Music Modeling on Piano-e-Competition (val)
1.835Val NLLTF with relative global attention
Evaluation results
| Method | Links | |
|---|---|---|
| TF with relative global attentionArchitecture=Music Transformer (Relative Global), Number of layers (L)=6, Filter size (fs)=2048, Maximum relative distance (r)=1024, Sequence length=20482018.09 | 1.835 | |
| TF with relative local attentionArchitecture=Music Transformer (Relative Local), Number of layers (L)=6, Filter size (fs)=1024, Maximum relative distance (r)=2048, Block size (bs)=512, Sequence length=20482018.09 | 1.84 | |
| Transformer (TF) baselineArchitecture=Transformer, Number of layers (L)=6, Hidden size (hs)=256, Attention size (att)=512, Filter size (fs)=2048, Maximum relative distance (r)=1024, Number of heads (h)=8, Sequence length=20482018.09 | 1.861 | |
| TF with local attentionArchitecture=Transformer, Number of layers (L)=8, Filter size (fs)=1024, Block size (bs)=512, Sequence length=20482018.09 | 1.863 | |
| LSTM with attentionArchitecture=LSTM, Number of layers (L)=3, Hidden size (hs)=1024, Attention size (att)=1024, Sequence length=20482018.09 | 1.959 | |
| PERFORMANCE RNN (LSTM)Architecture=LSTM, Number of layers (L)=3, Hidden size (hs)=1024, Sequence length=20482018.09 | 1.969 |