train_duration_dit_s.yaml

model:
  name: DiT-S
  input_size: 512
  embedding_vocab_size: 1024
  learn_sigma: true
optimization:
  constant_memory: false
  epochs: 1400
  global_batch_size: 512
  initial_input_size: 32
  learning_rate: 1.0e-4
  min_lr: 1.0e-5
  warmup_iters: 10000
  lr_decay_iters: 100000
  decay_lr: true
  weight_decay: 0.0
  max_grad_norm: 20.0
  betas:
    beta1: 0.9
    beta2: 0.999
  loss:
    num_timesteps: 1000
data:
  data_path: duration.npy
  data_dim: 1
  data_std: 0.067776896
  data_mean: 0.08663661
  normalize: true
training:
  enable_compile: true
  use_bfloat16: true
  use_block_mask: false
  seed: 42
  ckpt_every: 10_000
  log_every: 100
  results_dir: results/duration
  resume_from_ckpt: null
  wandb:
    enable: true
    project: diffusion-speech