configs/hypersim_a_plus_train.yaml

seed: 17
device: auto
dataset:
  name: hypersim
  root: data/hypersim_prepared/train
  image_size: [128, 160]
  context_views: 2
  target_views: 1
  length: 5841
  sample_stride: 5
  spatial_protocol: context_local_metric
  local_bounds_m: [[-6.0, -4.0, 0.05], [6.0, 4.0, 12.05]]
model:
  mode: fixed
  voxel_resolution: [48, 32, 48]
  image_feature_dim: 32
  state_feature_dim: 32
  refinement_blocks: 2
  n_samples: 64
  ray_chunk_size: 4096
training:
  output_dir: outputs/hypersim_a_plus_v2_cell_centered_train
  batch_size: 1
  learning_rate: 0.0002
  max_steps: 50000
  num_workers: 0
  amp: true
  amp_dtype: bfloat16
  log_every: 20
  checkpoint_every: 1000
  measurements: [rgb, depth, normal, point, visibility]
  weight_decay: 0.0001
  lr_schedule: warmup_cosine
  warmup_steps: 1000
  min_lr_ratio: 0.05
  ema_decay: 0.98