image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6-aligned-vae/config.yaml

name: michelangelo-image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6-170k
description: ''
tag: michelangelo-aligned-autoencoder+n4096+noise0.0+pfeat3+zeroemb0.0+normembFalse+lr5e-05+qkvbiasFalse+nfreq8+ln_postTrue
seed: 0
use_timestamp: true
timestamp: ''
exp_root_dir: outputs
exp_dir: outputs/michelangelo-image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6-170k
trial_name: michelangelo-aligned-autoencoder+n4096+noise0.0+pfeat3+zeroemb0.0+normembFalse+lr5e-05+qkvbiasFalse+nfreq8+ln_postTrue
trial_dir: outputs/michelangelo-image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6-170k/michelangelo-aligned-autoencoder+n4096+noise0.0+pfeat3+zeroemb0.0+normembFalse+lr5e-05+qkvbiasFalse+nfreq8+ln_postTrue
n_gpus: 8
resume: ./ckpts/3DNativeGeneration/michelangelo-image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6-170k.ckpt
data_type: objaverse-datamodule
data:
  root_dir: data/objaverse_clean/cap3d_high_quality_170k_images
  data_type: occupancy
  n_samples: 4096
  noise_sigma: 0.0
  load_supervision: false
  supervision_type: occupancy
  n_supervision: 10000
  load_image: true
  image_data_path: data/objaverse_clean/raw_data/images/cap3d_high_quality_170k
  image_type: mvrgb
  idx:
  - 0
  - 4
  - 8
  - 12
  - 16
  n_views: 4
  load_caption: false
  rotate_points: false
  batch_size: 32
  num_workers: 16
system_type: shape-diffusion-system
system:
  val_samples_json: val_data/mv_images/val_samples_rgb_mvimage.json
  z_scale_factor: 1.0
  guidance_scale: 7.5
  num_inference_steps: 50
  eta: 0.0
  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
  condition_model_type: clip-embedder
  condition_model:
    pretrained_model_name_or_path: openai/clip-vit-large-patch14
    encode_camera: true
    camera_embeds_dim: 32
    n_views: 4
    empty_embeds_ratio: 0.1
    normalize_embeds: false
    zero_uncond_embeds: true
  denoiser_model_type: simple-denoiser
  denoiser_model:
    input_channels: 64
    output_channels: 64
    n_ctx: 256
    width: 768
    layers: 6
    heads: 12
    context_dim: 1024
    init_scale: 1.0
    skip_ln: true
    use_checkpoint: true
  noise_scheduler_type: diffusers.schedulers.DDPMScheduler
  noise_scheduler:
    num_train_timesteps: 1000
    beta_start: 0.00085
    beta_end: 0.012
    beta_schedule: scaled_linear
    variance_type: fixed_small
    clip_sample: false
  denoise_scheduler_type: diffusers.schedulers.DDIMScheduler
  denoise_scheduler:
    num_train_timesteps: 1000
    beta_start: 0.00085
    beta_end: 0.012
    beta_schedule: scaled_linear
    clip_sample: false
    set_alpha_to_one: false
    steps_offset: 1
  loggers:
    wandb:
      enable: false
      project: JiangXin
      name: text-to-shape-diffusion+michelangelo-image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6-170k+michelangelo-aligned-autoencoder+n4096+noise0.0+pfeat3+zeroemb0.0+normembFalse+lr5e-05+qkvbiasFalse+nfreq8+ln_postTrue
  loss:
    loss_type: mse
    lambda_diffusion: 1.0
  optimizer:
    name: AdamW
    args:
      lr: 5.0e-05
      betas:
      - 0.9
      - 0.99
      eps: 1.0e-06
  scheduler:
    name: SequentialLR
    interval: step
    schedulers:
    - name: LinearLR
      interval: step
      args:
        start_factor: 1.0e-06
        end_factor: 1.0
        total_iters: 5000
    - name: CosineAnnealingLR
      interval: step
      args:
        T_max: 5000
        eta_min: 0.0
    milestones:
    - 5000
trainer:
  num_nodes: 2
  max_epochs: 100000
  log_every_n_steps: 5
  num_sanity_val_steps: 1
  check_val_every_n_epoch: 3
  enable_progress_bar: true
  precision: 16-mixed
  strategy: ddp_find_unused_parameters_true
checkpoint:
  save_last: true
  save_top_k: -1
  every_n_train_steps: 5000