[DOC] Ensure clear documentation of base API - model.predict, decoder, test_dataset in TFT

[ms042087]

I have trouble in understanding in how to predict the future value with some known future feature.

For example in the predict on new data section in the official tutorial: https://pytorch-forecasting.readthedocs.io/en/latest/tutorials/stallion.html#Predict-on-new-data.

1. From my understanding, encoder should have the length equal max_encoder_length. Decoder, which represents the future values, should have the length equal max_prediction_length.
2. In the decoder, I should only fill up the known future values. For the unknown future values, should I set it with 0 by pd.concat([encoder, decoder]).fillna(0.0) or NaN? I have tried to fill in NaN but the model.predict function does not accept NaN values. If I fill the unknown value of 0, will it affect the prediction or they will be ignored during inference?
3. If one of the target_col is in the time_varying_unknown_reals in the model training (i.e. for learning how the past values affect the future values), what should I set in the decoder? If I set them to 0, will the model predict the target_col value correctly?
4. If I understand correctly from [1], I should create a testing data set from from_dataset method first with predict = True. It will makes the last items in pd.concat([encoder, decoder]) with max_prediction_length, i.e., the decoder to be the predicted target.

Thanks a lot for your help in advance.

[1] https://pytorch-forecasting.readthedocs.io/en/stable/api/pytorch_forecasting.data.timeseries.TimeSeriesDataSet.html#pytorch_forecasting.data.timeseries.TimeSeriesDataSet.from_dataset

[moogoofoo]

I agree that the documentation is not at all clear that the target values are passed to the encoder. This should be made abundantly clear. This may be related to issue #1585

[fkiraly]

I see - I think we should turn this into a documentation issue first, to document the status quo in:

• docstrings of methods
• a complete usage tutorial

Once that's clear, based on that discussions on potential rearchitecture or interface changes can be had, if needed.