State-of-the-art Machine Learning research with Google tools
summary
Cloud-managed TPUs scale into TPU Pods that run state-of-the-art training and inference, while T5X provides a modular framework for configurable sequence-model training; an example model has over 11 billion parameters trained on the Colossal Clean Crawled Corpus. The TPU Research Cloud offers free access for researchers. A concrete application, Pathdreamer synthesizes 360° views up to 20 feet from a single observation to infer unseen rooms, is built with TensorFlow and Keras and is open source on GitHub. JAX runs the same NumPy-style code on GPU or TPU and interoperates with TensorFlow data pipelines.
AI-written summary · about
chapters (7)
related talks
42mCloud TPU Pods: AI Supercomputing for Large Machine Learning Problems
23mPyTorch on Google Cloud TPUs
12mAI-powered infrastructure:
13mTaking Machine Learning to the Next Level
28mSupercharge your training on TPUs
20mAI and Machine Learning for developers
27mTips and tricks for distributed large model training
37mEnd-2-End Machine Learning on GCP