configs/conditional_pose_diffusion.yaml

random_seed: 1

WANDB:
  project: StructDiffusion
  save_dir: ${base_dirs.wandb_dir}
  name: conditional_pose_diffusion

DATASET:
  data_root: ${base_dirs.data}
  vocab_dir: ${base_dirs.data}/type_vocabs_coarse.json

  # important
  use_virtual_structure_frame: True
  ignore_distractor_objects: True
  ignore_rgb: True

  # the following are determined by the dataset
  max_num_target_objects: 7
  max_num_distractor_objects: 5
  max_num_shape_parameters: 5
  # set to zeros because they are not used for now
  max_num_rearrange_features: 0
  max_num_anchor_features: 0

  num_pts: 1024
  filter_num_moved_objects_range:
  data_augmentation: False

DATALOADER:
  batch_size: 64
  num_workers: 8
  pin_memory: True

MODEL:
  # transformer encoder
  encoder_input_dim: 256
  num_attention_heads: 8
  encoder_hidden_dim: 512
  encoder_dropout: 0.0
  encoder_activation: relu
  encoder_num_layers: 8
  # output head
  structure_dropout: 0
  object_dropout: 0
  # pc encoder
  ignore_rgb: ${DATASET.ignore_rgb}
  pc_emb_dim: 256
  posed_pc_emb_dim: 80
  # pose encoder
  pose_emb_dim: 80
  # language
  word_emb_dim: 160
  # diffusion step
  time_emb_dim: 80
  # sequence embeddings
  # max_num_target_objects (+ max_num_distractor_objects if not ignore_distractor_objects)
  max_seq_size: 7
  max_token_type_size: 4
  seq_pos_emb_dim: 8
  seq_type_emb_dim: 8
  # virtual frame
  use_virtual_structure_frame: ${DATASET.use_virtual_structure_frame}

NOISE_SCHEDULE:
  timesteps: 200

LOSS:
  type: huber

OPTIMIZER:
  lr: 0.0001
  weight_decay: 0  #0.0001
  # lr_restart: 3000
  # warmup: 10

TRAINER:
  max_epochs: 200
  gradient_clip_val: 1.0
  gpus: 1
  deterministic: False
  # enable_progress_bar: False