You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi,
My problem: I want to determine a value for time t based on variables at time t and their associated time series, i.e. values at time t-1, t-2, and so on.
So for example: I have a dataset with 4 columns which store values for certain sensors (e.g. A,B,C,D). Additionally I have a datetime column. I want that my model predicts the value for sensor A at time t by only looking at Sensor B, C, D time series, so time t, t-1, t-2, t-3 ... . In general when I test my Model I have an input of size [sequence length, variables (without target)] and my output is my target variable, so one value.
I want to implement the whole thing with a transformer. However, when configuring your framework, I specify my Target column and in the relevant columns I also specify my Target column again. If I understand it correctly, the model learns additionally about the history of the target column. However, I do not want this in my example. ( For training I have the information of the target variable of course, but I don't want this information to influence the weights of the model, but only based on the other relevant columns).
Is there a solution to adapt the configuration for a SimpleTransformer, Informer or Crossformer to my mentioned problem?
The text was updated successfully, but these errors were encountered:
jannik-scheider
changed the title
Transformer for time series
Question: Transformer for time series
Oct 10, 2023
Hi,
My problem: I want to determine a value for time t based on variables at time t and their associated time series, i.e. values at time t-1, t-2, and so on.
So for example: I have a dataset with 4 columns which store values for certain sensors (e.g. A,B,C,D). Additionally I have a datetime column. I want that my model predicts the value for sensor A at time t by only looking at Sensor B, C, D time series, so time t, t-1, t-2, t-3 ... . In general when I test my Model I have an input of size [sequence length, variables (without target)] and my output is my target variable, so one value.
I want to implement the whole thing with a transformer. However, when configuring your framework, I specify my Target column and in the relevant columns I also specify my Target column again. If I understand it correctly, the model learns additionally about the history of the target column. However, I do not want this in my example. ( For training I have the information of the target variable of course, but I don't want this information to influence the weights of the model, but only based on the other relevant columns).
Is there a solution to adapt the configuration for a SimpleTransformer, Informer or Crossformer to my mentioned problem?
The text was updated successfully, but these errors were encountered: