# 1230 meeting note
###### tags: `meeting note`
- What did I do in this week
- I trained the Bidirectional LSTM
- sounds good on irealpro dataset
- after transfering to dense_jazz dataset, sounds good, too
- the output of inferencing on unseen data is also good
- however, there are some interesting thing
- there are some unnatural short notes jumping up and down
- there are many long notes, because we cannot make pitch repetition in output
- I modified some code:
- using tf.data.dataset API
- I rewrote the cost function to fit the masking
- Tasks
- try different hyperparameter on BD-LSTM
- piece length
- width and depth of NN
- unit number of LSTM cell
- change to GRU cell
- some direction of better input/output format
- add drums informations into input
- add velocity as output
- add onset/offset timing correction into output
- indicate whether the note is filled in the beat, or start slightly late, or stop early
- use regression to predict
- ideally, it can solve swing problem and repetition problem at the same time
- try new concept of model
- one note per output, which is more suitable for encoder-decoder architecture in my opinion
- train encoder-decoder rnn with attention
- I think it’s a great idea to treat it as a kind of translation problem
- perform experiments to prove the previous improvements are not meaningless
- train same epochs on LSTM, BD-LSTM and encoder-decoder and compare
- use the same model, compare the result of:
- training on biggest dataset directly
- training on small dataset first, and transfer to larger dataset
- How we do our presentation?