Hyperparameter tunning

What are the best practices of performing hyper-parameters of a LSTM, like trying out different hidden_sizes to see how they affect the performance of the model for a certain task?