musicxlstm - An xLSTM Model

Trained with Helibrunna by Dr. Tristan Behrens.

Configuration

training:
  model_name: musicxlstm
  batch_size: 256
  lr: 0.001
  lr_warmup_steps: 2000
  lr_decay_until_steps: 20000
  lr_decay_factor: 0.001
  weight_decay: 0.1
  amp_precision: bfloat16
  weight_precision: float32
  enable_mixed_precision: true
  num_epochs: 20
  output_dir: output/musicxlstm
  save_every_step: 100
  log_every_step: 10
  wandb_project: musicxlstm
  tensorboard_dir: runs
model:
  num_blocks: 2
  embedding_dim: 64
  mlstm_block:
    mlstm:
      num_heads: 1
  slstm_block:
    slstm:
      num_heads: 1
  slstm_at:
  - 1
  context_length: 256
  vocab_size: 119
dataset:
  hugging_face_id: TristanBehrens/js-fakes-4bars
tokenizer:
  type: whitespace