Not following the original training target?

[PipaFlores]

In the original paper, the composite model has two aims. First, to reconstruct the input, which is correctly implemented here (i guess). Second, to predict the following timesteps. The training is using backprop in both objectives, defined by two different losses.