mugen/config/solver/jasco/drums.yaml

# @package __global__

# This is the training loop solver
# for the base MusicGen model (text-to-music)
# on monophonic audio sampled at 32 kHz
defaults:
  - musicgen/default
  - /model: lm/musicgen_lm
  - override /dset: audio/default
  - override /conditioner: drums2music
  - _self_

lm_model: flow_matching
solver: jasco

autocast: true
autocast_dtype: float16

# EnCodec large trained on mono-channel music audio sampled at 32khz
# with a total stride of 640 leading to 50 frames/s.
# rvq.n_q=4, rvq.bins=2048, no quantization dropout
# (transformer_lm card and n_q must be compatible)
compression_model_checkpoint: //pretrained/facebook/encodec_32khz
# precomputed mean std accross training data sub-partition
compression_model_latent_std: 4.0102
compression_model_latent_mean: -0.0074
compression_model_framerate: 50
compression_model_latent_dim: 128 

efficient_attention_backend: xformers # restricted attention implementation supports only xformers at the moment

channels: 1
sample_rate: 32000

deadlock:
  use: true  # deadlock detection

dataset:
  segment_duration: 10
  batch_size: 336
  sample_on_weight: false  # Uniform sampling all the way
  sample_on_duration: false  # Uniform sampling all the way
  compression_model_framerate: ${compression_model_framerate}

optim:
  epochs: 500
  optimizer: adamw
  lr: 1e-4
  ema:
    use: true
    updates: 10
    device: cuda

logging:
  log_tensorboard: true

schedule:
  lr_scheduler: cosine
  cosine:
    warmup: 4000
    lr_min_ratio: 0.0
    cycle_length: 1.0

transformer_lm:
  causal: false
  skip_connections: false
  flow_dim: ${compression_model_latent_dim}
  drums_dim: ${conditioners.self_wav.drum_latents.out_dim}

generate:
  lm: 
    max_prompt_len: null
    max_gen_len: null
    remove_prompts: false
    cfg_coef_all: 3.0
    cfg_coef_txt: 1.0
    prompted_samples: false
    samples:
      prompted: false
      unprompted: true

conditioners:
  self_wav:
    drum_latents:
      compression_model_latent_dim: ${compression_model_latent_dim}
      compression_model_framerate: ${compression_model_framerate}
      segment_duration: ${dataset.segment_duration}