Slicing predictions for using `plot_prediction()` method with Temporal Fusion Transformer

[nicoliKim]

Hello,

First of all thanks a lot for putting together this nice libraries.

I have a question regarding on how to properly call the plot_prediction() method.
Assume I have a TemporalFusionTransformer model which I refer below as model.

To make predictions I do the following:

preds = model.predict(
            test_dataloader,
            mode='raw', 
            return_x=True,
            return_y=True, 
            mode_kwargs={"n_samples": n_samples}, 
            trainer_kwargs=dict(accelerator="mps")
            )

Now using these predictions, I want to plot the results.
Note that if I do

type(self.predictions.output.prediction)

this is a list of size 3where each element of the list is a different variable I am forecasting. If I analyze the shape of the first element of the list I get

self.predictions.output.prediction[0].shape

I get

torch.Size([1, 60, 3])

where the first dimension is the batch, second dimension is the prediction length and last dimension is the quantiles I obtain from the QuantileLoss loss function.

What I am struggling with is the following. I'd like to call the plot_prediction() method so that only one element of the list self.predictions.output.prediction is being used.

I hope the way I phrased the question makes sense. Please bear with me as I am new to the library and its workflow. If you have suggestions on how to do things differently that would be very much appreciated.

[phoeenniixx]

Hello @nicoliKim! I don't think right now there is any way to plot only one element of the raw predictions using plot_prediction(), you'll have to do it maunally.

But We are working on the rework of this package for the v2.0 release and we would keep this thing in mind for the new version!
You can also look at the beta version here

And if you have any feedback or suggestions for v2.0 please comment on #1736 :)

FYI @PranavBhatP, @fkiraly, @agobbifbk

[nicoliKim]

Thanks a lot @phoeenniixx for letting me know! Would be cool to see this being implemented.
As an alternative, let's assume I have 3 variables to predict: do you think that training 3 independent models (for each of these variables) is an option for then using the plot_predict() method at the current stage?

Do you expect that training one model to forecast three variables (i.e., setting target=['Var1', 'Var2', 'Var3']) compared to having three models to forecast a single variable may lead to significant differences in performance?

[fkiraly]

Do you expect that training one model to forecast three variables (i.e., setting target=['Var1', 'Var2', 'Var3']) compared to having three models to forecast a single variable may lead to significant differences in performance?

In general, it may perform differently if you do this. Probably worse since you are missing out on associations between the variables? Hard to say without trying though.