8000 [ENH] Interfacing `chronos-bolt` in `sktime` · Issue #7680 · sktime/sktime · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

[ENH] Interfacing chronos-bolt in sktime #7680

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
PranavBhatP opened this issue Jan 22, 2025 · 7 comments · Fixed by #7718
Closed

[ENH] Interfacing chronos-bolt in sktime #7680

PranavBhatP opened this issue Jan 22, 2025 · 7 comments · Fixed by #7718
Assignees
Labels
enhancement Adding new functionality module:forecasting forecasting module: forecasting, incl probabilistic and hierarchical forecasting

Comments

@PranavBhatP
Copy link
Contributor

Is your feature request related to a problem? Please describe.
The Chronos model has already been implemented in sktime as ChronosForecaster. Recently, an improved version of Chronos was developed namely Chronos-Bolt. The model has been to proven to have faster inference and better efficiency than the pre-exisiting Chronos model.
Refer blog

Describe the solution you'd like
Interfacing Chronos-Bolt seems to be a good option, although I'm not sure about whether this would be mean replacing the pre-exisiting ChronosForecaster or a standalone implementation?

Additional context
The need for the same was also expressed in the conversation linked to the below message.
Reference issue: #1736
sktime/pytorch-forecasting#1736 (comment)

@PranavBhatP PranavBhatP added the enhancement Adding new functionality label Jan 22, 2025
@PranavBhatP
Copy link
Contributor Author

I would like to work on this issue, @fkiraly.

@fkiraly
Copy link
Collaborator
fkiraly commented Jan 22, 2025

Sure!

Quick question, is this simply a different checkpoint that can be run with the current architecture, or does it require also different code?

A link to the hf weights would be appreciated.

@fkiraly fkiraly added the module:forecasting forecasting module: forecasting, incl probabilistic and hierarchical forecasting label Jan 22, 2025
@PranavBhatP
Copy link
Contributor Author

This is the one currently being interfaced:
https://huggingface.co/amazon/chronos-t5-tiny

The one I plan to interface:
https://huggingface.co/autogluon/chronos-bolt-tiny

@fkiraly
Copy link
Collaborator
fkiraly commented Jan 22, 2025

Do you know the answer to this?

Quick question, is this simply a different checkpoint that can be run with the current architecture, or does it require also different code?

@PranavBhatP
Copy link
Contributor Author
8000
PranavBhatP commented Jan 22, 2025

No, it is not a checkpoint. It seems to have a slightly different architecture over the original chronos-t5.

@PranavBhatP
Copy link
Contributor Author
PranavBhatP commented Jan 23, 2025

Hi, before I proceed with a PR, I wanted to confirm whether I should interface Chronos-Bolt as a standalone model in addition to the existing Deep Learning Based Forecasters? I'm still not clear in that aspect and wanted to know if this approach is relevant?

@geetu040
Copy link
Contributor

Hi, before I proceed with a PR, I wanted to confirm whether I should interface Chronos-Bolt as a standalone model in addition to the existing Deep Learning Based Forecasters? I'm still not clear in that aspect and wanted to know if this approach is relevant?

From the documentation amazon-science/chronos-forecasting we know that BaseChronosPipeline can be used to work with both the models, So I would suggest to make changes in the existing ChronosForecaster class in sktime to add support for bolt models as well

@fkiraly fkiraly moved this from Todo to In Progress in Dec 2024 - Mar 2025 mentee projects Jan 27, 2025
benHeid pushed a commit that referenced this issue Mar 20, 2025
)

#### Reference Issues/PRs
Fixes #7680 

#### What does this implement/fix? Explain your changes.
The PR extends the existing `ChronosForecaster` class to support the
latest `chronos-bolt` family of models, which operators upon the same
`BaseChronosPipeline` as `chronos`.

Links:
[Chronos-Bolt-Tiny](https://huggingface.co/amazon/chronos-bolt-tiny)

---------

Co-authored-by: PranavBhatP <pranavbhat2004@gmai.com>
Co-authored-by: Armaghan Shakir <raoarmaghanshakir040@gmail.com>
Spinachboul pushed a commit to Spinachboul/sktime that referenced this issue Mar 23, 2025
…time#7718)

#### Reference Issues/PRs
Fixes sktime#7680 

#### What does this implement/fix? Explain your changes.
The PR extends the existing `ChronosForecaster` class to support the
latest `chronos-bolt` family of models, which operators upon the same
`BaseChronosPipeline` as `chronos`.

Links:
[Chronos-Bolt-Tiny](https://huggingface.co/amazon/chronos-bolt-tiny)

---------

Co-authored-by: PranavBhatP <pranavbhat2004@gmai.com>
Co-authored-by: Armaghan Shakir <raoarmaghanshakir040@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
enhancement Adding new functionality module:forecasting forecasting module: forecasting, incl probabilistic and hierarchical forecasting
Development

Successfully merging a pull request may close this issue.

3 participants
0