File size: 3,755 Bytes
7caec3d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 | """Pull one fully-transformed sample through the real openpi data pipeline. Fails in seconds.
uv run python scripts/check_dataloader.py pi05_g1_pickplace
Run this BEFORE compute_norm_stats.py. It exercises exactly the path that norm-stat computation
and training use -- LeRobotDataset construction, delta_timestamps, repack, the policy transforms,
the delta-action transform and the model transforms -- but on a single sample with no workers, so
a wiring error surfaces immediately instead of inside a DataLoader worker after a minute of setup.
The class of bug this catches: `action_sequence_keys` defaults to ("actions",) while a LeRobot
dataset's column may be "action". That name is fed raw into delta_timestamps, so it is resolved
against the hf_dataset BEFORE any repack renames it, and the failure looks like
KeyError: Column actions not in the dataset. Current columns: [... 'action' ...]
buried in a torch worker traceback.
"""
import sys
import numpy as np
def main() -> int:
if len(sys.argv) < 2:
print(__doc__)
return 2
name = sys.argv[1]
from openpi.training import config as _config
from openpi.training import data_loader as _data_loader
cfg = _config.get_config(name)
data_config = cfg.data.create(cfg.assets_dirs, cfg.model)
print(f"config {cfg.name}")
print(f"repo_id {data_config.repo_id}")
print(f"action_sequence_keys{'':<0} {tuple(data_config.action_sequence_keys)}")
print(f"action_horizon {cfg.model.action_horizon}")
print(f"use_quantile_norm {data_config.use_quantile_norm}")
print(f"prompt_from_task {data_config.prompt_from_task}")
# Raw dataset first: this is where an action_sequence_keys mismatch blows up.
ds = _data_loader.create_torch_dataset(data_config, cfg.model.action_horizon, cfg.model)
print(f"\nraw dataset {len(ds)} samples")
raw = ds[0]
for k, v in sorted(raw.items()):
if hasattr(v, "shape"):
print(f" {k:<44} {tuple(v.shape)}")
if isinstance(raw.get("prompt"), str) or isinstance(raw.get("task"), str):
print(f" prompt/task{'':<33} {raw.get('prompt') or raw.get('task')!r}")
# Then the full transform stack. skip_norm_stats=True so this works before Step 8.
tds = _data_loader.transform_dataset(ds, data_config, skip_norm_stats=True)
item = tds[0]
print("\ntransformed sample (what the model actually receives):")
for k, v in sorted(item.items()):
if isinstance(v, dict):
for kk, vv in sorted(v.items()):
print(f" {k}/{kk:<38} {getattr(vv, 'shape', vv)}")
elif hasattr(v, "shape"):
print(f" {k:<44} {tuple(np.asarray(v).shape)} {np.asarray(v).dtype}")
else:
print(f" {k:<44} {v!r}")
problems = []
st, ac = np.asarray(item["state"]), np.asarray(item["actions"])
if st.shape != (cfg.model.action_dim,):
problems.append(f"state is {st.shape}, expected ({cfg.model.action_dim},)")
if ac.shape != (cfg.model.action_horizon, cfg.model.action_dim):
problems.append(f"actions is {ac.shape}, expected "
f"({cfg.model.action_horizon}, {cfg.model.action_dim})")
if not np.all(np.isfinite(st)) or not np.all(np.isfinite(ac)):
problems.append("non-finite values in state/actions")
if np.all(ac == 0):
problems.append("every action is exactly 0 -- the action column is probably not wired")
print()
if problems:
for p in problems:
print("FAIL", p)
return 1
print("PASS -- the pipeline produces well-formed model inputs. Safe to run compute_norm_stats.py.")
return 0
if __name__ == "__main__":
sys.exit(main())
|