passalis / dain Goto Github PK

In line 39 of train_utils.py

It should be:

loss = (train_loss / counter ).cpu().data.numpy()

otherwise train_loss is not used and loss seems to be a pretty small number.

Can you please double check?

where i can found the run_exp.py? thanks

Sorry but i am a little confused...
Why is the input shape being (n_samples, dim, n_feature_vectors),
instead of (n_samples, n_feature_vectors, dim)?
Where n_feature_vectors stands for the length of time window,
dim stands for feature count,
Am i wrong?

I try to use adaptive input normalization for time series data, and my window size is 20/100/200, but the result is not good.
While I use Z-Score got a nice result. Can you give me some suggestions? Is daptive input normalization suitable for special scenarios？
thanks

Hi, thanks for the repository and very interested in using dain to normalize my data. I am hoping there is a way to just normalize the data from raw -> normalize using dain for both training and production data (using same model)? Any help to do that would be very much appreciated. I am trying to follow the code but not sure even where to begin. TIA

One thing that confused me is we are already normalizing the input inside dain, but I am not sure if I should normalize the whole train data with something like z-score before dain. To me it feels like both should be ok, but am I missing something or is there a preferable way? Thanks.

I am reading the paper. I got little confused about the input. In equation (2) in paper, are we getting average over window size of a given ith time series data or whole length of the time series i? I suppose it should be over window size but could not be sure.

Thank you!