timeenc signals not passed to the model

[rosswolf]

Hey there, my apologies if I got this wrong. I've been stepping through the Units model, and it looks like you never actually use the time encoding features within the model training. Some of the params are passed in, but then they are left unused. Is that intentional, and/or did I miss somewhere they they are in fact being leveraged?

[gasvn]

Because some datasets are only used for training are not used during finetuning, so the prompt and task tokens for these datasets are not used. Also, one pretrained tower is removed after pretraining, so it's parameters are not used.

[rosswolf]

How can we fine tune with a dataset that incorporates time-enc signals then and have the model make use of it?

[gasvn]

Can you give more details about time-enc signals?

[rosswolf]

I know the time information (Month, day, year) for my timeseries data (encoded as timefeatures and returned from my Dataset in the 3rd & 4th parts of the tuple, aka seq_x_mark, seq_y_mark.) I would like the model to be able to learn trends about the data as it correlates to these signals. (Eg, temperature sensor data, with month/day/year, the temperature in winter is colder and the model can learn that if my input data has dates associated with it) and so I was expecting that by having my dataset return the timefeature information it would be taken into account