Problem in reproducing results with `--features MS` about autoformer HOT 3 CLOSED

Hi, thanks for your interest!
(1) Our paper does not provide the benchmark for "Multivariate to Univariate". The paper only shows the "Multivariate to Multivariate" and "Univariate to Univariate". So, "reproduce" is suitable in this situation.

(2) The code about MS can be found in './exp/exp_main.py' line76, 146, 224. The operation for picking the last dimension variate is done outside the model. Thus, it is normal that the model output is [bs x ts x 7].

(3) If you want to save the extra parameter in the last projector. You can ref to the following steps:

• change the --c_out to 1
• add trend_init=trend_init[:,:,-1:] to line 83 of './models/Autoformer.py'

Hope this will be helpful to you~

from autoformer.

Thank you for your quick response!

With regards to the code mentioned in './exp/exp_main.py' - my issue was that you picked the last dimension only for the targets (batch_y) and not for the model's outputs, which means that the loss criteria is based on dimensions 1 and 7, respectively.

Additionally, I agree with your suggestion in './models/Autoformer.py' but one needs to keep in mind that the dimension picking on trend_init=trend_init[:,:,-1:] should be done while considering the addition of mean in line 84.

Thanks again,

from autoformer.

Cool! Thanks for your suggestion.
I have updated the code for MS in 7f9ce3c
Hope this is helpful to you.

from autoformer.