Ideal structure of RNN

[Shafaq19]

The ideal structure is as follows:

Embedding layer (nn.Embedding) before the LSTM or GRU layer.
The fully-connected layer comes at the end to get our desired number of outputs.
Extra marks for not using a dropout after LSTM and before FC layer, as the drop out is already incorporated in the LSTMs, A lot of students will add it and then end up finding convergence difficult
You can try to add more than one fc:

init

self.fcc=nn.Linear(self.hidden_dim, self.hidden_dim)
self.fcc2=nn.Linear(self.hidden_dim,self.output_size)
....

forward

output,hidden=self.lstm(embedded,hidden)
lstm_output = output.contiguous().view(-1, self.hidden_dim)
output= self.fcc(output)
output=self.dropout(output)
output=self.fcc2(output)