thuman.yaml

model:
  base_learning_rate: 5e-5
  target: ldm.models.diffusion.morphable_diffusion.SyncMultiviewDiffusion
  params:
    view_num: 16
    image_size: 256
    cfg_scale: 2.0
    output_num: 8
    batch_view_num: 4
    finetune_unet: True
    drop_conditions: false
    projection: 'orthographic'
    use_spatial_volume: False
    clip_image_encoder_path: ./ckpt/ViT-L-14.pt
    target_elevation: 0

    scheduler_config: # 10000 warmup steps
      target: ldm.lr_scheduler.LambdaLinearScheduler
      params:
        warm_up_steps: [ 100 ]
        cycle_lengths: [ 100000 ]
        f_start: [ 0.02 ]
        f_max: [ 1.0 ]
        f_min: [ 1.0 ]

    unet_config:
      target: ldm.models.diffusion.attention.DepthWiseAttention
      params:
        volume_dims: [64, 128, 256, 512]
        image_size: 32
        in_channels: 8
        out_channels: 4
        model_channels: 320
        attention_resolutions: [ 4, 2, 1 ]
        num_res_blocks: 2
        channel_mult: [ 1, 2, 4, 4 ]
        num_heads: 8
        use_spatial_transformer: True
        transformer_depth: 1
        context_dim: 768
        use_checkpoint: True
        legacy: False

data:
  target: ldm.data.thuman.THumanDataset
  params:
    data_dir: /cluster/scratch/xiychen/data/thuman_2.1_preprocessed
    smplx_dir: /cluster/scratch/xiychen/data/thuman_smplx # a list of uids
    batch_size: 70 # batch size for a single gpu
    num_workers: 1

lightning:
  modelcheckpoint:
    params:
      every_n_train_steps: 2000 # we will save models every 1k steps
  callbacks:
    {}

  trainer:
    benchmark: True
    max_steps: 6000
    val_check_interval: 250 # we will run validation every 1k steps, the validation will output images to <log_dir>/<images>/val
    num_sanity_val_steps: 0
    precision: 32
    check_val_every_n_epoch: null
    accumulate_grad_batches: 1