configs/shape-autoencoder/michelangelo-l768-e64-ne8-nd16.yaml

exp_root_dir: "outputs"
name: "michelangelo-autoencoder/michelangelo-l768-e64-ne8-nd16"
tag: "${rmspace:n${data.n_samples}+${data.supervision_type}+rot${data.rotate_points}+noise${data.noise_sigma}+${system.shape_model.embed_type}+dsample${system.shape_model.use_downsample}+pfeat${system.shape_model.point_feats}+logits${system.loss.lambda_logits}+kl${system.loss.lambda_kl}+lr${system.optimizer.args.lr},_}"
seed: 0

data_type: "Objaverse-datamodule"
data:
  root_dir: ./data/objaverse

  load_geometry: True           # whether to load geometry
  geo_data_type: "tsdf" 
  n_samples: 10240
  load_supervision: True
  supervision_type: "occupancy" 
  n_supervision: 10240
  tsdf_threshold: 0.0078125     # threshold for truncating sdf values, used when input is sdf

  load_image: False             # whether to load images 
  load_caption: False           # whether to load captions

  batch_size: 8
  num_workers: 0

system_type: "shape-autoencoder-system"
system:
  sample_posterior: true
  
  shape_model_type: "michelangelo-autoencoder"
  shape_model:
    num_latents: 1024 # 1024
    embed_dim: 64
    point_feats: 3   # xyz + normal
    out_dim: 1 # only occupancy
    embed_type: "fourier"
    num_freqs: 8
    include_pi: false
    heads: 12
    width: 768
    num_encoder_layers: 8
    num_decoder_layers: 16
    use_ln_post: true
    init_scale: 0.25
    qkv_bias: true
    use_flash: true
    use_checkpoint: true
    use_downsample: true

  loggers:
    wandb:
      enable: false
      project: "CraftsMan"
      name: shape-autoencoder+${name}+${tag}

  loss:
    lambda_logits: 1.
    lambda_kl: 0.001

  optimizer:
    name: AdamW
    args:
      lr: 1.e-4
      betas: [0.9, 0.99]
      eps: 1.e-6

  scheduler:
    name: SequentialLR
    interval: step
    schedulers:
      - name: LinearLR
        interval: step
        args:
          start_factor: 1e-6
          end_factor: 1.0
          total_iters: 5000
      - name: CosineAnnealingLR
        interval: step
        args:
          T_max: 5000
          eta_min: 0.
    milestones: [5000]

trainer:
  num_nodes: 1
  max_epochs: 100000
  log_every_n_steps: 5
  num_sanity_val_steps: 1
  check_val_every_n_epoch: 600
  enable_progress_bar: true
  precision: 16-mixed

checkpoint:
  save_last: true
  save_top_k: -1
  every_n_train_steps: 5000