My current LSTM has a many-to-one structure (please see the pic below). On the top of LSTM layer, I added one dropout layer and one linear layer to get the final output, so in PyTorch it looks like
self.dropout = nn.Dropout(0.2)
self.ln = nn.Linear(hidden_size, 1)
h, c = self.lstm(x)
last_h = self.dropout(h[:, -1, :])
out = self.ln(last_h)
Now, I want to modify my LSTM to simulate many-to-many; let’s say many-to-10.
Q1. Is it correct if I add one dropout and one dense layer for the 10-cell outputs? Something like
self.dropout = nn.Dropout(0.2)
self.ln = nn.Linear(hidden_sizex10, 10)
h, c = self.lstm(x)
last_h = self.dropout(h[:, -10, :])
out = self.ln(last_h)
Q2. Or, should I add one dropout and dense layer for each cell output (i.e., 10 dropout and 10 dense layers)? In this case, how I can implement them in the code?