ipex-llm
ipex-llm copied to clipboard
[WIP] Chronos: How-to-Guides (How to train forecaster)
Description
Add How-to-Guides, and help users train forecaster.
1. Why the change?
Make our doc more friendly.
2. User API changes
No changed.
3. Summary of the change
Add a doc file docs/readthedocs/source/doc/Chronos/Overview/how_to_train_forecaster_on_one_node.md
4. How to test?
- [ ] Document test
5. New dependencies
We can follow such overview for this page, currently this draft is kind of like an API doc :)
Maybe use a jupyter notebook and transform to html could be a good choice, since we may need to show some output to our users.
Train Forecaster on one node
A paragraph quickly illustrate
-
Why we need to train a forecaster? (to learn the pattern (like the period, scale...) in history data)
-
Chronos only support CPU training (and acceleration on Intel CPU) for training,. Since time series data is not large, CPU training is enough for most cases.
Data preparation
We support 3 data representations, briefly illustrate what they are and a short code-snip how to generate them.
-
TSDataset
-
Numpy Ndarray
-
pytorch dataloader / tf dataset
Training
show them how to use fit
. A short code-snip would do.
validation
show them a code snip how to use "best_epoch" and use a paragraph to illustrate why validation is important.
automatically acceleration
leave this blank to me