Google announced a new open source system Monday that could speed the process for creating and training machine learning models within the firm’s TensorFlow library. Tensor2Tensor (T2T), unveiled via a blog post, is geared toward creating deep learning models in particular, and can be used for a variety of purposes.
T2T can be used to build models for processes such as text translation or parsing, as well as image captioning, the post said. It also allows users to create these models and explore their ideas “much faster than previously possible,” the post noted.
One of the main goals of T2T seems to be lowering the barrier to entry for users looking to experiment with deep learning and compare their findings against other work in the field. To help users get started, Google has also released a library of datasets, models, and configurations that offer insights with less foundational engineering work.
SEE: The Machine Learning and Artificial Intelligence Bundle (TechRepublic Academy)
In some experiments run by Google, and noted in the post, two T2T models beat out the GNMT+MoE, which was previously considered to be state-of-the-art. T2T is also able to accomplish this in a relatively short period of time. “Notably, with T2T you can approach previous state-of-the-art results with a single GPU in one day,” the post said.
Google used existing TensorFlow tools to build T2T, and the system will work to define which pieces a user may need to build their deep learning system. It also utilizes a standard interface among all aspects of a deep learning system, including datasets, models, optimizers, and different sets of hyperparameters, the post said. So, users can swap versions of these components out to see how they perform together. It is this modular architecture that is one of the core values of T2T.
Because of its modular nature, T2T isn’t confined to just one model or dataset, and architectures can be “defined in a few dozen lines of code,” the post said. Models can also be trained on multiple tasks from different domains with T2T. In one example cited by Google, a single deep learning model was successfully able to perform three distinct tasks at once.
Additionally, Google has built best practices into the T2T system. So, the configurations and “tricks of the trade” that worked well for Google in its research will be available to T2T users.
More information can be found on GitHub.
The 3 big takeaways for TechRepublic readers
- Google’s Tensor2Tensor (T2T) could make it faster and easier to train deep learning models on TensorFlow.
- T2T could lower the barrier for organizations looking to experiment with deep learning, offering a modular architecture that is easy to work with.
- Google is also including tips, tricks, and best practices along with the T2T library, to help users on their way to creating and using machine learning models.