I have implemented an LSTM model that have 2 LSTM layers, a dropout layer and a dense layer for predictions. I trained my LSTM model on 1000 XML files. Each file has 4 main markups with very simple fields in between the markups. My training data was acquired using a list of sequences. The sequence length is 3 and the step window is 1.
As model parameters, I have set :
learning rate: 0.001
batch size: 65
Number of iterations: 20
So for predictions, I give my model 3 words which are 3 XML markups chosen randomly and the model should generate the 10 next markups. What I don't understand is why my model predicts accurate results when the input seed is chosen randomly while when I give it a constant input seed, it does not predict accurately.