Google has announced a tool for customizing existing machine learning (ML) models to work with your data and applications. MediaPipe Model Maker provides a faster alternative to building and training a new ML model.

While the overall tool is new, it's actually based on several existing tools, specifically MediaPipe Solutions, TensorFlow Lite Task Library, and TensorFlow Lite Model Maker.

MediaPipe provides a low-code API to create and deploy advanced ML solutions across platforms, that in the new tool has been optimized to go beyond single model inference. MediaPipe also has a Model Maker low-code API to customize solutions using a users own data; and MediaPipe Studio to view and check solutions.

Google says Model Maker uses transfer learning for training ML models. This retrains existing models with new data, meaning a model re-uses a significant portion of the existing model logic, so training takes less time than training a new model, and can be done with less data.

Model Maker works on various types of models including object detection, gesture recognition, and classifiers for images, text, or audio data. The tool retrains models by removing the last few layers of the model that classify data into specific categories, and rebuilds those layers using new data you provide. Model Maker also supports some option to fine tune model layers to improve accuracy and performance.

The MediaPipe team says retraining a model using Model Maker usually makes the model smaller, and can be retrained with significantly less data than training a new model. The recommendation when retraining a model with new data is that you should have around 100 data samples for each trained class.

More Information

Google MediaPipe Model Maker

Related Articles

AI Advances At Google I/O

Google Adds Code Generation To Bard

Google's Large Language Model Takes Control

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.

Comments

Make a Comment or View Existing Comments Using Disqus

or email your comment to: comments@i-programmer.info