configs/model/dino_downsampler.yaml

arch: "BTSNet"
use_code: true
prediction_mode: default

predict_dino: true
dino_dims: 64  # == encoder.pca_dino_out

compensate_artifacts: true
flip_augmentation: true

encoder:
  type: "dinov2"
  mode: "downsample-prediction"                 # upsample-gt, downsample-pred
  decoder_arch: "dpt"
  # upsampler_arch: "multiscale-crop"           # multiscale-crop, nearest
  downsampler_arch: "featup"                    # featup, bilinear
  encoder_arch: "vit-b"                         # vit-s, vit-b
  version: "v1"                                 # v1, v2, reg, fit3d
  separate_gt_version: "v1"                     # v1, v2, reg, fit3d, None
  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  # == dino_dims
  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  # number of frames to render among v==8