sponsored by Deep Learning College @Modulabs, Seoul, South Korea contributor: Jinwoo Oh, Wonyoung Seo, Sumin Han
- Music Transformer: Generating Music with Long-Term Structure
- 2019 ICLR, Cheng-Zhi Anna Huang et al., Google Brain
- paper link
- paper review
- Re-producer : Yang-Kichang
- Customizer: Jinwoo Oh
To extend Chopin’s unfinished piano sonata, Canon in F minor, B. 129a, we trained the model with Chopin's other works in MAESTRO dataset.
-
Train with only Decoder wise ( only self-attention AR. )
$ python train.py --pickle_dir= './chopin_preroc/' --save_path= './result/'
-
learning rate : 0.0001
-
head size : 4
-
number of layers : 6
-
seqence length : 2048
-
embedding dim : 256 (dh = 256 / 4 = 64)
-
batch size : 1
Because Our memory didn't statisfy the requirement of model (Batch_size = 2), we define it 1.