configs/training/diffusion_ffhq512.yaml

model:
  target: models.unet.UNetModel
  params:
    image_size: 512
    in_channels: 3
    model_channels: 32
    out_channels: 6
    attention_resolutions: [32, 16, 8]
    dropout: 0
    channel_mult: [1, 2, 4, 8, 8, 16, 16]
    num_res_blocks: [1, 2, 2, 2, 2, 3, 4]
    conv_resample: True
    dims: 2
    use_fp16: False
    num_head_channels: 64
    use_scale_shift_norm: True
    resblock_updown: False
    use_new_attention_order: False

diffusion:
  target: models.script_util.create_gaussian_diffusion
  params:
    steps: 1000
    learn_sigma: True
    sigma_small: False
    noise_schedule: linear
    use_kl: False
    predict_xstart: False
    rescale_timesteps: False
    rescale_learned_sigmas: True
    timestep_respacing: ""

train:
  lr: 1e-4
  batch: [32, 4]   # batchsize for training and validation
  microbatch: 8
  use_fp16: False
  num_workers: 16
  prefetch_factor: 2
  iterations: 800000
  weight_decay: 0
  scheduler: step   # step or cosin
  milestones: [10000, 800000]
  ema_rates: [0.999]
  save_freq: 10000
  val_freq: 5000
  log_freq: [1000, 2000]

data:
  train:
    type: face
    params:
      ffhq_txt: ./datapipe/files_txt/ffhq512.txt
      out_size: 512
      transform_type: face