Conversation
|
AllenBaranov seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. You have signed the CLA already but the status is still pending? Let us recheck it. |
sarahmish
left a comment
sarahmish
left a comment
There was a problem hiding this comment.
@abaranov25 seems like Chronos2 pipeline did not work well on S-1 example, do you have an idea why? I'm curious if you use S-1-test would you get different results?
| }, | ||
| "mlstars.custom.timeseries_preprocessing.rolling_window_sequences#1": { | ||
| "target_column": 0, | ||
| "window_size": 256 |
There was a problem hiding this comment.
I'm curious why 256? Orion pipelines that are forecast based are usually set to 250. For timesfm, it was set to 256 because it needs to be a multiple of 32 according to their docs.
| "init_params": { | ||
| "mlstars.custom.timeseries_preprocessing.time_segments_aggregate#1": { | ||
| "time_column": "timestamp", | ||
| "interval": 600, |
There was a problem hiding this comment.
We typically set the example to interval=21600 since our example code uses NASA signals (MSL and SMAP)
orion/primitives/chronos2.py
Outdated
| # if sys.version_info < (3, 11): | ||
| # msg = ( | ||
| # '`timesfm` requires Python >= 3.11 and your ' | ||
| # f'python version is {sys.version}.\n' | ||
| # 'Make sure you are using Python 3.11 or later.\n' | ||
| # ) | ||
| # raise RuntimeError(msg) |
There was a problem hiding this comment.
If chronos works on python 3.9+ remove this comment completely
orion/primitives/chronos2.py
Outdated
| # raise RuntimeError(msg) | ||
|
|
||
| try: | ||
| from chronos import Chronos2Pipeline |
There was a problem hiding this comment.
add chronos to setup.py under the pretrained_requires section, then remove the try-catch since it will always be installed using pip install orion[pretrained]
orion/primitives/chronos2.py
Outdated
| if __name__ == "__main__": | ||
| chronos2 = Chronos2() | ||
| X = np.random.rand(100, 256, 10) | ||
| y = chronos2.predict(X) | ||
| print(y.shape) | ||
| print(y) No newline at end of file |
sarahmish
left a comment
sarahmish
left a comment
There was a problem hiding this comment.
@abaranov25 It would be nice to run the notebook locally then commit it so that the output is rendered into the notebook , then users can look at the github and immediately see the results of Chronos 2
tutorials/pipelines/chronos2.ipynb
Outdated
There was a problem hiding this comment.
This line needs to be updated
There was a problem hiding this comment.
Is there a specific reason why we are not using the timestamps we have from the data? Every input to orion has two columns: timestamp, values. You can use the first index of timestamp to set start_time.
If Chronos has been trained on major events, then timestamps can be valuable to know them, e.g. thanksgiving.
There was a problem hiding this comment.
I see, makes sense
| "mlstars.custom.timeseries_preprocessing.time_segments_aggregate#1": { | ||
| "time_column": "timestamp", | ||
| "interval": 21600 | ||
| } |
There was a problem hiding this comment.
This needs to be adjusted for all datasets
| } | |
| }, | |
| "orion.primitives.chronos2.Chronos2#1": { | |
| "time_interval": 21600 | |
| } |
There was a problem hiding this comment.
This line needs to be updated "The UniTS contains 6 primitives." -> "The Chronos2 pipeline contains 6 primitives."
| "mlstars.custom.timeseries_preprocessing.rolling_window_sequences#1": { | ||
| "target_column": 0, | ||
| "window_size": 250 | ||
| }, |
There was a problem hiding this comment.
| }, | |
| }, | |
| "orion.primitives.chronos2.Chronos2#1": { | |
| "time_interval": 21600 | |
| }, |
3 participants
Resolve #662
Adding Chronos-2 as a new pipeline on Orion