training:
  resume: False  # If True, must set hydra.run.dir accordingly
  pretrain_path: ""
  interval_visualize: 1000
  interval_save_checkpoint: 5000
  interval_delete_checkpoint: 10000
  interval_evaluate: 5000
  delete_all_checkpoints_after_training: False
  lr: 1e-4
  mixed_precision: True
  matmul_precision: high 
  max_iterations: 100000
  batch_size: 64
  num_workers: 8
  gpu_id: 0
  freeze_encoder: True
  seed: 0
  job_key: ""  # Use this for submitit sweeps where timestamps might collide
  translation_scale: 1.0
  regression: False
  prob_unconditional: 0
  load_extra_cameras: False
  calculate_intrinsics: False
  distort: False
  normalize_first_camera: True
  diffuse_origins_and_endpoints: True
  diffuse_depths: False
  depth_resolution: 1
  dpt_head: False
  full_num_patches_x: 16
  full_num_patches_y: 16
  dpt_encoder_features: True
  nearest_neighbor: True
  no_bg_targets: True
  unit_normalize_scene: False
  sd_scale: 2
  bfloat: True
  first_cam_mediod: True
  gradient_clipping: False
  l1_loss: False
  grad_accumulation: False
  reinit: False

model:
  pred_x0: True
  model_type: dit
  num_patches_x: 16
  num_patches_y: 16
  depth: 16
  num_images: 1
  random_num_images: True
  feature_extractor: dino
  append_ndc: True
  within_image: False
  use_homogeneous: True
  freeze_transformer: False
  cond_depth_mask: True

noise_scheduler:
  type: linear
  max_timesteps: 100
  beta_start: 0.0120
  beta_end: 0.00085
  marigold_ddim: False

dataset:
  name: co3d
  shape: all_train
  apply_augmentation: True
  use_global_intrinsics: True
  mask_holes: True
  image_size: 224

debug:
  wandb: True
  project_name: diffusionsfm
  run_name: 
  anomaly_detection: False

hydra:
  run:
    dir: ./output/${now:%m%d_%H%M%S_%f}${training.job_key}
  output_subdir: hydra