Pipelines of TSForecaster forces runtime error when trying to run them on new data

[laiki]

Hi,

base on the example 15_PatchTST_a_new_transformer_for_LTSF I used a TSShrinkDataFrame instance + other usefull objects in the preprocessing pipeline of my TSForecaster instance.
Unfortunatelly the example do not work caused by issue #861 (url for data download is wrong)

I stored the TSForecaster after the fit_one_cycle call (n_epochs set to one, just to get it running at least once) by its export function.
After reading it back, as shown in the mentioned example in chapter 'Inference - Preprocess dataframe' running the loader.transform(df) call fails.

learn.transform(valid_df)
...
  File ~/anaconda3/envs/py311/lib/python3.11/site-packages/tsai/data/preprocessing.py:1059 in transform
    X.loc[:, self.columns] = X[self.columns].astype(self.dt)

AttributeError: 'TSShrinkDataFrame' object has no attribute 'dt'

I removed the TSShrinkDataFrame object from the first pipeline and was able to perform the transform function of the 1st pipeline by
prep1 = learn.pipelines[0].transform(valid_df)

Unfortunatelly the 2nd pipeline using a TSStandardScaler instance fails as well

prep2 = learn.pipelines[1].transform(prep1)
...
TypeError: 'NoneType' object is not iterable

The message referes to tsai/data/preprocessing.py:1576 in transform
The object variables "columns", "mean" and "std" seem to be None.
From my point of view, loading the TSForecaster with
load_learner(...) or saving it with learn.export(...) is not making use of the parameter written by a fit/fit_transform call of the pipelines.
Might I need to store and restore the learner object a different way to get all it's parameters restored correctly?
Do I need to use a different way to get such pipelines loadable and usable to prepare the new data for the use with a trained model stored on the disk?

[laiki]

I identified, that I stored the pipelines before I fitted them :(