(T) I had the opportunity to attend this week IBM Index Developer conference at the Moscone Center. In particular, I attended on Tuesday the Open Community Meetings on TensorFlow (and Spark).
The presentations on TensorFlow were mostly given by Google developer advocates (while the presentations on Spark were given by Databricks and IBM Spark Technology Center in San Francisco).
Google has now internally 100 engineers working full-time on TensorFlow. In addition, 1,000 engineers, outside of Google, are contributing to TensorFlow. With so many contributors to an open source platform, no wonder how much the platform is expanding in so many areas:
TensorFlow Lite provides machine learning inference with low latency and small binary size for mobile, IoT, and embedded devices. It supports as well as hardware acceleration with the Android Neural Networks API. There are many use cases for TensorFlow Lite in particular device intelligence for IoT devices, camera, images, and voice interactions on mobile devices.
T2T integrates to TensorFlow widely used models and datasets for image classifications, language modeling, sentiment analysis, speech recognition, summarization, and translation:
- Datasets: ImageNet, CIFAR, MNIST, Coco, WMT, LM1B…
- Models: ResNet, RevNet, ShakeShake, Xception, SliceNet, Transformer, ByteNet, Neural GPU, LSTM…
- Tools: Cloud training, hyperparameters tuning, TPU…
TensorFlow with R
TensorFlow has now an R interface for the Keras API, Estimator API, and Core API.
KubeFlow aims to integrate TensorFlow to Google Kubernetes for:
- Easy, repeatable, and portable deployments on a diverse infrastructure (laptop, training cluster, and production cluster)
- Deploying and managing microservices
- Scaling based on demand
Note: The picture above is Google’s Cloud Tensor Processing Units (TPUs).
Copyright © 2005-2018 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.
Categories: Artificial Intelligence, Deep Learning, Machine Learning