Take for example, I am building an LSTM RNN which takes in 5 features that correlate to the number of sales of a product for that day. The LSTM takes in sequences of 10 days of these 5 features (it is unrolled to 10 time steps, each receiving 5 features as input), and then attempts to predict the number of sales for the 11th day.
As the network is unrolled into 10 time steps, it then outputs a figure at each. Should the figure output at time step i be considered to be its prediction for the number of sales at time step i + 1, or should the first 9 outputs be ignored, as all that is cared about is the last prediction - the prediction of sales on the 11th day (the output at time step 10).
Leading on from this, should my labels be sequences of 10 figures corresponding to the number of sales on day 2, 3, ..., 11, or just a single figure for day 11? In other words, when calculating the loss, and hence affecting backpropagation, should the error between the output at each of the 10 time steps and the corresponding label for that day be considered, or just the difference between the last output and the corresponding single label?
The same goes for the accuracy calculation - should only the very last day be considered, or all 10?
