configs/gslrm.yaml

# =============================================================================
# General Configuration
# =============================================================================
profile: false
debug: false

# =============================================================================
# Model Configuration
# =============================================================================
model:
  class_name: gslrm.model.gslrm.GSLRM
  
  # Image processing settings
  image_tokenizer:
    image_size: 512
    patch_size: 8
    in_channels: 9  # 3 RGB + 3 direction + 3 Reference

  # Transformer architecture
  transformer:
    d: 1024
    d_head: 64
    n_layer: 24
  
  # Gaussian splatting configuration
  gaussians:
    n_gaussians: 2  # 12288
    sh_degree: 0
    
    upsampler:
      upsample_factor: 1

  # Model behavior flags
  add_refsrc_marker: false
  hard_pixelalign: true
  use_custom_plucker: true
  clip_xyz: true

# =============================================================================
# Training Configuration
# =============================================================================
training:
  # Training runtime settings
  runtime:
    use_tf32: true
    use_amp: true
    amp_dtype: "bf16"
    torch_compile: false
    grad_accum_steps: 1
    grad_clip_norm: 1.0
    grad_checkpoint_every: 1
  
  # Dataset configuration
  dataset:
    dataset_path: "data_sample/gslrm/data_gslrm_train.txt"
    
    # View configuration
    maximize_view_overlap: true
    num_views: 8
    num_input_views: 6 # In training, we set it as 4. In inference, we set it as 6.
    target_has_input: true
    
    # Data preprocessing
    normalize_distance_to: 0.0
    remove_alpha: false
    background_color: "white"
  
  # Data loader settings
  dataloader:
    batch_size_per_gpu: 2
    num_workers: 4
    num_threads: 32
    prefetch_factor: 32
  
  # Loss function weights
  losses:
    l2_loss_weight: 1.0
    lpips_loss_weight: 0.0
    perceptual_loss_weight: 0.5
    ssim_loss_weight: 0.0
    pixelalign_loss_weight: 0.0
    masked_pixelalign_loss: true
    pointsdist_loss_weight: 0.0
    warmup_pointsdist: false
    distill_loss_weight: 0.0
  
  # Optimizer configuration (AdamW)
  optimizer:
    lr: 0.0001
    beta1: 0.9
    beta2: 0.95
    weight_decay: 0.05
    reset_lr: false
    reset_weight_decay: false
    reset_training_state: true
  
  # Training schedule
  schedule:
    num_epochs: 100000          # dataset epochs
    early_stop_after_epochs: 100000  # 40
    max_fwdbwd_passes: 20000    # forward/backward pass steps
    warmup: 500                 # parameter update steps
    l2_warmup_steps: 500
  
  # Checkpointing
  checkpointing:
    resume_ckpt: "checkpoints/gslrm/stage_2"
    checkpoint_every: 5000       # forward/backward pass steps
    checkpoint_dir: "checkpoints/gslrm/stage_3"
  
  # Logging and monitoring
  logging:
    print_every: 20             # forward/backward pass steps
    vis_every: 250              # forward/backward pass steps
    
    # Weights & Biases configuration
    wandb:
      project: "facelift_gslrm"
      exp_name: "stage_3"
      group: "facelift"
      job_type: "train"
      log_every: 50             # forward/backward pass steps
      offline: false


# =============================================================================
# Inference Configuration
# =============================================================================
inference:
  enabled: false
  output_dir: "outputs/inference/gslrm/stage_3"

# =============================================================================
# Validation Configuration
# =============================================================================
validation:
  enabled: true
  val_every: 5000
  output_dir: "outputs/validation/gslrm/stage_3"
  dataset_path: "data_sample/gslrm/data_gslrm_val.txt"