PatchTSMixer tutorial broken. ForecastDFDataset API change

[emanuelshalev]

in the tutorial this is how ForecastDFDataset is used:

train_dataset = ForecastDFDataset( time_series_processor.preprocess(train_data), id_columns=id_columns, timestamp_column="date", input_columns=forecast_columns, output_columns=forecast_columns, context_length=context_length, prediction_length=forecast_horizon, )

but input_columns and output_columns are no longer valid parameters. What is the correct usage?

[vijaye12]

We have deprecated ForecastDataset API and using TSP for dataset generation.

you can check the example here https://colab.research.google.com/github/IBM/tsfm/blob/tutorial/notebooks/tutorial/ttm_tutorial.ipynb

Also, you can also try using TTM (which is an advanced version of PatchTSMixer). Here is the model card https://huggingface.co/ibm-granite/granite-timeseries-ttm-v1

[millen11]

I am trying to use the regular PatchTST model per this documentation :https://huggingface.co/blog/patchtst. I run into the same problems as the original issue and have tried the links provided as a solution, however that had more issues. Any insight on a fix?

[osanseviero]

cc @kashif

[etpereira]

Same problem here... Could anyone help?

[kashif]

thanks for the report @etpereira let me have a look again

[etpereira]

@kashif , I was just looking at this link https://github.com/ibm-granite/granite-tsfm/blob/main/notebooks/hfdemo/patch_tsmixer_getting_started.ipynb

And discover a mode of doing it work:

The correct arguments are:

tsp.preprocess(train_data),
id_columns=id_columns,
target_columns=forecast_columns,
context_length=context_length,
prediction_length=forecast_horizon,

Thank you, anyway.