default.yaml

category: 'human_nerf'


##############################################3
## Network Specs

# modules
network_module: 'core.nets.human_nerf.network'
trainer_module: 'core.train.trainers.human_nerf.trainer'
lr_updater_module: 'core.train.trainers.human_nerf.lr_updaters.exp_decay'
optimizer_module: 'core.train.optimizers.human_nerf.optimizer'

# positional embedder -- canonical mlp
embedder:
  module: "core.nets.human_nerf.embedders.fourier"

# positional embedder -- non-rigid motion mlp
non_rigid_embedder:
  module: "core.nets.human_nerf.embedders.hannw_fourier"

# canonical mlp
canonical_mlp:
  module: 'core.nets.human_nerf.canonical_mlps.mlp_rgb_sigma'
  mlp_depth: 8         # layers in network
  mlp_width: 256       # channels per layer
  multires: 10         # log2 of max freq for positional encoding (3D location)
  i_embed: 0           # set 0 for default positional encoding, -1 for none

# motion weights volume
mweight_volume:
  module: 'core.nets.human_nerf.mweight_vol_decoders.deconv_vol_decoder'
  embedding_size: 256
  volume_size: 32
  dst_voxel_size: 0.0625

# non-rigid motion mlp
non_rigid_motion_mlp:
  module: 'core.nets.human_nerf.non_rigid_motion_mlps.mlp_offset'
  condition_code_size: 69
  mlp_width: 128
  mlp_depth: 6
  skips: [4]
  multires: 6       # log2 of max freq for positional encoding (3D location)
  i_embed: 0        # set 0 for default positional encoding, -1 for none
  kick_in_iter: 10000
  full_band_iter: 50000

# pose decoder
pose_decoder:
  module: 'core.nets.human_nerf.pose_decoders.mlp_delta_body_pose'
  embedding_size: 69
  mlp_width: 256
  mlp_depth: 4


##############################################3
## Data Configuration

train_keyfilter: ['rays',
                  'motion_bases', 'motion_weights_priors',
                  'cnl_bbox', 'dst_posevec_69']
test_keyfilter: ['rays', 'target_rgbs', 
                 'motion_bases', 'motion_weights_priors',
                  'cnl_bbox', 'dst_posevec_69', 'pts']

train:
  use_input_annot: False
  use_huggingface: False
  min_bodypart_pixels: 2000
  sil_loss_inside_only: False
  freeze_rigid_motion: False
  train_albedo_only: False
  resize_to_patchsize_by_mask: False
  use_patch_dict: False
  use_best_fit: False
  
  with_vit: False
  with_clip: False
  reid_model: ''
  start_frame: 0
  perturb: 1.        # only for training, set to 0. for no jitter, 1. for jitter
  batch_size: 1
  shuffle: True
  drop_last: False
  maxiter: 400000
  lr: 0.0005  # 5e-4
  lr_mweight_vol_decoder: 0.00005 # 5e-5
  lr_pose_decoder: 0.00005        # 5e-5
  lr_non_rigid_mlp: 0.00005       # 5e-5
  lrate_decay: 500
  optimizer: 'adam'
  log_interval: 20
  save_checkpt_interval: 2000
  save_model_interval: 50000
  ray_shoot_mode: 'patch'
  lossweights:
    lpips: 1.0
    mse: 0.2
  vit_resolution: 224


progress:
  batch_size: 1
  shuffle: False
  drop_last: False
  dump_interval: 5000

movement:
  batch_size: 1
  shuffle: False
  drop_last: False

freeview:
  batch_size: 1
  shuffle: False
  drop_last: False
  frame_idx: 0
  render_gt_view: False
  use_gt_camera: False
  dataset: ''

mesh:
  batch_size: 1
  shuffle: False
  drop_last: False
  frame_idx: 0
  render_gt_view: True
  dataset: ''
  for_tet: False

tpose:
  batch_size: 1
  shuffle: False
  drop_last: False


##############################################3
## Misc

sex: 'neutral'
total_bones: 24
bbox_offset: 0.3

load_net: latest
save_all: True    # save all checkpoints

patch:
  sample_subject_ratio: 0.8
  N_patches: 6
  size: 32      # [Patch] size of patch

N_samples: 128      # number of samples for each ray in coarse ray matching

perturb: 1.        # only for training, set to 0. for no jitter, 1. for jitter

netchunk_per_gpu: 300000 # number of pts sent through network in parallel, decrease if running out of memory
chunk: 32768   # 32768=1024*32, number of rays processed in parallel, decrease if running out of memory
n_gpus: 1

show_alpha: False  
show_truth: False
show_depth: False


textureless: False
ambient_light: 0.5
point_light: 0.5

segments:
  body: 236
  rightArm: 216
  leftArm: 196
  rightHand: 176
  leftHand: 156
  rightUpLeg: 136
  leftUpLeg: 116
  rightLeg: 96
  leftLeg: 76
  head: 56
  neck: 36
symmetric:
  leftArm: rightArm
  leftHand: rightHand
  leftUpLeg: rightUpLeg
  leftLeg: rightLeg
  body: body

fps: 60

voxel_size: [0.005, 0.005, 0.005]
mesh_th: 5.
render_th: 0.