shap-e/transmitter_config.yaml

encoder:
  cross_attention_dataset: pcl_and_multiview_pcl
  d_latent: 1024
  data_ctx: 1024
  fps_method: first
  heads: 8
  init_scale: 0.25
  inner_batch_size:
  - 16384
  - 20
  input_channels: 6
  latent_bottleneck:
    diffusion:
      schedule: inv_parabola
      schedule_args:
        power: 5.0
      timesteps: 1024
    diffusion_prob: 0.1
    name: clamp_diffusion_noise
  layers: 12
  max_depth: 9.0
  max_unrolls: 1
  min_unrolls: 1
  name: PointCloudPerceiverChannelsEncoder
  params_proj:
    init_scale: 1.0
    learned_scale: 0.0625
    name: channels
    use_ln: true
  patch_size: 8
  pointconv_hidden:
  - 1024
  - 1024
  pointconv_padding_mode: circular
  pointconv_patch_size: 8
  pointconv_samples: 64
  pointconv_stride: 4
  pos_emb: nerf
  use_depth: true
  use_pointconv: true
  width: 1024
name: Transmitter
renderer:
  grid_size: 128
  n_coarse_samples: 64
  n_fine_samples: 128
  name: NeRSTFRenderer
  nerstf:
    activation: swish
    d_hidden: 256
    density_activation: relu
    init_scale: 0.25
    initial_density_bias: 0.1
    insert_direction_at: 4
    meta_bias: false
    meta_parameters: true
    n_hidden_layers: 6
    n_meta_layers: 4
    name: MLPNeRSTFModel
    posenc_version: nerf
    separate_coarse_channels: true
    separate_nerf_channels: true
    trainable_meta: false
  separate_shared_samples: true
  void:
    background:
    - 0
    - 0
    - 0
    name: VoidNeRFModel
  volume:
    bbox_max:
    - 1.0
    - 1.0
    - 1.0
    bbox_min:
    - -1.0
    - -1.0
    - -1.0
    name: BoundingBoxVolume