File size: 4,108 Bytes
84af0a7 | 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 89 90 91 92 93 94 95 96 97 98 99 100 101 | """Load one saved checkpoint and run a single inference. The real test of a checkpoint.
Run from inside the openpi checkout:
# while training still owns the A100 -- CPU, slow (~1-2 min), proves loadability
JAX_PLATFORMS=cpu uv run python <this file> /data/checkpoints/.../pickplace_full_v1/7500
# after training exits -- GPU, fast
uv run python <this file> /data/checkpoints/.../pickplace_full_v1/29999
`check_checkpoints.py` verifies the files are structurally complete. This verifies they actually
deserialise into the model, that norm stats resolve, and that the policy emits a finite (50, 28)
action chunk for a plausible observation. A checkpoint can pass the structural check and still be
unusable -- wrong config, mismatched action_dim, missing norm stats for this exact config name.
Nothing here touches the robot. The action values are meaningless (the images are synthetic); only
their shape, dtype and finiteness are being checked.
"""
import argparse
import pathlib
import sys
import time
TASK = "pick octopus and place inside brown basket"
def main() -> int:
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("checkpoint", type=pathlib.Path, help="a step directory containing params/ and assets/")
ap.add_argument("--config", default="pi05_g1_pickplace")
ap.add_argument("--prompt", default=TASK)
ap.add_argument("--state-dim", type=int, default=43, choices=(28, 43))
ap.add_argument("--repeat", type=int, default=2, help="inferences to run; the 2nd shows warm latency")
a = ap.parse_args()
if not (a.checkpoint / "params").is_dir():
print(f"FAIL: {a.checkpoint}/params does not exist", file=sys.stderr)
return 2
import numpy as np
import jax
from openpi.policies import policy_config
from openpi.training import config as _config
print(f"jax backend : {jax.default_backend()} devices={jax.devices()}")
print(f"config : {a.config}")
print(f"checkpoint : {a.checkpoint}")
cfg = _config.get_config(a.config)
t0 = time.time()
policy = policy_config.create_trained_policy(cfg, a.checkpoint)
print(f"loaded : {time.time() - t0:.1f}s")
# Dataset-native keys -- create_trained_policy applies the repack transform, so this is the
# same dict the robot client sends (see runbook Step 12).
rng = np.random.default_rng(0)
obs = {
"observation.images.ego_view": rng.integers(0, 256, (480, 640, 3), dtype=np.uint8),
"observation.images.ego_left": rng.integers(0, 256, (480, 640, 3), dtype=np.uint8),
"observation.images.ego_right": rng.integers(0, 256, (480, 640, 3), dtype=np.uint8),
"observation.state": np.zeros(a.state_dim, dtype=np.float32),
"prompt": a.prompt,
}
for i in range(a.repeat):
t0 = time.time()
out = policy.infer(obs)
dt = time.time() - t0
act = np.asarray(out["actions"])
print(f"infer #{i + 1} : {dt:.2f}s shape={act.shape} dtype={act.dtype}")
problems = []
if act.shape != (50, 28):
problems.append(f"expected action chunk (50, 28), got {act.shape}")
if not np.all(np.isfinite(act)):
n = int((~np.isfinite(act)).sum())
problems.append(f"{n} non-finite value(s) in the action chunk")
if np.allclose(act, 0.0):
problems.append("every action is exactly 0.0 -- norm stats or params likely not loaded")
print(f"\nfirst step : {np.array2string(act[0], precision=3, max_line_width=200)}")
print(f"per-dim |max|: {np.abs(act).max(axis=0).round(3)}")
print(f"range : [{act.min():.3f}, {act.max():.3f}]")
if problems:
print("\nFAIL")
for p in problems:
print(f" {p}")
return 1
print("\nPASS -- checkpoint loads and produces a finite (50, 28) chunk.")
print("Action VALUES are not validated here; the images were noise. Only §14 rollouts tell you")
print("whether this checkpoint is any good.")
return 0
if __name__ == "__main__":
sys.exit(main())
|