| local_run_dir: ckpts |
| exp_name: hapo_stack_three_d0 |
| cache_dir: .cache |
| seed: 1 |
| warmup_steps: 10 |
| cache_reference_logprobs: false |
| lr: 5.0e-05 |
| correct_ratio: 0.5 |
| interaction_ratio: 0.25 |
| wandb: |
| enabled: true |
| entity: null |
| project: hapo |
| minimum_log_interval_secs: 1.0 |
| use_lora: true |
| batch_size: 8 |
| eval_batch_size: 8 |
| n_epochs: 10 |
| grad_accumulation_steps: 1 |
| save_steps: 500 |
| action_space: 7 |
| do_first_eval: false |
| eval_every: 2000 |
| debug: false |
| intermediate_checkpoints: true |
| model: |
| name_or_path: /mnt/bn/xwk-robotlearning/workspace/openvla/ckpts/base_policy/stack_three_d0_300/adapter_dir |
| base_path: /mnt/bn/xwk-robotlearning/.cache/huggingface/hub/models--openvla--openvla-7b/snapshots/31f090d05236101ebfc381b61c674dd4746d4ce0 |
| tokenizer_name_or_path: null |
| load_from: null |
| from_checkpoint: null |
| block_name: LlamaDecoderLayer |
| policy_dtype: bfloat16 |
| reference_dtype: bfloat16 |
| max_grad_norm: 5.0 |
| v_head_max_grad_norm: 0.1 |
| max_length: 1024 |
| max_prompt_length: 512 |
| activation_checkpointing: true |
| batch_size: 8 |
| gradient_accumulation_steps: 1 |
| eval_batch_size: 8 |
| attn_implementation: flash_attention_2 |
| use_peft: true |
| load_lora_from: null |
| peft: |
| lora_r: 64 |
| lora_alpha: 256 |
| lora_dropout: 0.05 |
| target_modules: all-linear |
| reward_model: |
| path: null |
| model_class: AutoModelForBradleyTerry |
| dtype: float32 |
| attn_implementation: flash_attention_2 |
| loss: |
| beta: 0.1 |
| trainer: KTOTrainer |
| dataloader: UnpairedPreferenceDataLoader |
| desirable_weight: 1 |
| undesirable_weight: 1.0 |
| bc_weight: 1 |
| kl_max: 5 |
| kl_min: -5 |
| token_distance: 1 |
| trainer: |
| _target_: trainer.ddp_robotic_trainer.HAPOTrainer |
| dataset: |
| _target_: dataset.balance_hapo_dataset.BalanceHAPODataset |
| kwargs: |
| data_root_dir: robotic_data/ |
| interaction_root_dir: robotic_data/mimicgen_interaction_data/mimicgen_stack_three_d0_interaction |
| data_name: mimicgen |
| data_list_names: stack_three_d0.json |
| interaction_list_names: interaction_data_list.json |
| demo_number: 50 |
| image_shape: 224 |
| previous_K: 10 |
| eval_dataset: |
| _target_: dataset.eval_mimicgen.EvalMimicgen |
| kwargs: |
| data_root_dir: robotic_data/ |
| interaction_root_dir: robotic_data/mimicgen_interaction_data/mimicgen_stack_three_d0_interaction |
| data_name: mimicgen |
| data_list_names: stack_three_d0.json |
| interaction_list_names: interaction_data_list.json |
| demo_number: 0 |
| image_shape: 224 |
|
|
