So I have trained an LSTM model which takes time-series (C-MAPSS dataset) data that have sliding window formulation (somewhat similar to the picture below)
model input : (batch_size x sliding_window x feature_size)
model output: (batch_size x sliding_window_outputs)
, which is good for training the model. Helps in the reduction of MSE loss and r2 loss seems alright. Now I have run into another complication, i.e., because of such dataset/dataloaders formulation, the predictions are also going to be sliding window technique (because test dataset also has to be in the format since model accepts such dimensionality) and cannot be really used on test dataset to check the model performance in the end.
Can anybody suggest how to obtain predictons just for the time-series data and not according to the sliding window approach? Basically, how to formulate the test-data loaders or any post testing modification so that I can get predictions serially and not sliding window fashion?
Thank you
