[ENH] Add Tide to test framework of ptf-v2

[phoeenniixx]

Adds Tide model from dsipts to ptf-v2

Created a new folder tide_dsipts in tide that contains all the necessary parts for the tide

[phoeenniixx]

@agobbifbk, @fkiraly, I think this PR is ready for review ( I tried to keep the code almost same as tide in dsipts)

[phoeenniixx]

(still need to add docstrings, the current docstrings are also copied from dsipts, I will add them today as well)

[fkiraly]

I think this PR is ready for review

then please click the button "ready for review" (below), I almost did not see this

[agobbifbk]

pytorch-forecasting/pytorch_forecasting/models/tide/tide_dsipts/_tide_v2.py

Line 165 in 89939e8

if "x_num_past" not in batch:

Here I have some doubts: are you sure all the keys are in the batch dict: encoder_cont,decoder_cont,encoder_cat,decoder_cat? I see in the D2 that you put some defaults if this information is missing (see

pytorch-forecasting/pytorch_forecasting/data/data_module.py

Line 491 in 89939e8

decoder_cont = (

)

This is has not been tested, for example here (

pytorch-forecasting/pytorch_forecasting/layers/_dsipts/_sub_nn.py

Line 58 in 89939e8

cat_n_embd = self.get_cat_n_embd(cat_vars)

) I don't know what could be the behavior of this operation in case of a 0-shaped tensor. Are your tests finishing without any error?

[phoeenniixx]

Are your tests finishing without any error?

Yes,
From what I can understand, embedding_cat_variables is to get embeddings for categorical variables, but as we discussed, we are for now, not passing any categoricals, so I think this layer is not used, that's why there is no issue?

The data i am using, doesnot pass any categoricals:

pytorch-forecasting/pytorch_forecasting/tests/_data_scenarios.py

Line 189 in d931c3b

categorical_features = []

[agobbifbk]

Ok, I see. Even though there is the key in the dict there is no category to loop on. It should work :-)

[phoeenniixx]

exactly :)

[fkiraly]

why are we doing this? This seems like a significant change to the internal contract of TimeSeriesDataSet

[phoeenniixx]

We discussed it some months back, that there was no target_past in EncoderDecoderDataModule and some models do require it. TSLibDataModule already implements it as history_target (see here). Tide also needs it, and we were already calculating target_past for the target_scale, so I just renamed the variable and added it to the return

(if you see line 509, we are still returning target_scale as well)

[fkiraly]

• see above, it seems like there is a significant change to EncoderDecoderTimeSeriesDataModule in terms of the dict return. The target_scale variable might not be assigned, and its source is being changed. Can you please explain?
• Can you add the new layers in folders sorted by type of layer rather than by name of dependency (dsipts)?