# @package __global__

# WARNING: This is a base configuration file shared across ALL solvers in AudioCraft
# Please don't update this file directly. Instead use distinct configuration files
# to override the below configuration.
solver: ???

fsdp:
  use: false  # should we use FSDP.
  param_dtype: float16  # equivalent to autocast_dtype for FSDP.
  reduce_dtype: float32  # gradient averaging dtype, float32 will give max stability.
  buffer_dtype: float32  # dtype used for buffers, we don't have much buffers, so let's leave it.
  sharding_strategy: shard_grad_op  # can be shard_grad_op or full_shard.
                                    # full_shard will use less memory but slower ??
  per_block: true  # If True, uses nested FSDP.

profiler:
  enabled: false

deadlock:
  use: false
  timeout: 600

dataset:
  batch_size: ???
  num_workers: 10
  segment_duration: null
  num_samples: null
  return_info: false
  shuffle: false
  sample_on_duration: true
  sample_on_weight: true
  min_segment_ratio: 0.5
  train:
    num_samples: null
    shuffle: true
    shuffle_seed: 0  # if you want to sample the data differently.
    permutation_on_files: false
  valid:
    num_samples: null
  evaluate:
    num_samples: null
  generate:
    num_samples: null
    return_info: true

checkpoint:
  save_last: true
  save_every: null
  keep_last: null
  keep_every_states: null

generate:
  every: null
  path: 'samples'
  audio:
    format: 'mp3'
    strategy: 'clip'
    sample_rate: null
  lm:
    use_sampling: false
    temp: 1.0
    top_k: 0
    top_p: 0.0
evaluate:
  every: null
  num_workers: 5
  truncate_audio: null
  fixed_generation_duration: null  # in secs
  metrics:
    base: true  # run default evaluation (e.g. like train/valid stage)

optim:
  epochs: ???
  updates_per_epoch: null
  lr: ???
  optimizer: ???
  adam:
    betas: [0.9, 0.999]
    weight_decay: 0.
  ema:
    use: false  # whether to use EMA or not
    updates: ${optim.updates_per_epoch}  # frequency of updates of the EMA
    device: cpu  # device for EMA, can be put on GPU if more frequent updates
    decay: 0.99  # EMA decay value, if null, no EMA is used

schedule:
  lr_scheduler: null
  step:
    step_size: null
    gamma: null
  exponential:
    lr_decay: null
  cosine:
    warmup: null
    lr_min_ratio: 0.0
    cycle_length: 1.0
  polynomial_decay:
    warmup: null
    zero_lr_warmup_steps: 0
    end_lr: 0.0
    power: 1
  inverse_sqrt:
    warmup: null
    warmup_init_lr: 0.0
  linear_warmup:
    warmup: null
    warmup_init_lr: 0.0