CUDNN RNN implementation

Accelerated Computing GPU-Accelerated Libraries
1

The cudnn RNN implementation takes a dropout descriptor but has either zero documentation on exactly how this descriptor is used.

I am guessing the description is here:

But I am not certain. How is dropout used in the cudnn RNN implementation?

2

I guess further along those lines it would be ideal of the dropout supported multiplicative noise instead of bernoulli dropout as there are some cases where it appears that multiplicative noise centered around [1,user-chosen-std-dev] has some nicer effects on generalization.