NEWS.md

In business-science/modeltime.gluonts: 'GluonTS' Deep Learning

modeltime.gluont 0.3.1

Installation Support

Windows Conflicting Dependencies

Improved support for conflicting package dependencies on Windows Operating Systems. Solution is to separate the installation process into two stages, which happens inside of install_gluonts(). #32

pytorch-lightning 1.3.8 depends on numpy>=1.17.2
mxnet 1.7.0.post1 depends on numpy<1.17.0 and >=1.8.2

New Uninstall Function

Users can now uninstall_gluonts().

modeltime.gluonts 0.3.0

Support for GluonTS 0.8.0 and Pytorch Backend:

Modeltime GluonTS now support gluonts 0.8.0. Simply run install_gluonts() to upgrade. The upgraded support makes modeltime.gluonts incompatible with earlier versions of GluonTS (e.g. gluonts 0.6.3). The solution is to upgrade to gluonts 0.8.0, which requires:

• gluonts==0.8.0

• mxnet~=1.7

Additionally, GluonTS 0.8.0 now supports pytorch as a backend. Use install_gluonts(include_pytorch = TRUE) to simplify installation of the PyTorch backend. Pytorch backend requirements:

• torch~=1.6.0

• pytorch-lightning~=1.1

New Algorithms

Pytorch DeepAR

A new engine has been added to deep_ar() that enables the Pytorch backend using set_engine("torch"). This requires the Python packages pytorch and pytorch-lightning. Use install_gluonts(include_pytorch = TRUE) to simplify installation.

GP Forecaster Algorithm

A new function, gp_forecaster(), integrates the Gaussian Process Estimator from GluonTS.

Deep State Algorithm

A new function, deep_state(), integrates the Deep State Estimator from GluonTS.

Tutorials

• We've updated the Installation Guide. This includes revised requirements for installation, upgrading to modeltime.gluonts >= 0.3.0, troubleshooting installation, python environment requirements, and custom python environments.

• We've updated the Getting Started Guide to go through a DeepAR example.

• We've update the GPU Setup Instructions to cover Modeltime >=0.3.0.

Improvements

• install_gluonts(): Gains two new parameters to help with upgrading:
  1. fresh_install: If TRUE, will remove prior installations of the r-glounts conda environment to setup for a fresh installation. This can be useful if errors appear during upgrades. Default: FALSE.
  2. include_pytorch: If TRUE, will install torch. Needed for Torch implementation of deep_ar(). Default: FALSE.

Breaking Changes

• GluonTS <= 0.8.0. The modeltime.gluonts package version >= 0.2.2.9000 is not compatible with gluonts < 0.8.0. To fix, simply upgrade to gluonts 0.8.0.

modeltime.gluonts 0.2.2

Dials Params

• NBEATS Models: Adding Dials helpers #14

modeltime.gluonts 0.2.1

Improvements made to connect with the GluonTS Python Environment on Startup.

modeltime.gluonts 0.2.0

New Vignettes

• Managing GluonTS Environments

• Using GPUs with GluonTS

New Features

• Internal Scaling by Group: After significant testing it appears that some data sets return better results when the data is scaled by time series "id" (group). To help facilitate this, a new option is available scale by id: scale = TRUE.

• Custom Python Environments: Provide an option for setting a Custom Python Environment by supplying a GLUONTS_PYTHON environment variable. Before running library(modeltime.gluonts) use Sys.setenv(GLUONTS_PYTHON = 'path/to/python') to set the path of your python executable in a Conda or Virtual Environment that has 'gluonts', 'mxnet', 'numpy', 'pandas' and 'pathlib' available as dependencies.

Fixes & Improvements

• GluonTS 0.6.3 Upgrade: install_gluonts() now uses gluonts==0.6.3. This upgrade improves forecast accuracy.
• CRAN Comment - Add SystemRequirements: GluonTS.
• CRAN Comment - Fix .onLoad message to provide options for configuring the Python Environment.

modeltime.gluonts 0.1.0

• Models: Initial Release incorporates 2 GluonTS Algorithms:

  • deep_ar(): Integrates GluonTS DeepAREstimator
  • nbeats(): Integrates N-BEATS & N-BEATS Ensemble Estimators

• New Vignette: Getting Started

• Website: Modeltime GluonTS

business-science/modeltime.gluonts documentation built on Jan. 20, 2024, 3:59 a.m.