[BUG] Calling save() after fit() throw pickle error for FFT models

[fabioprocaci]

Describe the bug
save() method does not work after fit() method in some cases.
AttributeError: Can't pickle local object 'FFT.fit.<locals>.<lambda>'

To Reproduce

from darts.datasets import EnergyDataset
from darts.dataprocessing.transformers import Scaler, MissingValuesFiller



df = EnergyDataset().load().pd_dataframe()
df_day_avg = df.groupby(df.index.astype(str).str.split(" ").str[0]).mean().reset_index()
filler = MissingValuesFiller()
scaler = Scaler()
series = scaler.fit_transform(
    filler.transform(
        TimeSeries.from_dataframe(
            df_day_avg, "time", ["generation hydro run-of-river and poundage"]
        )
    )
).astype(np.float32)



train, val = series.split_after(pd.Timestamp("20170901"))


model = FFT(nr_freqs_to_keep=2, trend='linear', trend_poly_degree=2)
model.fit(train)


model.save('test.pht')

Expected behavior

The model should be saved properly as test.pht

System (please complete the following information):

• Python version: 3.7.13
• darts version 0.23.1

Additional context

In another code a co-worker got a similar error :
AttributeError: Can't pickle local object 'EvaluationLoop.advance.<locals>.batch_to_device'

It was trying model.save() method after fitting a TFTModel. I couldn't reproduce the error on my machine though ...

[madtoinou]

Hi,

After a quick investigation, it appears that "linear" is not a supported value for the trend argument of FTT.fit(). If you replace it by "poly" or "exp", the save() method works just fine. For your example more specifically, trend="poly" trend_poly_degree=1 should result in a linear trend function.

The error seems to be linked to the trend_function being lambda x : 0 when trend is anything but "poly" or "exp", not sure of how to solve it atm...