qshou-coder's picture
Add final 10k-step pi0.5 stack-cups JAX checkpoint and inference files
9e627f0 verified
Raw History Blame Contribute Delete
4.9 kB
"""Explicit pi0.5 JAX LoRA recipe and raw-joint-anchor deployment adapter."""
import dataclasses
import os
from pathlib import Path
import numpy as np
from openpi.models.pi0_config import Pi0Config
from openpi.training import config, optimizer, weight_loaders
from openpi import transforms
ROOT = Path(__file__).resolve().parent.parent
BASE = Path(os.environ.get('OPENPI_PI05_BASE', '/path/to/pi05_base'))
REPO_ID = 'galaxea/stack_cups_all81_20hz'
MASK = np.asarray([True] * 6 + [False] + [True] * 6 + [False])
@dataclasses.dataclass(frozen=True)
class GalaxeaData(config.LeRobotAlohaDataConfig):
raw_anchor_in_adapter: bool = False
def create(self, assets_dirs, model_config):
data = super().create(assets_dirs, model_config)
if self.raw_anchor_in_adapter:
outputs = [x for x in data.data_transforms.outputs if not isinstance(x, transforms.AbsoluteActions)]
data = dataclasses.replace(data, data_transforms=transforms.Group(
inputs=data.data_transforms.inputs, outputs=outputs))
return data
def make_config(*, smoke=False, raw_anchor_in_adapter=False):
model = Pi0Config(pi05=True, action_dim=32, action_horizon=50, max_token_len=200,
paligemma_variant='gemma_2b_lora', action_expert_variant='gemma_300m_lora')
data = GalaxeaData(
repo_id=REPO_ID, adapt_to_pi=False, default_prompt='Stack the cups.',
raw_anchor_in_adapter=raw_anchor_in_adapter,
assets=config.AssetsConfig(asset_id='galaxea_r1_lite_stack_cups'),
base_config=config.DataConfig(prompt_from_task=True),
repack_transforms=transforms.Group(inputs=[transforms.RepackTransform({
'images': {'cam_high': 'observation.images.head_left',
'cam_left_wrist': 'observation.images.wrist_left',
'cam_right_wrist': 'observation.images.wrist_right'},
'state': 'observation.state', 'actions': 'action', 'prompt': 'prompt',
})]),
)
return config.TrainConfig(
name='pi05_stack_cups_lora', exp_name='smoke' if smoke else 'stack-cups-all81-jax-lora-h50-10k',
project_name='galaxea-real-baselines', model=model, data=data,
weight_loader=weight_loaders.CheckpointWeightLoader(str(BASE / 'params')),
freeze_filter=model.get_freeze_filter(), ema_decay=None,
batch_size=8 if smoke else 256, num_workers=0 if smoke else 8,
num_train_steps=3 if smoke else 10000, seed=0,
lr_schedule=optimizer.CosineDecaySchedule(warmup_steps=500, peak_lr=5e-5,
decay_steps=10000, decay_lr=5e-6),
optimizer=optimizer.AdamW(b1=.9, b2=.95, weight_decay=1e-10, clip_gradient_norm=1.0),
assets_base_dir=str(ROOT / 'assets'), checkpoint_base_dir=str(ROOT / 'checkpoints'),
save_interval=2000, keep_period=1, log_interval=1 if smoke else 20,
fsdp_devices=1, overwrite=False, resume=False, wandb_enabled=not smoke,
policy_metadata={'robot': 'galaxea_r1_lite', 'task': 'stack_cups', 'fps': 20,
'physical_action_dim': 14, 'prediction_horizon': 50,
'joint_order': 'left_arm6,left_gripper,right_arm6,right_gripper',
'arm_units': 'radians', 'gripper_units': 'recorded high-level position command',
'output_contract': 'absolute_joint_targets; raw joint feedback anchor added once'},
)
class RawJointAnchorPolicy:
"""Wrap only a delta-output policy; never wrap already absolute output."""
def __init__(self, delta_policy):
self.delta_policy = delta_policy
@property
def metadata(self):
return self.delta_policy.metadata
def reset(self):
reset=getattr(self.delta_policy,'reset',None)
if reset is not None:
return reset()
def infer(self, observation):
raw_state = np.asarray(observation['state'], dtype=np.float32).copy()
if raw_state.shape != (14,) or not np.isfinite(raw_state).all():
raise ValueError('Expected finite physical 14-D state in left6,grip,right6,grip order')
result = self.delta_policy.infer(observation)
actions = np.asarray(result['actions'], dtype=np.float32).copy()
if actions.shape != (50, 14) or not np.isfinite(actions).all():
raise ValueError(f'Expected finite 50x14 delta/absolute-gripper predictions: {actions.shape}')
actions[:, MASK] += raw_state[None, MASK]
return {**result, 'actions': actions}
def create_galaxea_policy(checkpoint_dir, *, num_steps=10):
from openpi.policies.policy_config import create_trained_policy
cfg = make_config(raw_anchor_in_adapter=True)
delta_policy = create_trained_policy(cfg, checkpoint_dir, sample_kwargs={'num_steps': num_steps})
return RawJointAnchorPolicy(delta_policy)