dataset:
  type: old_KITTI_360
  data_path: /home/fb20user12/datasets/KITTI-360/
  pose_path: /home/fb20user12/datasets/KITTI-360/data_poses
  split_path: datasets/kitti_360/splits/sscbench
  image_size:
  - 192
  - 640
  data_stereo: true
  data_fisheye: true
  data_fc: 2
  fisheye_offset:
  - 10
  - 11
  - 12
  - 13
  - 14
  - 15
  - 16
  - 17
  - 18
  - 19
  - 20
  - 21
  - 22
  - 23
  - 24
  - 25
  - 26
  - 27
  - 28
  - 29
  - 30
  - 31
  - 32
  - 33
  - 34
  - 35
  - 36
  - 37
  - 38
  - 39
  - 40
  stereo_offset:
  - 0
  is_preprocessed: true
  fisheye_rotation: -15
  data_segmentation: true
  steps_per_epoch: 5164
model:
  arch: BTSNet
  use_code: true
  prediction_mode: default
  predict_dino: true
  dino_dims: 64
  compensate_artifacts: true
  flip_augmentation: true
  encoder:
    type: dinov2
    mode: downsample-prediction
    decoder_arch: dpt
    downsampler_arch: featup
    encoder_arch: vit-b
    version: v2
    separate_gt_version: v2
    encoder_freeze: false
    flip_avg_gt: false
    dim_reduction_arch: mlp
    num_ch_enc:
    - 64
    - 64
    - 128
    - 256
    intermediate_features:
    - 3
    - 6
    - 9
    decoder_out_dim: 256
    dino_pca_dim: 64
    image_size:
    - 192
    - 640
    key_features: false
  code:
    num_freqs: 6
    freq_factor: 1.5
    include_input: true
  decoder_heads:
  - type: resnet
    name: normal_head
    freeze: false
    args:
      n_blocks: 0
      d_hidden: 128
  final_prediction_head: normal_head
  encoding_strategy:
    name: default
    args: {}
  eval_encoding_strategy:
    name: default
    args: null
  loss_renderer_strategy:
    name: kitti_360
    args: null
  eval_loss_renderer_strategy:
    name: single_renderer
    args:
      shuffle_frames: false
      all_frames: true
  inv_z: true
  learn_empty: false
  code_mode: z
  n_frames_render: 4
  sample_radius_3d: 0.5
renderer:
  n_coarse: 32
  n_fine: 0
  n_fine_depth: 0
  depth_std: 1.0
  sched: []
  white_bkgd: false
  lindisp: true
  hard_alpha_cap: true
  eval_batch_size: 65536
  render_mode: volumetric
  normalize_dino: true
training:
  optimizer:
    type: adam
    args:
      lr: 0.0005
      betas:
      - 0.9
      - 0.999
      eps: 1.0e-08
      weight_decay: 0.0
      amsgrad: false
  loss:
  - type: stego
    random_weight: 0.7681858818342623
    knn_weight: 0.40262895957374445
    self_weight: 0.2478902214214737
    random_shift: 0.8167245534406465
    knn_shift: 0.11229105513980008
    self_shift: 0.5397087004143353
    pointwise: false
  num_epochs: 1
  epoch_length: 1000
  resume_from: /home/fb20user12/code/bts-thesis-dino/out/paper-runs/kitti-dinov2/training_checkpoint_210000.pt
  continue: false
  checkpoint_every: 5000
  log_every_iters: 250
  ray_sampler:
    z_near: 3
    z_far: 80
    sample_mode: image
validation:
  validation:
    metrics:
    - type: seg
      agg_type: unsup_seg
      args:
        n_classes: 19
        gt_classes: 19
    - type: stego
      agg_type: concat
    subset:
      type: random
      args:
        size: 32
    save_best:
      metric: stego_cluster_weighted_miou
      update_model: true
      dry_run: false
    log_loss: false
    global_step:
      type: trainer iteration
    events:
    - type: ITERATION_COMPLETED
      args:
        every: 100
  visualization_seg:
    metrics: {}
    subset:
      type: range
      args:
        start: 300
        end: 301
    visualize:
      input_imgs: null
      dino_gt: null
      batch_dino_gt: null
      batch_dino_artifacts: null
      segs_gt: null
      segs_pred: null
      batch_reconstructed_dino: null
      batch_dino_features_kmeans: null
      depth: null
    log_loss: false
    global_step:
      type: trainer iteration
    events:
    - type: STARTED
      args: null
    - type: ITERATION_COMPLETED
      args:
        every: 100
downstream:
  type: segmentation
  n_classes: 19
  gt_classes: 19
  input_dim: 768
  code_dim: 64
  knn_neighbors: 4
  buffer_size: 256
  patch_sample_size: 576
  mode: 3d
sweep:
  study_name: sweep_dinov2
  storage_url: sqlite:////home/fb20user12/sweeps/optuna_study.db
  direction: maximize
  n_trials: 50
  start_original_param: true
  hparams:
  - key: training.loss.0.random_weight
    method: suggest_float
    kwargs:
      name: random_weight
      low: 0.0
      high: 1.0
  - key: training.loss.0.knn_weight
    method: suggest_float
    kwargs:
      name: knn_weight
      low: 0.0
      high: 1.0
  - key: training.loss.0.self_weight
    method: suggest_float
    kwargs:
      name: self_weight
      low: 0.0
      high: 1.0
  - key: training.loss.0.random_shift
    method: suggest_float
    kwargs:
      name: random_shift
      low: 0.6
      high: 1.0
  - key: training.loss.0.knn_shift
    method: suggest_float
    kwargs:
      name: knn_shift
      low: 0.0
      high: 0.4
  - key: training.loss.0.self_shift
    method: suggest_float
    kwargs:
      name: self_shift
      low: 0.2
      high: 0.6
training_type: downstream_training
mode: nvs
seed: 66
backend: null
nproc_per_node: null
with_amp: true
name: training
batch_size: 8
gradient_accum_factor: 1
num_workers: 8
output:
  path: out/sweep_dinov2/sweep_66
  unique_id: sweep_66
  original_path: out/sweep_dinov2/
  original_unique_id: sweep
eval_seed: 65
cuda device name: Tesla V100-SXM3-32GB