diff --git a/src/b1k/training/config.py b/src/b1k/training/config.py index 52dfb7d..b5d5bca 100644 --- a/src/b1k/training/config.py +++ b/src/b1k/training/config.py @@ -92,6 +92,9 @@ class DataConfig: # Episodes index to use for training episodes_index: List[int] | None = None + # B1K task names to train on. If None, all 50 challenge tasks are used. + tasks: List[str] | None = None + class GroupFactory(Protocol): def __call__(self, model_config: _model.BaseModelConfig) -> _transforms.Group: @@ -376,6 +379,106 @@ _CONFIGS = [ save_interval=500, keep_period=2000, ), + TrainConfig( + # Fine-tune the 1st-place checkpoint_3 (tasks 4,27,31-33,35-39,41,46,49) on task 37 only. + name="pi_behavior_b1k_task37_clean_a_trumpet", + exp_name="ft_from_ckpt3_lr5e-6", + project_name="B1K", + model=pi_behavior_config.PiBehaviorConfig( + action_horizon=30, + action_dim=32, + use_correlated_noise=True, + correlation_beta=0.5, + use_fast_auxiliary=True, + fast_loss_weight=0.05, + fast_encoded_dims="0:6,7:23", + fast_vocab_size=1024, + max_fast_tokens=200, + use_kv_transform=True, + use_knowledge_insulation=False, + subtask_loss_weight=0.1, + freeze_vision_backbone=True, + ), + data=LeRobotB1KDataConfig( + repo_id="IliaLarchenko/behavior_224_rgb", + base_config=DataConfig( + prompt_from_task=False, + behavior_dataset_root="~/data/behavior_224_rgb", + use_per_timestamp_norm=True, + tasks=["clean_a_trumpet"], + ), + use_delta_joint_actions=True, + use_fast_tokenization=True, + ), + # Low LR since we start from an already task-adapted checkpoint; peak 2.5e-5 drove action_loss + # from ~0.10 up to ~0.19 during warmup and it never recovered. + lr_schedule=_optimizer.CosineDecaySchedule( + warmup_steps=200, + peak_lr=5e-6, + decay_steps=10_000, + decay_lr=5e-7, + ), + num_flow_samples=15, + weight_loader=weight_loaders.PiBehaviorWeightLoader("/root/models/checkpoint_3/params"), + num_train_steps=10_000, + batch_size=16, # 32 OOMs on one A100-80GB (needs ~81GB) + assets_base_dir="./outputs/assets", + checkpoint_base_dir="./outputs/checkpoints", + num_workers=32, + save_interval=1000, + keep_period=None, + wandb_enabled=False, + ), + TrainConfig( + # Fine-tune the 1st-place checkpoint_3 (tasks 4,27,31-33,35-39,41,46,49) on task 38 (spraying_for_bugs) only, 25k steps. + name="pi_behavior_b1k_task38_spraying_for_bugs", + exp_name="ft_from_ckpt3_lr5e-6_25k", + project_name="B1K", + model=pi_behavior_config.PiBehaviorConfig( + action_horizon=30, + action_dim=32, + use_correlated_noise=True, + correlation_beta=0.5, + use_fast_auxiliary=True, + fast_loss_weight=0.05, + fast_encoded_dims="0:6,7:23", + fast_vocab_size=1024, + max_fast_tokens=200, + use_kv_transform=True, + use_knowledge_insulation=False, + subtask_loss_weight=0.1, + freeze_vision_backbone=True, + ), + data=LeRobotB1KDataConfig( + repo_id="IliaLarchenko/behavior_224_rgb", + base_config=DataConfig( + prompt_from_task=False, + behavior_dataset_root="~/data/behavior_224_rgb", + use_per_timestamp_norm=True, + tasks=["spraying_for_bugs"], + ), + use_delta_joint_actions=True, + use_fast_tokenization=True, + ), + # Low LR since we start from an already task-adapted checkpoint; peak 2.5e-5 drove action_loss + # from ~0.10 up to ~0.19 during warmup and it never recovered. + lr_schedule=_optimizer.CosineDecaySchedule( + warmup_steps=200, + peak_lr=5e-6, + decay_steps=25_000, + decay_lr=5e-7, + ), + num_flow_samples=15, + weight_loader=weight_loaders.PiBehaviorWeightLoader("/root/models/checkpoint_3/params"), + num_train_steps=25_000, + batch_size=16, # 32 OOMs on one A100-80GB (needs ~81GB) + assets_base_dir="./outputs/assets", + checkpoint_base_dir="./outputs/checkpoints", + num_workers=12, + save_interval=2500, + keep_period=2500, + wandb_enabled=False, + ), ] if len({config.name for config in _CONFIGS}) != len(_CONFIGS): diff --git a/src/b1k/training/data_loader.py b/src/b1k/training/data_loader.py index 3dd8fb2..9e7738e 100644 --- a/src/b1k/training/data_loader.py +++ b/src/b1k/training/data_loader.py @@ -65,7 +65,7 @@ def create_behavior_dataset(data_config: _config.DataConfig, action_horizon: int if seed is None: seed = int(time.time() * 1000) % (2**32) logging.info(f"Using random seed for BehaviorLeRobotDataset: {seed}") - tasks = [ + tasks = data_config.tasks or [ "picking_up_trash", # difficulty: 2 "putting_away_Halloween_decorations", # difficulty: 3 "cleaning_up_plates_and_food", # difficulty: 3.5 diff --git a/src/b1k/training/weight_loaders.py b/src/b1k/training/weight_loaders.py index ed3a614..8a642bd 100644 --- a/src/b1k/training/weight_loaders.py +++ b/src/b1k/training/weight_loaders.py @@ -11,6 +11,7 @@ import flax.traverse_util import numpy as np import orbax.checkpoint as ocp +import openpi.models.model as _model import openpi.shared.array_typing as at import openpi.shared.download as download @@ -40,9 +41,8 @@ class PiBehaviorWeightLoader(WeightLoader): # Load checkpoint params_path = download.maybe_download(self.params_path) - # Load directly with PyTreeCheckpointer (handles both old and new checkpoint formats) - with ocp.PyTreeCheckpointer() as ckptr: - restored = ckptr.restore(params_path) + # Newer orbax requires explicit restore shardings; restore_params supplies them. + restored = _model.restore_params(params_path, restore_type=np.ndarray) # Handle nested 'params' key (from some checkpoint formats) if isinstance(restored, dict) and "params" in restored: