The architecture used in model training.

[ellurunaresh]

Hi all,
I have trained g2p model for my en-US dataset. I was trying to understand the internal details of model training. I want somebody to clarify the below questions?

1. Are we using dense layers with RELU as the activation function in model training?. If yes, I have observed in the graph having "conv1" unit in the feed forward network.
   "encoder/layer_0/ffn/conv1/kernel/Initializer/random_uniform". What is "conv1" unit here?
2. Can we use RNN with GRU cell instead of FFN. How much improvement can we expect?

Please reply to the above questions.