code-Psi0-yth-v21 / baselines /gr00t-n1.6 /openloop_eval_http.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
211ae03 verified
Raw History Blame Contribute Delete
4.25 kB
import dataclasses
import json
import pathlib
from typing import Any
import urllib.request
import numpy as np
import pyarrow.parquet as pq
import imageio.v3 as iio
from psi.deploy.helpers import RequestMessage, ResponseMessage
def _load_tasks(tasks_path: pathlib.Path) -> dict[int, str]:
tasks = {}
if not tasks_path.exists():
return tasks
with tasks_path.open("r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
rec = json.loads(line)
tasks[int(rec.get("task_index", 0))] = rec.get("task", "")
return tasks
def _load_frame(video_path: pathlib.Path, frame_index: int) -> np.ndarray:
if not video_path.exists():
raise FileNotFoundError(f"Video not found: {video_path}")
return iio.imread(video_path, index=frame_index)
@dataclasses.dataclass
class Args:
host: str = "127.0.0.1"
port: int = 5555
dataset_path: str = "/hfm/data/simple/G1WholebodyBendPick-v0-psi0"
episode_index: int = 0
target_index: int = 0
video_key: str = "observation.rgb_head_stereo_left"
def main() -> None:
import tyro
args = tyro.cli(Args)
dataset_root = pathlib.Path(args.dataset_path).resolve()
meta_dir = dataset_root / "meta"
info = json.loads((meta_dir / "info.json").read_text())
chunks_size = int(info.get("chunks_size", 1000))
data_path_fmt = info.get("data_path", "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet")
tasks = _load_tasks(meta_dir / "tasks.jsonl")
ep_index = args.episode_index
chunk_id = ep_index // chunks_size
parquet_path = dataset_root / data_path_fmt.format(
episode_chunk=chunk_id,
episode_index=ep_index,
)
table = pq.read_table(parquet_path)
df = table.to_pandas()
if args.target_index >= len(df):
raise ValueError(f"target_index {args.target_index} out of range (len={len(df)})")
row = df.iloc[args.target_index]
instruction = ""
if "task" in row:
instruction = str(row["task"])
elif "task_index" in row:
instruction = tasks.get(int(row["task_index"]), "")
if "observation.proprio_joint_positions" not in row or "observation.amo_policy_command" not in row:
raise KeyError(
"Dataset is missing observation.proprio_joint_positions / observation.amo_policy_command. "
"Use the original simple dataset or add these fields in conversion."
)
proprio = np.asarray(row["observation.proprio_joint_positions"], dtype=np.float32)
command = np.asarray(row["observation.amo_policy_command"], dtype=np.float32)
video_path = dataset_root / "videos" / f"chunk-{chunk_id:03d}" / args.video_key / f"episode_{ep_index:06d}.mp4"
frame = _load_frame(video_path, args.target_index)
request = RequestMessage(
image={args.video_key: frame},
instruction=instruction,
history={},
state={
"proprio_joint_positions": proprio,
"amo_policy_command": command,
},
condition={},
gt_action=[],
dataset_name="simple",
timestamp=str(row.get("timestamp", "")),
)
payload = json.dumps(request.serialize()).encode("utf-8")
url = f"http://{args.host}:{args.port}/act"
req = urllib.request.Request(url, data=payload, headers={"Content-Type": "application/json"})
with urllib.request.urlopen(req) as resp:
resp_payload = json.loads(resp.read().decode("utf-8"))
if "status" in resp_payload:
raise RuntimeError(f"Server error: {resp_payload['status']}")
response = ResponseMessage.deserialize(resp_payload)
action = response.action
print(f"got action shape={action.shape} err={response.err}")
gt_action = np.asarray(row["action"], dtype=np.float32)
print(f"gt action shape={gt_action.shape}")
if gt_action.size > 0:
print(
f"gt action: {np.array2string(gt_action.reshape(-1)[:10], precision=4, suppress_small=True)}"
)
if action.size > 0:
print(
f"pred action: {np.array2string(action.reshape(-1)[:10], precision=4, suppress_small=True)}"
)
if __name__ == "__main__":
main()