Specifications for how RNN needs input
nielsrolf opened this issue · comments
- Will one RNN be trained for all stores, or will they be store specific?
- How should the data be split into train/val/test? Split the stores, or give only the first part of each time series as train data?
- How should a train batch look like? Should it be a single time series for one store? If yes, the test and validation data can't be passed at once and the RNN's won't be compatible with our current
AbstractForecaster
. We could define a different interface for them (and maybe mergeAbstractForecaster
withFeedForward
; the latter adds tensorflow placeholder and session logic mainly)
I think training data should include all stores or it will reduce the size of dataset
- yes one network on all the stores, although it would make sense to have an identifier of stores in the feature space.
- the training data should be totally independant of the test data meaning, let's assume we have data from date 1.1.17 to 1.5.17 the network should be trained on for example 2 months of data then tested on the other 3. In other words, split the data per date train the model on older dates then predict the newer.
- the train batch could be for the data for each day/month.