Posted On: Jul 13, 2018
Amazon SageMaker announces several enhancements to the built-in DeepAR, BlazingText, and Linear Learner algorithms. Chainer 4.1 is now supported on the pre-configured containers within Amazon SageMaker.
DeepAR is typically used for forecasting in use cases such as improved supply chain with better product demand forecasts. Many data sets come with incomplete data leading to incorrect forecasts. With DeepAR in SageMaker, missing values are now handled within the model, making forecasting easier and more accurate by utilizing the recurrent neural networks (RNN) model. The second enhancement for DeepAR algorithm is the ability to support custom time-varying features such as seasonality patterns that vary at different levels of a hierarchy across different time series. Third, DeepAR supports grouping of time series with multiple attributes, also known as multiple groupings. With this enhancement, DeepAR can learn group-specific behavior such as seasonality patterns for better forecasts. Lastly, a new notebook that can show how to process a real-world dataset on Amazon SageMaker with DeepAR has been released. This dataset consists of the hourly electricity consumption of 370 customers that has been used academic publications such as “DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks”. For more information on DeepAR in Amazon SageMaker, see the documentation here.
BlazingText provides an optimized implementation of the Word2Vec algorithm to take advantage of GPU hardware. The algorithm learns high-quality distributed vector representations of words in a large collection of documents. This is used in natural language processing (NLP) tasks such as sentiment analysis and entity recognition. The first enhancement with BlazingText in SageMaker enables generation of meaningful vectors for out-of-vocabulary (OOV) words that do not appear in the training dataset. Second, high-speed multi-class and multi-label text classification is supported with BlazingText. The goal of text classification is to automatically classify the text documents into one or more defined categories. BlazingText can now train a text classification model on more than a billion words in a couple of minutes. For more details on BlazingText in Amazon SageMaker, see the documentation here.
The Linear Learner algorithm in Amazon SageMaker now supports multi-class classification, in addition to binary classification and linear regression. This is a task where the outputs are known to be in a finite set of labels. As an example, emails could be classified as inbox, work, personal etc. Linear Learner can now be used for such datasets. Details of Linear Learner can be found here.
Amazon SageMaker’s pre-configured containers now support Chainer 4.1. A key feature in this version is Layer-wise Adaptive Rate Scaling (LARS) that allows you to train networks with large batch sizes.
All these enhancements are now available in Amazon SageMaker in the US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland), Europe (Frankfurt), Asia Pacific (Tokyo), Asia Pacific (Seoul), and Asia Pacific (Sydney) AWS regions.