Default Initialization of RNNs

Hi there,

I am trying to find out what is the default intialization for all RNN parameters in DyNet. I have a model in PyTorch that is consistently underperforming its counterpart in DyNet and I want to find out why. I couldn't find any specific information neither by looking at the docs nor the C++ code (which doesn't imply that isn't there, of course). I'd appreciate any hints!

Review the DyNet documentation and C++ code to locate how default initialization is defined for RNN parameters. Compare the relevant behavior with the PyTorch model described in the issue. Done would be a clear, documented account of the defaults, if the project decides to address the question.

Written by the indexing model from the issue text.