New question regarding the multistep prediction strategy

[rodgdutra]

I was reading and debugging the multi-step implementation to understand it better. I've come across an interesting thing, seams like the features and labels in the training and evaluation are the same. This behavior is correct ? I thought that in a multi step prediction problem the input features is delayed in relation to the wanted labels, this way we have a window of past behavior of the data and we are aiming to predict the future behavior of the data.

[rodgdutra]

One can observe this in the lines:

transformer-time-series-prediction/transformer-multistep.py

Lines 91 to 92 in 570d39b

	train_seq = np.append(input_data[i:i+tw][:-output_window] , output_window * [0])
	train_label = input_data[i:i+tw]

[oliverguhr]

In this case, it is correct. Depending on the training mode, the loss is calculated over the output (prediction window) or the whole sequence:

transformer-time-series-prediction/transformer-multistep.py

Lines 177 to 180 in 570d39b

	if calculate_loss_over_all_values:
	total_loss += criterion(output, target).item()
	else:
	total_loss += criterion(output[-output_window:], target[-output_window:]).item()

You can use this line to change the behaviour:

transformer-time-series-prediction/transformer-multistep.py

Line 14 in 570d39b

calculate_loss_over_all_values = False

Depending on the task, it is sometimes beneficial to train the model, to also predict the correct value for the input features. You can also read a bit about this flag in this issue: #6

Feel free to reopen this issue if you have any questions.