Does TFT model use the ground truth of target during testing?

[zzhuqshun]

In this neural network structure, I want the model to do the train and validation without using the historical target values and to make the prediction directly through the covariates, so I set these to avoid model using historical target information:

• time_varying_unknown_reals=[]
• time_varying_known_reals=[cov1, cov2, cov3]

In the test, the model requires to give the ground truth of the target value, not nan or fixed value, which means the model still refers to the ground truth of the target value to make the prediction. But in the real case, it should be predict the target value by covariates, and it is not possible to know the ground truth of the target in advance.

My questions would be:

• Does the model use the ground truth of the target during training and testing?
• Should I redefine the training process? Use scheduled sampling to control the amount of ground truth input to the model during each training session, to keep the training process and the testing process in the exact same situation.
• In short: how should I make TFT do the autoregressiv training with predicted values?