justintiensmith's picture
Enable progress logging in Robometer runner
c93abf7 verified
Raw
History Blame Contribute Delete
4.43 kB
"""Run Robometer on one episode and publish only derived artifacts.
This script is intended for a bounded Hugging Face Job. The source dataset is
never opened for writing; all uploads target the duplicated dataset repository.
"""
from __future__ import annotations
import logging
import os
import subprocess
import sys
from pathlib import Path
from huggingface_hub import HfApi
import lerobot.rewards.robometer.compute_rabc_weights as robometer_compute
DATASET_REPO_ID = os.getenv(
"ROBOMETER_DATASET_REPO_ID",
"justintiensmith/Ordering_Constrained_Black_Bin_Sequencing_Robometer",
)
REWARD_MODEL_ID = os.getenv("ROBOMETER_MODEL_ID", "lerobot/Robometer-4B")
EPISODE = int(os.getenv("ROBOMETER_EPISODE", "0"))
CAMERA_KEY = os.getenv("ROBOMETER_CAMERA_KEY", "observation.images.middle")
BATCH_SIZE = int(os.getenv("ROBOMETER_BATCH_SIZE", "4"))
OUTPUT_ROOT = Path(
os.getenv(
"ROBOMETER_OUTPUT_DIR",
"/vol/dissolve/justin/outputs/robometer_black_bin_episode_000",
)
)
PROGRESS_PATH = OUTPUT_ROOT / "robometer_progress.parquet"
VIDEO_DIR = OUTPUT_ROOT / "videos"
def main() -> None:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s %(levelname)s %(message)s",
)
if EPISODE != 0:
raise ValueError(
"This optimized runner currently supports episode 0 only because the in-tree "
"Robometer helper uses global dataset indices."
)
OUTPUT_ROOT.mkdir(parents=True, exist_ok=True)
VIDEO_DIR.mkdir(parents=True, exist_ok=True)
# The in-tree helper normally constructs LeRobotDataset without an episode
# filter. Restricting it here prevents the job from downloading unrelated
# videos. Episode 0 starts at dataset index 0, so its local/global indices
# are identical after filtering.
original_dataset_class = robometer_compute.LeRobotDataset
def episode_dataset(repo_id: str, download_videos: bool = True):
return original_dataset_class(
repo_id,
episodes=[EPISODE],
download_videos=download_videos,
)
robometer_compute.LeRobotDataset = episode_dataset
try:
progress_path = robometer_compute.compute_robometer_progress(
dataset_repo_id=DATASET_REPO_ID,
reward_model_path=REWARD_MODEL_ID,
output_path=str(PROGRESS_PATH),
device="cuda",
batch_size=BATCH_SIZE,
num_subsampled_frames=4,
episodes=[EPISODE],
image_key=CAMERA_KEY,
)
finally:
robometer_compute.LeRobotDataset = original_dataset_class
api = HfApi()
api.upload_file(
path_or_fileobj=str(progress_path),
path_in_repo="robometer/episode_000/progress.parquet",
repo_id=DATASET_REPO_ID,
repo_type="dataset",
commit_message="Add Robometer progress for episode 0",
)
lerobot_repo = Path(
os.getenv(
"LEROBOT_REPO",
str(Path(robometer_compute.__file__).resolve().parents[4]),
)
)
overlay_script = lerobot_repo / "examples/dataset/create_progress_videos.py"
if not overlay_script.exists():
raise FileNotFoundError(
f"Could not find the LeRobot overlay script at {overlay_script}. "
"Set LEROBOT_REPO to the LeRobot source checkout."
)
subprocess.run(
[
sys.executable,
str(overlay_script),
"--repo-id",
DATASET_REPO_ID,
"--episode",
str(EPISODE),
"--camera-key",
CAMERA_KEY,
"--progress-file",
str(progress_path),
"--output-dir",
str(VIDEO_DIR),
],
check=True,
)
video_path = VIDEO_DIR / (
"justintiensmith_Ordering_Constrained_Black_Bin_Sequencing_Robometer_"
"ep0_progress.mp4"
)
if not video_path.exists():
raise FileNotFoundError(f"Expected rendered video was not created: {video_path}")
api.upload_file(
path_or_fileobj=str(video_path),
path_in_repo="robometer/episode_000/middle_progress.mp4",
repo_id=DATASET_REPO_ID,
repo_type="dataset",
commit_message="Add Robometer visualization for episode 0",
)
print(f"PROGRESS_ARTIFACT={progress_path}")
print(f"VIDEO_ARTIFACT={video_path}")
if __name__ == "__main__":
main()