File size: 8,159 Bytes
6a08a28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import os
import numpy as np
import shutil
import argparse
import cv2
import h5py
import dataclasses
from pathlib import Path
from typing import Any, Literal
import random
from tqdm import tqdm

from XPolicyLab.utils.load_file import load_yaml, load_json
from XPolicyLab.utils.process_data import decode_image_bit

from lerobot.common.datasets.lerobot_dataset import HF_LEROBOT_HOME
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ROOT_PATH = Path(__file__).parent.parent.parent.parent.parent

CAMERA_ALIASES = {
    "cam_head": "cam_high",
    "cam_left_wrist": "cam_left_wrist",
    "cam_right_wrist": "cam_right_wrist",
}

@dataclasses.dataclass(frozen=True)
class DatasetConfig:
    use_videos: bool = False
    tolerance_s: float = 0.0001
    image_writer_processes: int = 0
    image_writer_threads: int = 1
    video_backend: str | None = None


DEFAULT_DATASET_CONFIG = DatasetConfig()

def create_empty_dataset(
    repo_id: str,
    robot_type: str,
    fps: int,
    mode: Literal["video", "image"] = "image",
    *,
    dataset_config: DatasetConfig = DEFAULT_DATASET_CONFIG,
    robot_action_dim_info: dict = None,
) -> LeRobotDataset:
    
    MOTORS = [
        *[f"left_{i}" for i in range(robot_action_dim_info["arm_dim"][0])],
        *[f"left_ee_{i}" for i in range(robot_action_dim_info["ee_dim"][0])],
        *[f"right_{i}" for i in range(robot_action_dim_info["arm_dim"][1])],
        *[f"right_ee_{i}" for i in range(robot_action_dim_info["ee_dim"][1])]
    ]

    features = {
        "observation.state": {
            "dtype": "float32",
            "shape": (len(MOTORS),),
            "names": MOTORS,
        },
        "action": {
            "dtype": "float32",
            "shape": (len(MOTORS),),
            "names": [MOTORS],
        },
    }

    for camera_name in CAMERA_ALIASES.values():
        features[f"observation.images.{camera_name}"] = {
            "dtype": mode,
            "shape": (3, 480, 640),
            "names": ["height", "width", "channels"],
        }

    output_path = HF_LEROBOT_HOME / repo_id
    if output_path.exists():
        shutil.rmtree(output_path)

    return LeRobotDataset.create(
        repo_id=repo_id,
        fps=fps,
        robot_type=robot_type,
        features=features,
        use_videos=dataset_config.use_videos,
        tolerance_s=dataset_config.tolerance_s,
        image_writer_processes=dataset_config.image_writer_processes,
        image_writer_threads=dataset_config.image_writer_threads,
        video_backend=dataset_config.video_backend,
    )

def _load_compressed_images(group: h5py.Group, key: str) -> np.ndarray:
    return np.asarray(decode_image_bit(group[key]))

def _make_action_from_state(state: np.ndarray) -> np.ndarray:
    action = np.empty_like(state, dtype=np.float32)
    if len(state) == 1:
        action[0] = state[0]
        return action

    action[:-1] = state[1:]
    action[-1] = state[-1]
    return action

def load_data(ep_path) -> dict[str, Any]:
    with h5py.File(ep_path, "r") as ep:
        right_state = np.concatenate(
            [ep["state/right_arm_joint_states"][:], ep["state/right_ee_joint_states"][:][:, None]],
            axis=1,
        )
        left_state = np.concatenate(
            [ep["state/left_arm_joint_states"][:], ep["state/left_ee_joint_states"][:][:, None]],
            axis=1,
        )
        state = np.concatenate([left_state, right_state], axis=1).astype(np.float32)
        action = _make_action_from_state(state)

        images = {}
        for source_name, output_name in CAMERA_ALIASES.items():
            if source_name in ep["vision"]:
                images[output_name] = _load_compressed_images(ep["vision"][source_name], "colors")
        try:
            instructions = ep["instructions"]
        except KeyError:
            instructions = None

    return {
        "images": images,
        "state": state,
        "action": action,
        "velocity": None,
        "effort": None,
        "timestamps": None,
        "instructions": instructions,
    }

def main():
    parser = argparse.ArgumentParser(description="Process some episodes.")
    parser.add_argument("bench_name", type=str, help="Dataset bench name (e.g., RoboDojo)")
    parser.add_argument("ckpt_name", type=str, help="Run name; also selects raw task dir under data/<bench>/")
    parser.add_argument("env_cfg_type", type=str, help="Environment config type (e.g., arx_x5)")
    parser.add_argument("action_type", type=str, help="Action type for artifact naming (e.g., joint)")
    parser.add_argument(
        "expert_data_num",
        type=str,
        nargs="?",
        default=None,
        help="Optional number of episodes to process; non-numeric values are treated as raw_task_dirs.",
    )
    parser.add_argument(
        "raw_task_dirs",
        type=str,
        nargs="?",
        default=None,
        help="Optional raw task dir or comma-separated dirs under data/<bench>/; defaults to ckpt_name.",
    )
    parser.add_argument(
        "--mode",
        type=str,
        choices=["video", "image"],
        default="image",
        help="Whether to store images as videos or individual image files",
    )
    parser.add_argument(
        "--instruction",
        type=str,
        default="Do your job.",
        help="Default instruction when not present in HDF5",
    )
    args = parser.parse_args()

    bench_name = args.bench_name
    ckpt_name = args.ckpt_name
    env_cfg_type = args.env_cfg_type
    action_type = args.action_type
    repo_id = f"{bench_name}-{ckpt_name}-{env_cfg_type}-{action_type}"
    mode = args.mode
    instruction = args.instruction
    expert_data_num = None
    raw_task_dirs_arg = args.raw_task_dirs
    if args.expert_data_num is not None:
        try:
            expert_data_num = int(args.expert_data_num)
        except ValueError:
            if args.raw_task_dirs is not None:
                raise ValueError("raw_task_dirs was provided twice.") from None
            raw_task_dirs_arg = args.expert_data_num
    raw_task_dirs = [item.strip() for item in (raw_task_dirs_arg or ckpt_name).split(",") if item.strip()]

    env_cfg = load_yaml(os.path.join(ROOT_PATH, "./env_cfg", f"{env_cfg_type}.yml"))
    robot_type = env_cfg['config']['robot']

    robot_action_dim_info = robot_action_dim_info = load_json(os.path.join(ROOT_PATH, "env_cfg/robot", "_robot_info.json"))[robot_type]

    dataset = create_empty_dataset(
        repo_id=repo_id,
        robot_type=robot_type,
        fps=50, # pi default to 50
        mode=mode,
        dataset_config=DEFAULT_DATASET_CONFIG,
        robot_action_dim_info=robot_action_dim_info,
    )

    episode_files = []
    for raw_task_dir in raw_task_dirs:
        load_data_dir = Path(ROOT_PATH) / "data" / str(bench_name) / raw_task_dir / str(env_cfg_type)
        task_episode_files = sorted(load_data_dir.glob("data/episode_*.hdf5"))
        if not task_episode_files:
            task_episode_files = sorted(load_data_dir.glob("*.hdf5"))
        episode_files.extend(task_episode_files)
    if expert_data_num is not None:
        episode_files = episode_files[:expert_data_num]
    for ep_file in tqdm(episode_files, desc="Processing episodes", unit="episode"):
        try:
            data = load_data(ep_file)
            num_frames = data["state"].shape[0]

            for i in range(num_frames):
                frame = {
                    "observation.state": data["state"][i],
                    "action": data["action"][i],
                    "task": instruction if data["instructions"] is None else random.choice(data["instructions"]),
                }
                for camera_name, images in data["images"].items():
                    frame[f"observation.images.{camera_name}"] = images[i]

                dataset.add_frame(frame)
            
            dataset.save_episode()
            dataset.hf_dataset = dataset.create_hf_dataset()
            tqdm.write(f"Finished {ep_file.name} with {num_frames} frames")
        except Exception as e:
            tqdm.write(f"Error processing episode {ep_file}: {e}")

if __name__ == "__main__":
    main()