|
|
--- |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
Selection of trained models for the RhythGen project: https://github.com/efraimdahl/RhythGen |
|
|
|
|
|
# Models |
|
|
• **LAS:** In-attention conditioning with syncopation labels on the Lieder dataset. |
|
|
|
|
|
• **LMR2:** Attention modulation with spectral weight profiles profiles and a learnable scale (initialized at 10) on the Lieder dataset. |
|
|
|
|
|
• **LB:** Baseline NotaGen (small) model without any conditioning, finetuned on the Lieder dataset. |
|
|
|
|
|
• **HAS2:** In-attention conditioning with syncopation labels on the RAG-RH dataset. |
|
|
|
|
|
• **RAS2:** In-attention conditioning with syncopation scores and voice masking on the RAG- |
|
|
collection. |
|
|
|
|
|
• **RB:** Baseline NotaGen model without any conditioning, finetuned on the RAG-collection. |
|
|
|
|
|
|
|
|
# Base Model |
|
|
All models are fintuned from [NotaGen small](https://huggingface.co/ElectricAlexis/NotaGen/blob/main/weights_notagen_pretrain_p_size_16_p_length_2048_p_layers_12_c_layers_3_h_size_768_lr_0.0002_batch_8.pth). |
|
|
|
|
|
# Datasets |
|
|
We fine-tuned NotaGen-small on a corpus of 1000-1500 pieces from either the [Lieder Dataset](https://github.com/OpenScore/Lieder) which is public or the [RAG Collection](https://dspace.library.uu.nl/bitstream/handle/1874/354841/OdekerkenVolkKoops2017.pdf?sequence=1) which is available on request. |
