vae_pretrained/config.yaml

exp_root_dir: "outputs"
name: "michelangelo-autoencoder/l256-e64-ne8-nd16"
tag: michelangelo-autoencoder+n4096+noise0.0+pfeat3+normembFalse+lr5e-05+qkvbiasFalse+nfreq8+ln_postTrue
resume: ./ckpts/vae_pretrained/model.ckpt
seed: 0
data_type: "objaverse-datamodule"
data:
  root_dir: 'data/objaverse-MIX'
  data_type: "occupancy" 
  n_samples: 4096
  noise_sigma: 0.
  
  load_supervision: True
  supervision_type: "occupancy" 
  n_supervision: 10000

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

  batch_size: 8
  num_workers: 16

system_type: "shape-autoencoder-system"
system:
  sample_posterior: true
  
  shape_model_type: "michelangelo-aligned-autoencoder"
  shape_model:
    num_latents: 256
    embed_dim: 64
    point_feats: 3
    out_dim: 1
    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: false
    use_flash: true
    use_checkpoint: 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: 5e-05
      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
  # val_check_interval: 200
  check_val_every_n_epoch: 1
  enable_progress_bar: true
  precision: 16-mixed

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