config.json

callbacks:
  ema:
    _target_: flower.callbacks.ema.EMA
    decay: 0.999
    start_step: 0
    save_ema_weights_in_callback_state: true
    evaluate_ema_weights_instead: true
    power: 0.6666666666666666
    inv_gamma: 1.0
    min_value: 0.0
    max_value: 0.9999
  rollout_lh:
    _target_: flower.rollout.libero_rollout.RolloutLibero
    _recursive_: false
    env_cfg:
      _target_: flower.wrappers.hulc_wrapper.HulcWrapper
    skip_epochs: ${rollout_lh_skip_epochs}
    benchmark_name: ${libero_benchmark}
    rollout_freq: 10
    num_videos: 0
    num_sequences: 50
    max_steps: 520
    empty_cache: false
    debug: false
    n_eval: 50
    num_procs: 10
    use_mp: false
    task_embedding_format: clip
    device: ${device}
  checkpoint:
    _target_: pytorch_lightning.callbacks.ModelCheckpoint
    save_top_k: 1
    verbose: true
    monitor: eval_lh/avg_seq_len
    mode: max
    dirpath: saved_models
    filename: '{epoch:02d}_{eval_lh/avg_seq_len:.2f}'
    every_n_epochs: ${callbacks.rollout_lh.rollout_freq}
datamodule:
  transforms:
    train:
      rgb_static:
      - _target_: torchvision.transforms.Resize
        size: 112
        antialias: true
      - _target_: flower.utils.transforms.RandomShiftsAug
        pad: 10
      - _target_: flower.utils.transforms.ScaleImageTensor
      - _target_: torchvision.transforms.Normalize
        mean:
        - 0.48145466
        - 0.4578275
        - 0.40821073
        std:
        - 0.26862954
        - 0.26130258
        - 0.27577711
      rgb_gripper:
      - _target_: torchvision.transforms.Resize
        size: 112
        antialias: true
      - _target_: flower.utils.transforms.RandomShiftsAug
        pad: 4
      - _target_: flower.utils.transforms.ScaleImageTensor
      - _target_: torchvision.transforms.Normalize
        mean:
        - 0.48145466
        - 0.4578275
        - 0.40821073
        std:
        - 0.26862954
        - 0.26130258
        - 0.27577711
    val:
      rgb_static:
      - _target_: torchvision.transforms.Resize
        size: 112
        antialias: true
      - _target_: flower.utils.transforms.ScaleImageTensor
      - _target_: torchvision.transforms.Normalize
        mean:
        - 0.48145466
        - 0.4578275
        - 0.40821073
        std:
        - 0.26862954
        - 0.26130258
        - 0.27577711
      rgb_gripper:
      - _target_: torchvision.transforms.Resize
        size: 112
        antialias: true
      - _target_: flower.utils.transforms.ScaleImageTensor
      - _target_: torchvision.transforms.Normalize
        mean:
        - 0.48145466
        - 0.4578275
        - 0.40821073
        std:
        - 0.26862954
        - 0.26130258
        - 0.27577711
  _target_: flower.datasets.libero_data_module.LiberoDataModule
  _recursive_: false
  root_data_dir: ${root_data_dir}
  action_space: 7
  shuffle_val: false
  benchmark_name: ${libero_benchmark}
  observation_space:
    rgb_obs:
    - agentview_rgb
    - eye_in_hand_rgb
    depth_obs: []
    state_obs:
    - gripper_states
    - joint_states
    actions:
    - rel_actions
    language:
    - language
  proprioception_dims: None
  datasets:
    lang_dataset:
      _target_: flower.datasets.libero_dataset.LiberoMultitaskDataset
      key: lang
      benchmark_name: ${libero_benchmark}
      batch_size: ${batch_size}
      proprio_state: ${datamodule.proprioception_dims}
      obs_space: ${datamodule.observation_space}
      num_workers: ${num_workers}
      action_seq_len: ${act_seq_len}
      obs_seq_len: ${obs_seq_len}
      split_ratio: 0.0
model:
  _target_: flower.models.flower.FLOWERVLA
  _recursive_: false
  vlm_path: microsoft/Florence-2-large
  freeze_florence: false
  freeze_vision_tower: false
  vlm_prompt_style: default
  token_dropout: 0.1
  multistep: ${multistep}
  num_sampling_steps: 4
  lowdim_obs_dim: 7
  action_dim: 7
  act_window_size: 10
  load_pretrained: true
  pretrained_model_path: /home/hk-project-sustainebot/ft4740/code/flower_vla_policy/logs/runs/2025-02-05/10-17-02/360000_model_weights.pt
  use_second_view: true
  second_view_key: image_wrist
  action_type_adaln: true
  use_causal_attention: true
  use_cross_attn: true
  use_adaln_cond: false
  use_readout_token: false
  use_proprio: false
  return_act_chunk: false
  sampling_type: uniform
  dit_dim: 1024
  n_heads: 16
  n_layers: 18
  attn_pdrop: 0.1
  resid_pdrop: 0.1
  mlp_pdrop: 0.1
  use_rope: true
  use_nope: false
  query_seq_len: 100
  rope_theta: 32.0
  optimizer_type: adamw
  optimizer:
    _target_: torch.optim.AdamW
    transformer_weight_decay: 0.05
    learning_rate: 2.0e-05
    betas:
    - 0.9
    - 0.99
  lr_scheduler:
    lr_scheduler:
      init_lr: 2.0e-05
      init_lr_scale: 0.1
      final_lr_scale: 0.5
      total_steps: 50000
      phase_ratio: (0.05, 0.1, 0.85)
      lr: 2.0e-05
root_data_dir: /home/yagmurlu/code/MoDE_Calvin/dataset/task_ABC_D
lang_folder: lang_clip_resnet50
log_dir: ./logs
slurm: false
seed: 242
device: cuda
batch_size: 8
devices: 4
goal_window_size: 1
act_dim: 7
proprio_dims: 9
obs_dim: 512
goal_dim: 512
obs_seq_len: 1
act_seq_len: 10
multistep: ${act_seq_len}
p_last_state: 0
max_epochs: 30
rollout_lh_skip_epochs: 9
num_workers: 1
benchmark_name: ${libero_benchmark}
libero_benchmark: libero_goal
trainer:
  devices: ${devices}
  precision: bf16-mixed
  max_epochs: ${max_epochs}
  sync_batchnorm: true
  accelerator: gpu
  strategy: ddp
  limit_train_batches: 1000
  limit_val_batches: 4
logger:
  _target_: pytorch_lightning.loggers.WandbLogger
  save_dir: .
  name: logger
  group: mode
  log_model: false
  project: ${libero_benchmark}
  entity: bennoq
  id: ???