# 1209 meeting note
###### tags: `meeting note`

- Jay’s part:
- input2midi, midi2output are basically completed, but there exist some problem:
- no velocity information
- no swing
- no position for ‘rest’, so i use the position originally D5 (the highest note)
- no onset (where a note begin) information, so repetition of the same pitch will be considered as a single note
- I found another data
- [JAAH Dataset](https://github.com/MTG/JAAH)
- Jazz Audio-Aligned Harmony Dataset
- contains audio, and annotations(chord, beat, structure...)
- 113 tracks
- I think we can use [BassUNet](https://github.com/jakobabesser/bassunet) to generate bass part, then use the transformed annotation as input, and the generated bass part as ground truth
- may be we can use [jazzChordTranscription](https://github.com/geduran/jazzChordTranscription) to generate more data in the future
- I think the current model could be improved by:
- Enlarging the piece_length
- "piece_length=16" covers only one bar, which is kinda small
- Using all data training
- you can unzip the zip files to train model with all data
- make the data loading part able to obtain data from different directory
- you can input a whole piece to generate a complete bass midi by:
- ```python
import midi_np_translation.output2midi as output2midi

piece_data = slice_per_step(np.load("test_input/AfroBlues.MID.npy"))
output2midi.output_to_midi(model.predict(piece_data).reshape(-1, 52), "input_midi/jazz_standards/AfroBlues.MID")

      - Using the argument "stateful" of LSTM to preserve whole piece information?
      - I have tried using all data to train the model, and find that almost all ouput are "rest"`
          - https://drive.google.com/file/d/1ev0t-GjZi_uYXa5OPCwyGEWJRbQ0sY3c/view?usp=sharing
- tasks in the next iteration
  - process new data (JAAH)
  - design new input/output format
  - improve the current model