action_dim: 7
action_tokenizer_exist: false
class_dropout_prob: 0.0
data_root_dir: /media/liuzhuoyang/data/rlbench/rlds/12tasks_selected_keyframe_pointcloud_0814
future_action_window_size: 0
hf_token: .hf_token
image_aug: false
is_resume: false
llm_vision_layers: 8
load_all_data_for_training: true
past_action_window_size: 0
pretrained_checkpoint: /media/liuzhuoyang/new_vla/Rec_Diff_beta/exp/exp_12tasks_selected_keyframe_pointcloud_0814_Pretrainrtx0818e1_FreezeVistrue_Window0_Difftrue_Recfalse_Contrastive_Vislayer8_1024_0403_0820/checkpoints/step-005415-epoch-300-loss=1.4414.pt
recon_image: true
recon_pointcloud: true
repeated_diffusion_steps: 4
resume_epoch: null
resume_step: null
run_id: exp_12tasks_selected_keyframe_pointcloud_0814_Pretraindiff300_FreezeVistrue_Window0_Difftrue_PCtrue_Rectrue_Contrastive_Vislayer8_1024_0403_0822
run_id_note: null
run_root_dir: /media/liuzhuoyang/new_vla/Rec_Diff_beta/exp
save_interval: 100
seed: 42
trackers:
- jsonl
- wandb
use_contrastive: true
use_diff: true
use_ema: false
use_pointcloud: true
use_reconstruction: true
use_roi: false
vla:
  base_vlm: prism-dinosiglip-224px+7b
  data_mix: rlbench
  enable_gradient_checkpointing: true
  enable_mixed_precision_training: true
  epochs: 300
  expected_world_size: 32
  freeze_llm_backbone: false
  freeze_vision_tower: true
  global_batch_size: 256
  learning_rate: 2.0e-05
  lr_scheduler_type: constant
  max_grad_norm: 1.0
  max_steps: null
  per_device_batch_size: 8
  reduce_in_full_precision: true
  shuffle_buffer_size: 10000
  train_strategy: fsdp-full-shard
  type: prism-dinosiglip-224px+oxe+diffusion
  unfreeze_last_llm_layer: false
  vla_id: prism-dinosiglip-224px+oxe+diffusion
  warmup_ratio: 0.0
  weight_decay: 0.0
wandb_entity: liumail2023-peking-university
wandb_project: one_model_vla_sft