Changelog
Source:NEWS.md
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:-
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. -
include_pytorch
: If TRUE, will install torch. Needed for Torch implementation of deep_ar(). Default: FALSE.
-
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 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 runninglibrary(modeltime.gluonts)
useSys.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 usesgluonts==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
CRAN release: 2020-11-30
-
Models: Initial Release incorporates 2 GluonTS Algorithms:
New Vignette: Getting Started
Website: Modeltime GluonTS