Study of Dependency on number of LSTM units for Character based Text Generation models

Long Short Term Memory cells (LSTMs) are used to make character-based generation models. Single dimensional Convolutional LSTM networks are also used in sequential data processing. There is a significant effect of number of LSTM cells on the quality of the generation as the amount of overfitting and underfitting depends upon the number of units. We study the effect of the number of LSTM units and we find that increased number of LSTM cells initially improves the prediction. However increasing the number of LSTM cells above a certain number causes poorer results. The vocabulary of the corpus produces better results if the size is reduced.