I also referred the blog you mentioned time_sequence_prediction on pytorch forums. I am using a similar approach for one my problem where there are no features but data itself. I have implemented it here, but I am not sure what I have implemented is correct. My questions are below:

1. from the code self.lstm1 = nn.LSTMCell(1, 51) , though the sine waves data i.e. input_t is 997*1 and as the nn.LSTMCell provides information about first argument as input_size, why it shouldn't be (997, 51)
2. When I trained my model I got the MSELoss around 139 at 15th epoch, why this is so high? though I am yet to plot the predicted graph.
3. Whats the sequence length in sine-wave prediction example?