mugen/config/solver/magnet/audio_magnet_16khz.yaml

# @package __global__

# This is the training loop solver
# for the base audio-MAGNeT model (text-to-sound)
# on monophonic audio sampled at 16 kHz
# using a similar EnCodec+LM setup to MAGNeT
defaults:
  - audiogen/default
  - /model: lm/audiogen_lm
  - override /dset: audio/default
  - _self_

lm_model: transformer_lm_magnet
solver: audio_magnet

autocast: true
autocast_dtype: float16

# EnCodec large trained on mono-channel music audio sampled at 16khz
# with a total stride of 320 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: //reference/bd44a852/checkpoint.th

channels: 1
sample_rate: 16000

deadlock:
  use: true  # deadlock detection

dataset:
  batch_size: 128  # matching AudioGen paper setup (256 * mix_p=0.5 = 128)
  num_workers: 10
  segment_duration: 10
  min_segment_ratio: 1.0
  sample_on_weight: false  # Uniform sampling all the way
  sample_on_duration: false  # Uniform sampling all the way
  external_metadata_source: null
  # sample mixing augmentation at train time
  train:
    batch_size: 256  # matching AudioGen paper setup
    aug_p: 0.5  # perform audio mixing 50% of the time
    mix_p: 0.5  # proportion of batch items mixed together
                # important: note that this will reduce the
                # actual batch size used at train time
                # which will be equal to mix_p * batch_size
    mix_snr_low: -5
    mix_snr_high: 5
    mix_min_overlap: 0.5

optim:
  epochs: 100
  optimizer: adamw
  lr: 5e-4
  ema:
    use: true
    updates: 10
    device: cuda

logging:
  log_tensorboard: true

schedule:
  lr_scheduler: inverse_sqrt
  inverse_sqrt:
    warmup: 3000
    warmup_init_lr: 0.0

codebooks_pattern:
  modeling: parallel
  parallel:
    empty_initial: -1
  
transformer_lm:
  card: 2048
  causal: false
  subcodes_context: 5
  compression_model_framerate: 50  # NOTE: Must match the actual frame rate of the used compression model 
  segment_duration: 0
  span_len: -1

masking:
  span_len: 3

generate:
  lm: 
    max_prompt_len: null
    max_gen_len: null
    remove_prompts: false
    use_sampling: true
    temp: 3.5
    top_k: 0
    top_p: 0.8
    max_cfg_coef: 20.0
    min_cfg_coef: 1.0
    decoding_steps: [20, 10, 10, 10]
    anneal_temp: true
    span_scoring: 'max'
    span_arrangement: 'nonoverlap'
    prompted_samples: false
    samples:
      prompted: false
      unprompted: true