I would like to implement this method in TensorFlow.jl
which should automatically adjust the size of the minibatch in SGD. I am curious about the method, since in security the signals are typically very weak and therefore I would expect that large batch sizes would be needed. Looking at the source code
it seems like they need to identify consumers of a tensor (line 79 of gradient_moment.py).
I wonder if it is possible to do so in?