pedrolarben / timeseriesforecasting-deeplearning Goto Github PK

Dear author，after reading your paper "An Experimental Review on Deep Learning Architectures for Time Series Forecasting", I learned a lot as a beginner on dealing with time series problem with Python. Therefore, I try to download the datasets of trainNN5 presented in "train": https://www.dropbox.com/s/x29hkn72ja7e7v2/train.csv?dl=1.
But me myself feel coufused about the function of _moving_windows_preprocessing, especially :param train :param core. in preprocessing.py.

I use df to name the DataFrame of NN5 (no header) and print out the loop:
for i, ts in list(enumerate(df)):
print(i,ts)
but it shows 0 0 1 1. Sorry to bother you, I think I have a problem with the understanding of what ‘train’ should I use .
By the way, your paper was so impressive that give me insight on DeepLearning.

Dear author, I just finished reading "An Experimental Review on Deep Learning Architectures for Time Series Forecasting" and I want to reproduce some experiments in my own code with pytorch. However, I can't seem to get my metric "WAPE" to work properly, that is my "WAPE" was suspiciously low. From your code, I understand that you denormalize output from the model, as well as the labels back to raw data first. Then, you calculate "WAPE" for each outputs-and-targets vector pair (for y, o in zip(actual, predicted):) and store it in a list (res[]). Lastly, we average over the list for the final result, which is the WAPE per sample. Is this result the one reported in Table 11 of your Review? Thanks in advance.

Hello!

It seems that DATASETS = json.load("../data/datasets.json") and other json.load calls in generate_data.py do not work properly. This function should take not a filename but an object.

P.S. Thank you for a nice repository!