How to choose the activation function of the last layer in LSTM autoencoder?

[logfeng]

Thanks for your great article and code, I've learned a lot from that! Here I have a question.
Why we choose the activation function of the last layer as relu in LSTM autoencoder, as there is a part of data less than 0 after standardization, causing the reconstruction error not fall close to 0. Is it better to choose tanh instead?
Thanks a lot!