configs/hypersim_er_v5_appearance2x_train.yaml

seed: 17
device: auto
dataset:
  name: hypersim
  root: data/hypersim_er_prepared/train
  image_size: [128, 160]
  context_views: 4
  target_views: 2
  length: 16924
  sample_stride: 2
  spatial_protocol: context_local_metric
  local_bounds_m: [[-6.0, -4.0, 0.05], [6.0, 4.0, 12.05]]
model:
  mode: fixed
  state_architecture: evidence_residual
  voxel_resolution: [48, 32, 48]
  image_feature_dim: 48
  state_feature_dim: 48
  refinement_blocks: 4
  evidence_temperature: 0.1
  observed_residual_floor: 0.1
  completion_residual_scale: 4.0
  appearance_resolution_scale: 2
  n_samples: 64
  ray_chunk_size: 4096
training:
  output_dir: outputs/hypersim_er_v5_appearance2x_train
  batch_size: 1
  learning_rate: 0.0002
  max_steps: 150000
  num_workers: 0
  amp: true
  amp_dtype: bfloat16
  log_every: 20
  checkpoint_every: 5000
  measurements: [rgb, depth, normal, point, visibility]
  gradient_accumulation: 2
  weight_decay: 0.0001
  lr_schedule: warmup_cosine
  warmup_steps: 3000
  min_lr_ratio: 0.05
  ema_decay: 0.98
  evidence_residual_weight: 0.001