init/config.yaml

dataset:
  view_sampler:
    name: boundedv2
    num_target_views: 6
    num_context_views: 8
    min_distance_between_context_views: 24
    max_distance_between_context_views: 45
    max_distance_to_context_views: 0
    context_gap_warm_up_steps: 10000
    target_gap_warm_up_steps: 0
    initial_min_distance_between_context_views: 20
    initial_max_distance_between_context_views: 30
    initial_max_distance_to_context_views: 0
    extra_views_sampling_strategy: farthest_point
    target_views_replace_sample: false
  name: dl3dv
  roots:
  - datasets/dl3dv-480p-chunks
  make_baseline_1: false
  augment: true
  image_shape:
  - 256
  - 448
  background_color:
  - 0.0
  - 0.0
  - 0.0
  cameras_are_circular: false
  baseline_epsilon: 0.001
  max_fov: 100.0
  skip_bad_shape: true
  near: 0.01
  far: 200.0
  baseline_scale_bounds: false
  shuffle_val: true
  test_len: -1
  test_chunk_interval: 1
  sort_target_index: true
  sort_context_index: true
  train_times_per_scene: 1
  test_times_per_scene: 1
  ori_image_shape:
  - 270
  - 480
  overfit_max_views: 148
  use_index_to_load_chunk: false
  mix_tartanair: false
  no_mix_test_set: true
  load_depth: false
  center_pose: false
  opencv_pose_format: null
  overfit_to_scene: null
  min_views: 0
  max_views: 0
  highres: false
  mix_re10k: false
  re10k_min_view_dist: 40
  re10k_max_view_dist: 300
  pose_align_first_view: false
  scale_extrinsics: 1.0
  metric_scale_align_dl3dv: false
  load_remain_context: false
  num_remain_context: 8
  random_crop: false
  min_size:
  - 384
  - 512
  max_size:
  - 512
  - 960
scene_trainer:
  scene_initializer:
    name: resplat
    num_depth_candidates: 128
    num_surfaces: 1
    gaussians_per_pixel: 1
    gaussian_adapter:
      gaussian_scale_min: 0.5
      gaussian_scale_max: 0.3
      sh_degree: 3
      exp_scale: false
      softplus_scale: true
      clamp_min_scale: 1.0e-06
      scale_detach_depth: false
      exp_scale_bias: 4.0
      no_rotate_sh: true
      no_sh_mask: true
      init_rotation_identity: false
    d_feature: 128
    visualizer:
      num_samples: 8
      min_resolution: 256
      export_ply: false
    unimatch_weights_path: pretrained/gmdepth-scale1-resumeflowthings-scannet-5d9d7964.pth
    multiview_trans_attn_split: 2
    costvolume_unet_feat_dim: 128
    costvolume_unet_channel_mult:
    - 1
    - 1
    - 1
    costvolume_unet_attn_res: []
    depth_unet_feat_dim: 64
    depth_unet_attn_res: []
    depth_unet_channel_mult:
    - 1
    - 1
    - 1
    downscale_factor: 4
    shim_patch_size: 4
    local_mv_match: 2
    monodepth_vit_type: vits
    supervise_intermediate_depth: true
    return_depth: true
    num_scales: 1
    upsample_factor: 8
    lowest_feature_resolution: 8
    depth_unet_channels: 128
    grid_sample_disable_cudnn: false
    large_gaussian_head: false
    color_large_unet: false
    init_sh_input_img: true
    feature_upsampler_channels: 64
    gaussian_regressor_channels: 256
    unet_gaussian_regressor: false
    resnet_gaussian_regressor: false
    train_depth_only: false
    pt_head: true
    pt_heads: 1
    init_pt_with_mv_attn: false
    init_pt_with_mv_attn_lowres: false
    pt_head_channels: null
    pt_head_concat_img: false
    pt_head_conv: false
    multi_scale_pt: false
    attn_proj_channels: 64
    fps_num_samples: null
    knn_samples: 16
    post_norm: false
    no_rpe: true
    no_knn_attn: false
    num_blocks: 4
    pt_downsample: 0
    fps_agg_func: attn
    subsample_method: fps
    add_pt_residual: true
    pt_pred_residual_position: false
    freeze_depth: false
    use_gt_depth: false
    separate_depth_color: false
    separate_depth_type: small
    separate_depth_gaussian_scale: false
    sample_log_depth: true
    bilinear_upsample_depth: false
    no_upsample_depth: false
    return_lowres_depth: false
    foundationstereo: false
    fstereo_num_refine: 1
    lvsm_gaussian_regressor: false
    lvsm_layers: 6
    latent_gs: true
    latent_downsample: 4
    fixed_latent_size: true
    latent_gs_img_interp: area
    dpt_head_depth: false
    latent_dpt_upsampler: false
    latent_dpt_upsampler_no_concat: false
    light_dpt_feature: false
    avgpool_depth: false
    nearest_down_depth: false
    predict_scale: false
    norm_by_points: false
    no_pred_depth_range: false
    point_dist_init_gaussian_scale: false
    resizeconv_upsampler: false
    depth_pred_half_res: false
    use_amp: true
    pt_head_amp: true
    use_fsdp: false
    use_checkpointing: false
    init_use_checkpointing: false
    new_gaussian_order: true
    rotate_quat_to_world: false
    refine_rotate_quat_to_world: false
    refine_no_use_covariance: false
    latent_new_reshape: false
    ptv3: false
    ptv3_grid_size: 0.1
    no_pixel_offset: false
    init_gaussian_multiple: 1
    deform_sample_depth: false
    deform_sample_depth_debug: false
  scene_optimizer:
    name: depthsplat
    no_refine_rotation: false
    no_refine_mean: false
    no_render_error: false
    refine_sh_only: false
    num_basic_refine_blocks: 4
    num_refine_blocks: 1
    refine_lpips_error: false
    refine_pool_vgg_features: false
    refine_use_all_vgg_features: false
    refine_vit_feature: false
    refine_resnet_feature: false
    no_freeze_resnet_feature: false
    shallow_resnet_feature: false
    resnet_feature_layers: 18
    refine_convnext_feature: false
    convnext_feature_size: small
    refine_concat_feature: false
    refine_concat_feature_cosine: false
    refine_cosine_feature: false
    refine_add_feature: false
    refine_concat_rgb_feature_error: false
    concat_init_state: true
    replace_init_state: false
    state_channels: 0
    refine_block_rmsnorm: false
    refine_block_layernorm: false
    pt_qk_norm: false
    norm_pt_block: false
    refine_gaussian_multiple: 1
    refine_residual_init_state: false
    clamp_refine_max_scale: 3.0
    refine_no_rpe: false
    gaussian_head_multiple: 1
    update_head_concat_img: false
    update_head_act: gelu
    update_attn_proj_channels: null
    update_no_knn_attn: false
    update_no_tran_block_norm: false
    update_tran_block_act: gelu
    multi_gaussian_scale_smaller: false
    init_gaussian_multiple: 1
    no_pixel_offset: false
    refine_condition_pt_feature: false
    render_error_no_abs: false
    reinit_gaussian_when_refine_multiple: false
    refine_same_num_points: false
    render_rgb_error_no_shuffle: false
    refine_cache_resnet_feature: false
    refine_view_pool_resnet_feature: false
    refine_global_pool_resnet_feature: false
    pt_heads: 1
    refine_with_mv_attn: false
    refine_with_mv_attn_lowres: false
    refine_no_mv_attn: false
    mv_attn_conv_with_norm: false
    refine_mv_shuffle_attn: false
    refine_mv_attn_with_pos_enc: false
    refine_shuffle_attn_no_norm: false
    refine_mv_unimatch_attn: false
    refine_knn_samples: 16
    refine_multi_scale_pt: false
    refine_input_alpha: false
    refine_input_depth: false
    refine_input_depth_smooth_error: false
    refine_input_error: true
    refine_input_zero_gaussian: false
    refine_input_zero_state: false
    refine_output_scale_mag: false
    cross_attn_render_error: false
    cross_attn_render_error_blocks: 1
    cross_attn_with_mlp: false
    radii_averaged_render_error: false
    cross_attn_additional_render_error: false
    num_intermediate_views: 8
    render_error_mv_attn: false
    render_error_mv_attn_blocks: 2
    render_error_mv_attn_lowres: false
    mv_attn_proj_channels: 0
    lowres_attn_down_factor: 4
    lowres_attn_pos_enc: false
    render_error_mv_attn_swin: false
    render_error_num_views: 0
    render_error_remain_context: false
    render_error_merge_remain_context: false
    render_error_warp_remain_context: false
    render_error_random_num_remain_context: false
    render_error_num_remain_context_test: 0
    render_error_warp_input_view: false
    window_local_refine: false
    window_global_refine: false
    window_local_global_refine: false
    update_window_size: 0
    local_gaussian_render: false
    render_error_half_res: false
    prune_opacity: 0.0
    prune_radii: 0
    refine_input_gradient: false
    refine_input_gradient_log: false
    refine_input_gradient_log_clip_deltas: 0.001
    refine_input_gradient_scale: 1.0
    gradient_update_scale: 1.0
    refine_gradient_with_ssim_loss: false
    refine_input_gradient_same_loss: false
    train_min_refine: 0
    train_max_refine: 0
    local_global_update: false
    num_global_update: 0
    train_global_update_only: false
    random_update_with_size: false
    ptv3: false
    ptv3_grid_size: 0.1
    use_amp: true
    pt_head_amp: true
    pt_update_amp: true
    use_fsdp: false
    use_checkpointing: false
    init_use_checkpointing: false
    debug_refine_update_module: true
    debug_refine_mlp: false
    debug_refine_mlp_layers: 0
    debug_refine_mlp_fixed_init: false
  decoder:
    name: gsplat
    scale_invariant: false
  initializer:
    num_depth_candidates: 128
    costvolume_unet_feat_dim: 128
    costvolume_unet_channel_mult:
    - 1
    - 1
    - 1
    costvolume_unet_attn_res:
    - 4
    gaussians_per_pixel: 1
    depth_unet_feat_dim: 32
    depth_unet_attn_res:
    - 16
    depth_unet_channel_mult:
    - 1
    - 1
    - 1
    - 1
    - 1
    shim_patch_size: 16
  use_fsdp: false
  train_scene_init: true
  train_scene_opt: false
  num_update_steps: 0
loss:
  mse:
    weight: 1.0
  lpips:
    weight: 0.5
    apply_after_step: 0
    perceptual_loss: true
test:
  postprocessing:
    __target__: src.model.postprocessing.PostProcessCfg
    enabled: false
    name: none
    steps: 0
    compute_metrics_every: 0
    lr: 0
    scheduler: null
    scheduler_warm_up_ratio: 0.0
  output_path: null
  compute_scores: true
  eval_time_skip_steps: 0
  save_image: false
  save_video: false
  save_gt_image: false
  save_input_images: false
  save_depth: false
  save_depth_npy: false
  save_depth_concat_img: false
  save_gaussian: false
  no_align_to_view: false
  save_point_cloud: false
  render_chunk_size: null
  stablize_camera: false
  stab_camera_kernel: 50
  render_input_views: false
  inference_window_size: null
  profile_model: false
  save_colmap_train_test_views: false
  ori_colmap_data_path: null
  adam_optimizer_step: 0
  save_every_freq: null
  save_every_steps: null
  dec_chunk_size: 30
wandb:
  project: unified-dl3dv-8views
  entity: placeholder
  name: dl3dv
  mode: online
  id: null
  notes: null
  tags:
  - dl3dv
  - 270x480
mode: train
data_loader:
  train:
    num_workers: 10
    persistent_workers: true
    batch_size: 2
    seed: 1234
  test:
    num_workers: 4
    persistent_workers: false
    batch_size: 1
    seed: 2345
  val:
    num_workers: 1
    persistent_workers: true
    batch_size: 1
    seed: 3456
meta_optimizer:
  lr: 0.0002
  lr_monodepth: 2.0e-06
  lr_depth: 0.0
  warm_up_steps: 2000
  weight_decay: 0.01
  warm_up_ratio: 0.01
  adamw_8bit: false
checkpointing:
  load: null
  every_n_train_steps: 1000
  save_top_k: 5
  pretrained_model: null
  pretrained_model_rel_dir: ${checkpoint_rel_dir:${checkpointing.pretrained_model}}
  pretrained_monodepth: null
  pretrained_mvdepth: null
  pretrained_depth: pretrained/depthsplat-depth-small-352x640-samplelogdepth-b0ebc084.pth
  pretrained_scale_predictor: null
  pretrained_depth_teacher: null
  no_strict_load: false
  resume: false
  no_resume_upsampler: false
  partial_load: false
  freeze_mono_vit: false
  resume_update_module: null
train:
  depth_mode: null
  extended_visualization: false
  print_log_every_n_steps: 100
  eval_model_every_n_val: 2
  eval_data_length: 999999
  eval_deterministic: false
  eval_time_skip_steps: 3
  eval_save_model: true
  l1_loss: true
  intermediate_loss_weight: 0.9
  no_viz_video: false
  eval_depth: false
  forward_depth_only: false
  train_ignore_large_loss: 0.0
  no_log_projections: true
  no_log_video: true
  depth_loss_weight: 0.0
  log_depth_loss: true
  depth_smooth_loss_weight: 0.01
  depth_smooth_loss_nonorm: false
  depth_smooth_loss_weight_nvs: 0.0
  monodepth_loss_weight: 0.0
  depth_teacher_loss_weight: 0.0
  viz_depth_teacher: false
  eval_render_depth: false
  render_depth_loss_weight: 0.0
  viz_render_depth: false
  use_gt_depth_range: false
  depth_range_from_disparity: false
  max_disparity: 128.0
  min_disparity: 4.0
  img_warp_loss_weight: 0.0
  warp_loss_start_step: 5000
  loss_on_input_views: false
  train_window_size: null
  half_res_lpips_loss: false
  viz_depth_separate: false
seed: 111123
meta_trainer:
  max_steps: 100000
  val_check_interval: 0.5
  gradient_clip_val: 0.5
  num_sanity_val_steps: 2
  eval_index: null
  limit_test_batches: 1.0
  limit_train_batches: 1.0
  num_nodes: 1
output_dir: !!python/object/apply:pathlib.PosixPath
- checkpoints
- optgs
- unified-dl3dv-8views
- init
use_plugins: false
log_slurm_id: true