v1.2.6

google-research/timesfmv1.2.6Dec 31, 2024by rajatsen91

AI Summary

This release introduces support for TimesFM-2.0 models with new 500m checkpoint variants available in both JAX and PyTorch formats. It also sets the median head as the default point forecaster and includes bug fixes for PyTorch decoding, though cached decoding is not yet supported in either backend.

Key Highlights

Added support for TimesFM-2.0 models
Median head set as the default point forecaster
New TimesFM-2.0-500m checkpoints available (JAX and PyTorch)
Bug fixes in PyTorch decoding
No cached decoding support in both JAX and PyTorch

New Features

TimesFM-2.0 model support with hparam configuration
Median head as default point forecaster
TimesFM-2.0-500m-jax checkpoint
TimesFM-2.0-500m-pytorch checkpoint

Full Release Notes

Changes:
----
1. Add support for TimesFM-2.0 models.
2. Set the median head as the default point forecaster.

PyPI Release:
----
v1.2.6: Support for TimesFM-2.0 models.
- Add hparam support for TimesFM-2.0 models.
- Some bug fixes in pytorch decoding.
- Right now we do not support cached decoding in both jax and pytorch.

Checkpoints:
----
The TimesFM-2.0 checkpoints are available on Hugging Face:
- https://huggingface.co/google/timesfm-2.0-500m-jax
- https://huggingface.co/google/timesfm-2.0-500m-pytorch

Full Changelog:
----
https://github.com/google-research/timesfm/commits/v1.2.6