Recently in the text generated by the paper, the puzzling over how the new design,
At noon today suddenly thought of as the text generated basic seq2seq can accomplish the reconstruction of the text, so don't design the new network in the process of reconstruction, just want to achieve the loss and to oneself, is automatically adjusted as a result, they want
That if you can, so what is the point of design of the model structure,
I think it is not ok, design is various structure can better guide the direction of the network to find faster convergence, consult, please contact