Suppose I have an RNN, and I would like to do truncated backpropagation through time. Typically you can do truncated backprop at every K interval, but instead I want to backprop K steps at every time step, and conceptually I would like to dynamically insert a stop_gradient K steps before. Does Tensorflow 2 support this kind of operation? Thanks!
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…