invalid frequency error while using DeepAREstimator on data created using PandasDataset

[mohammednaseefop]

I'm using PandasDataset to create the model ready data as follows

from gluonts.dataset.pandas import PandasDataset
ds = PandasDataset.from_long_dataframe(df, target="Expense", item_id="Property ID")

ds.freq returns 'AS-JAN' as my frequency.

later when I try the following code I'm getting an error

from gluonts.mx import DeepAREstimator, Trainer

estimator = DeepAREstimator(
    freq=ds.freq, prediction_length=2, trainer=Trainer(epochs=1)
)
predictor = estimator.train(ds)
predictions = predictor.predict(ds)

---------------------------------------------------------------------------
Exception                                 Traceback (most recent call last)
[<ipython-input-128-26b28388962d>](https://localhost:8080/#) in <module>
      1 from gluonts.mx import DeepAREstimator, Trainer
      2 
----> 3 estimator = DeepAREstimator(
      4     freq=ds.freq, prediction_length=2, trainer=Trainer(epochs=1)
      5 )

2 frames
[/usr/local/lib/python3.8/dist-packages/gluonts/time_feature/lag.py](https://localhost:8080/#) in get_lags_for_frequency(freq_str, lag_ub, num_lags)
    141         )
    142     else:
--> 143         raise Exception("invalid frequency")
    144 
    145     # flatten lags list and filter

Exception: invalid frequency

Why am I getting the error? I am guessing not all the frequency strings are supported by functions in GluonTS.
How do I fix this issue?

[jaheba]

Hey, what is ds.freq?

[mohammednaseefop]

I'm following the gluonts document for pandas.Data Frame based dataset with my dataset.
I took the code from there.

ds = PandasDataset.from_long_dataframe(df, target="Expense", item_id="Property ID")
ds.freq

ds.freq returns the frequency of the timeseries. In the gluonts document the frequency of the data is hourly so it returns "H".

[jaheba]

Sorry, maybe I misread your question initially, but you write:

ds.freq returns 'AS-JAN' as my frequency.

Now you claim it is H.

---

To me it looks like we don't normalise the frequency. AS-JAN should just become A, which get_lags_for_frequency relies on. As a quick fix, you should be able to just pass freq="A" to DeepAR.

[mohammednaseefop]

Thanks!
Just to clarify . My custom data's frequency is 'AS-JAN'. In the gluonts document they are using a data at hourly frequency.

[jaheba]

Got it! Let me get the issue fixed in #2645 :)