TTI / Dev /testing /prepare_test_demo_single_task.py
JosephBai's picture
Upload folder using huggingface_hub
857c2e9 verified
"""
Prepare test demos for LIBERO-10 task suite.
The demonstration that has the shortest frames sequence is regarded as the expert demonstration.
When preparing the test demo, we exclude the expert demonstration.
For the rest of the demonstrations, we randomly sample up to 40 demonstrations (configurable).
For each demonstration, we sample 4 frames with equal interval (configurable), including the initial frame and the success frame.
We save the frames as PNGs in a folder.
Usage:
# Process all LIBERO-10 tasks
python prepare_test_demo_single_task.py --process-all-tasks --libero-data-dir <data_dir> --output-root <output_root>
# Process a single task
python prepare_test_demo_single_task.py --task_name <task_name> --hdf5_path <path_to_hdf5> --output_dir <output_dir>
Examples:
# Process all LIBERO-10 tasks
python prepare_test_demo_single_task.py \
--process-all-tasks \
--libero-data-dir /home/zechen/Data/Robo/LIBERO_Regen/libero_10_regen \
--output-root toy_test_demos_LIBERO_10 \
--num-demos 40 \
--frames-per-demo 4
# Process a single task
python prepare_test_demo_single_task.py \
--task-name KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it \
--hdf5-path /path/to/KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it.hdf5 \
--output-dir /path/to/output \
--num-demos 40 \
--frames-per-demo 4
"""
import argparse
import json
import random
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import h5py
import numpy as np
from PIL import Image
from tqdm import tqdm
# LIBERO-10 task list
LIBERO_10_TASKS = [
"KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it",
"KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it",
"KITCHEN_SCENE6_put_the_yellow_and_white_mug_in_the_microwave_and_close_it",
"KITCHEN_SCENE8_put_both_moka_pots_on_the_stove",
"LIVING_ROOM_SCENE1_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket",
"LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket",
"LIVING_ROOM_SCENE2_put_both_the_cream_cheese_box_and_the_butter_in_the_basket",
"LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate",
"LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate",
"STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy",
]
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Prepare test demos for LIBERO-10 task suite")
# Mode selection
parser.add_argument("--process-all-tasks", action="store_true", help="Process all LIBERO-10 tasks")
# Arguments for processing all tasks
parser.add_argument("--libero-data-dir", type=Path, help="Directory containing all LIBERO-10 HDF5 files (required with --process-all-tasks)")
parser.add_argument("--output-root", type=Path, help="Root output directory (required with --process-all-tasks)")
# Arguments for processing a single task
parser.add_argument("--task-name", help="Name of the task (for single task mode)")
parser.add_argument("--hdf5-path", type=Path, help="Path to the HDF5 file (for single task mode)")
parser.add_argument("--output-dir", type=Path, help="Output directory (for single task mode)")
# Common arguments
parser.add_argument("--num-demos", type=int, default=40, help="Number of test demos to sample (default: 40)")
parser.add_argument("--frames-per-demo", type=int, default=4, help="Number of frames per demo (default: 4)")
parser.add_argument("--random-seed", type=int, default=42, help="Random seed for reproducibility (default: 42)")
args = parser.parse_args()
# Validate arguments
if args.process_all_tasks:
if not args.libero_data_dir:
parser.error("--libero-data-dir is required when using --process-all-tasks")
if not args.output_root:
parser.error("--output-root is required when using --process-all-tasks")
else:
if not args.task_name:
parser.error("--task-name is required for single task mode")
if not args.hdf5_path:
parser.error("--hdf5-path is required for single task mode")
if not args.output_dir:
parser.error("--output-dir is required for single task mode")
return args
def load_demo_arrays(demo_group: h5py.Group):
"""Load frames and dones arrays from a demo group."""
frames = demo_group["obs/agentview_rgb"]
dones_ds = demo_group.get("dones")
dones = np.asarray(dones_ds[:]) if dones_ds is not None else None
return frames, dones
def first_success_index(dones: Optional[np.ndarray], total_frames: int) -> int:
"""Find the index of the first success frame."""
if dones is None:
return total_frames - 1
indices = np.where(dones == 1)[0]
return int(indices[0]) if indices.size > 0 else total_frames - 1
def find_expert_demo(data_group: h5py.Group) -> Optional[Tuple[str, int, int]]:
"""Find the expert demo (shortest successful trajectory).
Returns:
Tuple of (demo_name, total_frames, success_index) or None if no expert found
"""
best_name: Optional[str] = None
best_total: Optional[int] = None
best_success: Optional[int] = None
for demo_name in sorted(data_group.keys()):
demo_group = data_group[demo_name]
frames, dones = load_demo_arrays(demo_group)
total_frames = frames.shape[0]
success_idx = first_success_index(dones, total_frames)
if success_idx <= 0:
continue
if best_name is None or total_frames < best_total:
best_name = demo_name
best_total = total_frames
best_success = success_idx
if best_name is None:
return None
return best_name, best_total, best_success
def save_frame(array: np.ndarray, file_path: Path) -> None:
"""Save a frame array as PNG with vertical flip."""
file_path.parent.mkdir(parents=True, exist_ok=True)
Image.fromarray(array).transpose(Image.FLIP_TOP_BOTTOM).save(file_path)
def process_single_task(
task_name: str,
hdf5_path: Path,
output_dir: Path,
num_demos: int,
frames_per_demo: int,
random_seed: int,
) -> bool:
"""Process a single task and generate test demos.
Args:
task_name: Name of the task
hdf5_path: Path to the HDF5 file
output_dir: Output directory for this task
num_demos: Number of demos to sample
frames_per_demo: Number of frames per demo
random_seed: Random seed for reproducibility
Returns:
True if successful, False otherwise
"""
output_dir.mkdir(parents=True, exist_ok=True)
if not hdf5_path.exists():
print(f"Error: HDF5 file not found: {hdf5_path}")
return False
print(f"\n{'='*80}")
print(f"Processing task: {task_name}")
print(f"HDF5 file: {hdf5_path}")
print(f"Output directory: {output_dir}")
print(f"{'='*80}")
with h5py.File(hdf5_path, "r") as handle:
data_group = handle.get("data")
if data_group is None:
print("Error: No data group found in HDF5 file")
return False
# Step 1: Find the expert demo (shortest successful trajectory)
print("\nFinding expert demonstration...")
expert_info = find_expert_demo(data_group)
if expert_info is None:
print("Error: No successful demonstrations found")
return False
expert_name, expert_frames, expert_success = expert_info
print(f"Expert demo: {expert_name} ({expert_frames} frames, success at {expert_success})")
# Step 2: Collect all non-expert successful demos
demo_names = sorted(data_group.keys())
candidate_demos: List[Tuple[str, int, int]] = []
for demo_name in demo_names:
if demo_name == expert_name:
continue
demo_group = data_group[demo_name]
frames, dones = load_demo_arrays(demo_group)
total_frames = frames.shape[0]
success_idx = first_success_index(dones, total_frames)
if success_idx <= 0:
continue
candidate_demos.append((demo_name, total_frames, success_idx))
print(f"\nFound {len(candidate_demos)} candidate test demos (excluding expert)")
if len(candidate_demos) == 0:
print("Error: No test demos available after excluding expert")
return False
# Step 3: Randomly sample up to num_demos
random.seed(random_seed) # For reproducibility
sampled_demos = random.sample(candidate_demos, min(num_demos, len(candidate_demos)))
print(f"Sampling {len(sampled_demos)} test demos")
# Step 4: Export frames for each sampled demo
manifest: List[Dict] = []
for demo_name, total_frames, success_idx in tqdm(sampled_demos, desc="Exporting demos"):
demo_group = data_group[demo_name]
frames, _ = load_demo_arrays(demo_group)
# Sample frames_per_demo frames with equal intervals
# Including initial frame (0) and success frame (success_idx)
if frames_per_demo < 2:
print(f"Warning: frames_per_demo must be at least 2, setting to 2")
frames_per_demo = 2
indices = np.linspace(0, success_idx, num=frames_per_demo, dtype=int)
# Ensure we have the initial and success frames
indices = sorted(set(indices))
if 0 not in indices:
indices.insert(0, 0)
if success_idx not in indices:
indices.append(success_idx)
# Create demo folder
demo_folder = output_dir / f"{task_name}_{demo_name}"
demo_folder.mkdir(parents=True, exist_ok=True)
# Save frames
frame_paths: List[str] = []
for i, frame_idx in enumerate(indices):
frame_path = demo_folder / f"frame_{i:04d}.png"
save_frame(frames[frame_idx], frame_path)
frame_paths.append(str(frame_path.relative_to(output_dir)))
# Save demo metadata
demo_metadata = {
"task_name": task_name,
"demo_name": demo_name,
"total_frames": int(total_frames),
"success_index": int(success_idx),
"sampled_frame_indices": [int(idx) for idx in indices],
"frame_paths": frame_paths,
}
manifest.append(demo_metadata)
# Save individual demo metadata
metadata_path = demo_folder / "metadata.json"
with metadata_path.open("w", encoding="utf-8") as f:
json.dump(demo_metadata, f, indent=2)
# Step 5: Save manifest
manifest_path = output_dir / f"{task_name}_test_manifest.json"
manifest_data = {
"task_name": task_name,
"expert_demo": expert_name,
"num_test_demos": len(sampled_demos),
"frames_per_demo": frames_per_demo,
"demos": manifest,
}
with manifest_path.open("w", encoding="utf-8") as f:
json.dump(manifest_data, f, indent=2)
print(f"\nSuccessfully exported {len(sampled_demos)} test demos")
print(f"Output directory: {output_dir}")
print(f"Manifest: {manifest_path}")
return True
def find_hdf5_file(data_dir: Path, task_name: str) -> Optional[Path]:
"""Find the HDF5 file for a given task name.
Tries different naming patterns commonly used in LIBERO datasets.
"""
# Try different naming patterns
patterns = [
f"{task_name}.hdf5",
f"{task_name}_demo.hdf5",
f"{task_name}_demos.hdf5",
]
for pattern in patterns:
candidate = data_dir / pattern
if candidate.exists():
return candidate
return None
def main() -> None:
args = parse_args()
if args.process_all_tasks:
# Process all LIBERO-10 tasks
libero_data_dir = args.libero_data_dir.expanduser()
output_root = args.output_root.expanduser()
if not libero_data_dir.exists():
print(f"Error: LIBERO data directory not found: {libero_data_dir}")
return
output_root.mkdir(parents=True, exist_ok=True)
print("="*80)
print("PROCESSING ALL LIBERO-10 TASKS")
print("="*80)
print(f"Data directory: {libero_data_dir}")
print(f"Output root: {output_root}")
print(f"Number of demos per task: {args.num_demos}")
print(f"Frames per demo: {args.frames_per_demo}")
print(f"Random seed: {args.random_seed}")
print(f"Total tasks to process: {len(LIBERO_10_TASKS)}")
print("="*80)
successful_tasks = []
failed_tasks = []
for idx, task_name in enumerate(LIBERO_10_TASKS, 1):
print(f"\n[{idx}/{len(LIBERO_10_TASKS)}] Processing: {task_name}")
# Find HDF5 file
hdf5_path = find_hdf5_file(libero_data_dir, task_name)
if hdf5_path is None:
print(f" [ERROR] HDF5 file not found for task: {task_name}")
failed_tasks.append(task_name)
continue
# Create task output directory
task_output_dir = output_root / task_name
# Process the task
success = process_single_task(
task_name=task_name,
hdf5_path=hdf5_path,
output_dir=task_output_dir,
num_demos=args.num_demos,
frames_per_demo=args.frames_per_demo,
random_seed=args.random_seed,
)
if success:
successful_tasks.append(task_name)
else:
failed_tasks.append(task_name)
# Print summary
print("\n" + "="*80)
print("PROCESSING COMPLETE")
print("="*80)
print(f"Successfully processed: {len(successful_tasks)}/{len(LIBERO_10_TASKS)} tasks")
print(f"Failed: {len(failed_tasks)}/{len(LIBERO_10_TASKS)} tasks")
if failed_tasks:
print("\nFailed tasks:")
for task in failed_tasks:
print(f" - {task}")
print(f"\nOutput directory: {output_root}")
print("="*80)
else:
# Process a single task
success = process_single_task(
task_name=args.task_name,
hdf5_path=args.hdf5_path.expanduser(),
output_dir=args.output_dir.expanduser(),
num_demos=args.num_demos,
frames_per_demo=args.frames_per_demo,
random_seed=args.random_seed,
)
if not success:
print("\nProcessing failed!")
exit(1)
if __name__ == "__main__":
main()