Model structure/topology


Hi, I am using an NMT architecture with 4 encoder(GNMT) layers and 4 decoder layers architecture. But in the training options, I saw that there is option of convolutional layers (with default size =2). I wanted to ask, does the gnmt encoder option overrides the convolutional layers? Or there are 4 LSTM layers on top of convolutional layers?

Sorry for troubling, I am not familiar with Lua so its difficult for me to investigate this.

(Guillaume Klein) #2


Some options are model dependent. In this case, convolutional layers are only used when -encoder_type is set to cnn.


Ok, many thanks guillaumekin!!