.”
Another goal is spurring research that would make current “brittle” AI models more adaptive by synthesizing “multiple kinds of understanding into a more context-based, multimodal understanding,” the company added.
Along with the open-source code release to GitHub, Facebook said it would also contribute additional tasks, tools, data sets and reference models.
Basing Pythia on the PyTorch further extends the reach of the emerging deep learning framework that is currently challenging TensorFlow for primacy among developers. According to the release notes, PyTorch serves as the foundation for the “modular [Pythia] framework for supercharging vision and language research….”
Facebook earlier released image and natural language processing libraries along with an AI reinforcement learning platform called Horizon.