Parameter tuning

[junwei-pan]

Which parameter set do you use for those results on the souldcloud? I use the following parameters and the result music is not so good, show I increase the number of epochs?

maxlen = 128 # Max input length
num_hidden_units = 10 # Number of hidden units of LSTM
step = 8
num_char_pred = 17 * 30 # Number of predicted char
num_layers = 2
num_epoch = 30
batch_size = 128
diversity = 0.9

Here is the code, thanks a lot.

[keunwoochoi]

My code has every parameter as I used. I think you should vary diversity like 0.8, 0.9, 1.0, 1.1, 1.2 and increase num_hidden_units. Given the vocabulary size, 10 seems quite too small.

[junwei-pan]

Actually I use 128 for the num_hidden_units. How many epochs do you think I should use? Is 30 too small?

[keunwoochoi]

30 seems fine. On the soundcloud, LSTMetallica Drums - 10 - 30, 0.50 means 30 epochs and 0.50 diversity. But it would be worth running more than 30.

[junwei-pan]

OK, I see. I will try more diversities as you suggested.