Raffael-Kultyshev's picture
Update dataset to DynamicIntelligence/humanoid-robots-training-dataset
fb66e8f verified
raw
history blame
13.6 kB
import json
import os
import logging
import sys
import html
from pathlib import Path
from typing import Dict, List, Optional
from functools import lru_cache
import gradio as gr
import pandas as pd
import plotly.graph_objects as go
import plotly.io as pio
from huggingface_hub import snapshot_download, HfApi
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger = logging.getLogger(__name__)
DEFAULT_DATASET_ID = os.getenv(
"DATASET_ID", "DynamicIntelligence/humanoid-robots-training-dataset"
)
LOCAL_DATASET_DIR = Path("dataset_cache")
HF_TOKEN = os.getenv("HF_TOKEN")
JOINT_ALIASES = {
"wrist": "Wrist",
"thumb_tip": "Thumb Tip",
"index_mcp": "Index MCP",
"index_tip": "Index Tip",
}
JOINT_NAME_MAP = {
"wrist": "WRIST",
"thumb_tip": "THUMB_TIP",
"index_mcp": "INDEX_FINGER_MCP",
"index_tip": "INDEX_FINGER_TIP",
}
METRIC_LABELS = {
"x_cm": "X (cm)",
"y_cm": "Y (cm)",
"z_cm": "Z (cm)",
"yaw_deg": "Yaw (°)",
"pitch_deg": "Pitch (°)",
"roll_deg": "Roll (°)",
}
PLOT_GRID = [
["x_cm", "y_cm", "z_cm"],
["yaw_deg", "pitch_deg", "roll_deg"],
]
PLOT_ORDER = [metric for row in PLOT_GRID for metric in row]
CUSTOM_CSS = """
:root, .gradio-container, body {
background-color: #050a18 !important;
color: #f8fafc !important;
font-family: 'Inter', 'Segoe UI', system-ui, sans-serif;
}
.side-panel {
background: #0f172a;
padding: 20px;
border-radius: 18px;
border: 1px solid #1f2b47;
min-height: 100%;
}
.stats-card ul {
list-style: none;
padding: 0;
margin: 0;
font-size: 0.92rem;
}
.stats-card li {
margin-bottom: 10px;
color: #e2e8f0;
}
.stats-card span {
display: inline-block;
margin-right: 6px;
color: #7dd3fc;
}
.episodes-title {
margin: 18px 0 8px;
font-size: 0.78rem;
text-transform: uppercase;
letter-spacing: 0.14em;
color: #94a3b8;
}
.episode-list .gr-form {
padding: 0;
}
.episode-list .gr-form > div {
gap: 0;
}
.episode-list input[type="radio"] {
display: none;
}
.episode-list label {
background: transparent !important;
border: none !important;
color: #cbd5f5 !important;
padding: 3px 0 !important;
justify-content: flex-start;
font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", "Courier New", monospace;
font-size: 0.9rem;
text-decoration: underline;
}
.episode-list label:hover {
color: #67e8f9 !important;
cursor: pointer;
}
.episode-list input[type="radio"]:checked + label {
color: #facc15 !important;
font-weight: 700;
margin-left: -2px;
}
.main-panel {
padding-top: 8px;
}
.instruction-card {
background: #0f172a;
padding: 18px 20px;
border-radius: 18px;
border: 1px solid #1f2b47;
}
.instruction-label {
font-size: 0.75rem;
letter-spacing: 0.12em;
text-transform: uppercase;
color: #94a3b8;
margin-bottom: 10px;
}
.instruction-text {
font-size: 1.1rem;
line-height: 1.5;
}
.video-card {
background: #0f172a;
border: 1px solid #1f2b47;
border-radius: 18px;
padding: 18px 20px;
margin-top: 18px;
}
.video-title {
font-size: 0.78rem;
text-transform: uppercase;
letter-spacing: 0.18em;
color: #94a3b8;
margin-bottom: 8px;
}
.video-panel video {
border-radius: 12px;
border: 1px solid #1f2b47;
background: #030712;
}
.download-button button {
border-radius: 999px;
border: 1px solid #334155;
background: #1e293b;
color: #f8fafc;
font-size: 0.85rem;
padding: 8px 24px;
}
.download-button button:hover {
border-color: #67e8f9;
color: #67e8f9;
}
.plots-wrap {
margin-top: 18px;
}
.plots-wrap .gr-row {
gap: 16px;
}
.plot-html {
background: #111a2c;
border-radius: 12px;
padding: 10px;
border: 1px solid #1f2b47;
min-height: 320px;
}
.plot-html iframe {
width: 100%;
height: 300px;
border: none;
}
"""
@lru_cache(maxsize=1)
def get_dataset_revision(repo_id: str) -> Optional[str]:
try:
info = HfApi(token=HF_TOKEN).repo_info(repo_id=repo_id, repo_type="dataset")
return info.sha
except Exception as exc:
logger.warning(f"Could not fetch dataset revision for {repo_id}: {exc}")
return None
@lru_cache(maxsize=2)
def get_dataset_root(repo_id: str, revision: Optional[str]) -> Path:
local_path = snapshot_download(
repo_id=repo_id,
repo_type="dataset",
local_dir=LOCAL_DATASET_DIR,
local_dir_use_symlinks=False,
revision=revision,
token=HF_TOKEN,
)
return Path(local_path)
@lru_cache(maxsize=2)
def load_info(repo_id: str, revision: Optional[str]) -> Dict:
root = get_dataset_root(repo_id, revision)
info_path = root / "meta" / "info.json"
with open(info_path, "r", encoding="utf-8") as f:
return json.load(f)
def resolve_path(root: Path, template: str, episode_chunk: int, episode_index: int) -> Path:
if isinstance(template, dict):
rgb_template = template.get("rgb")
if rgb_template is None:
raise ValueError("RGB template missing from metadata")
return root / rgb_template.format(episode_chunk=episode_chunk, episode_index=episode_index)
return root / template.format(episode_chunk=episode_chunk, episode_index=episode_index)
@lru_cache(maxsize=64)
def load_episode(repo_id: str, episode_index: int, revision: Optional[str]) -> Dict:
info = load_info(repo_id, revision)
root = get_dataset_root(repo_id, revision)
episode_meta = next((ep for ep in info["episodes"] if ep["episode_index"] == episode_index), None)
if not episode_meta:
raise ValueError(f"Episode {episode_index} not found in metadata")
chunk = episode_meta["episode_chunk"]
parquet_path = resolve_path(root, info["data_path"], chunk, episode_index)
if not parquet_path.exists():
raise FileNotFoundError(f"Parquet file not found: {parquet_path}")
df = pd.read_parquet(parquet_path)
timestamps, state_df = build_state_dataframe(df)
rgb_path = resolve_path(root, info["video_path"], chunk, episode_index)
instruction = (
episode_meta.get("language_instruction")
or (
df["language_instruction"].dropna().iloc[0]
if "language_instruction" in df.columns and not df["language_instruction"].isna().all()
else info.get("task", "Tape roll to bowl")
)
)
return {
"timestamps": timestamps,
"state_df": state_df,
"rgb_path": rgb_path,
"instruction": instruction,
}
def build_state_dataframe(df: pd.DataFrame) -> (List[float], pd.DataFrame):
if "frame_idx" not in df.columns or "timestamp_s" not in df.columns:
raise ValueError("Episode parquet missing frame timing information.")
frame_times = (
df[["frame_idx", "timestamp_s"]]
.drop_duplicates("frame_idx")
.set_index("frame_idx")
.sort_index()
)
frame_indices = frame_times.index.to_list()
state_df = pd.DataFrame(index=frame_indices)
for alias, joint_name in JOINT_NAME_MAP.items():
joint_df = (
df[df["joint_name"] == joint_name]
.set_index("frame_idx")
.sort_index()
.reindex(frame_indices)
)
for metric in METRIC_LABELS.keys():
if metric in joint_df.columns:
state_df[f"{alias}_{metric}"] = joint_df[metric].astype(float)
state_df.reset_index(drop=True, inplace=True)
timestamps = frame_times["timestamp_s"].to_list()
return timestamps, state_df
def build_plot_fig(data: Dict, metric: str) -> go.Figure:
timestamps = data["timestamps"]
state_df = data["state_df"]
fig = go.Figure()
for alias, label in JOINT_ALIASES.items():
col_name = f"{alias}_{metric}"
if col_name not in state_df.columns:
continue
fig.add_trace(
go.Scatter(
x=timestamps,
y=state_df[col_name],
mode="lines",
name=label,
)
)
fig.update_layout(
margin=dict(l=20, r=20, t=30, b=20),
height=250,
template="plotly_dark",
legend=dict(orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1),
xaxis_title="Time (s)",
yaxis_title=METRIC_LABELS[metric],
)
fig.update_xaxes(showgrid=True, gridwidth=0.5, gridcolor="rgba(255,255,255,0.1)")
fig.update_yaxes(showgrid=True, gridwidth=0.5, gridcolor="rgba(255,255,255,0.1)")
return fig
def build_plot_html(data: Dict, metric: str) -> str:
fig = build_plot_fig(data, metric)
return pio.to_html(fig, include_plotlyjs="cdn", full_html=False)
def format_episode_label(idx: int) -> str:
return f"Episode {idx:02d}"
def parse_episode_label(label: str) -> int:
return int(label.replace("Episode", "").strip())
def format_instruction_html(text: str) -> str:
safe_text = html.escape(text)
return (
'<div class="instruction-card">'
'<p class="instruction-label">Language Instruction</p>'
f'<p class="instruction-text">{safe_text}</p>'
"</div>"
)
def build_interface():
revision = get_dataset_revision(DEFAULT_DATASET_ID)
info = load_info(DEFAULT_DATASET_ID, revision)
episode_indices = sorted(ep["episode_index"] for ep in info["episodes"])
if not episode_indices:
raise RuntimeError("No episodes found in dataset metadata.")
default_idx = episode_indices[0]
default_label = format_episode_label(default_idx)
default_data = load_episode(DEFAULT_DATASET_ID, default_idx, revision)
default_video = str(default_data["rgb_path"])
default_instruction = default_data["instruction"]
default_figs = {metric: build_plot_html(default_data, metric) for metric in METRIC_LABELS.keys()}
total_frames = sum(ep.get("num_frames", 0) for ep in info["episodes"])
fps = info.get("fps", 30.0)
stats_html = f"""
<div class="stats-card">
<ul>
<li><span>Number of samples/frames:</span> {total_frames:,}</li>
<li><span>Number of episodes:</span> {len(episode_indices)}</li>
<li><span>Frames per second:</span> {fps:.1f}</li>
</ul>
</div>
"""
theme = gr.themes.Soft(
primary_hue="cyan", secondary_hue="blue", neutral_hue="slate"
).set(
body_background_fill="#0c1424",
body_text_color="#f8fafc",
block_background_fill="#111a2c",
block_title_text_color="#f8fafc",
input_background_fill="#151f33",
border_color_primary="#1f2b47",
shadow_drop="none",
)
with gr.Blocks(theme=theme, css=CUSTOM_CSS) as demo:
gr.Markdown("# Humanoid Robots Hand Pose Viewer")
gr.Markdown(
"Visualize RGB + 6DoF hand trajectories for all Moving_Mini tasks "
"(humanoid-robots-training-dataset)."
)
with gr.Row(equal_height=True):
with gr.Column(scale=1, min_width=260, elem_classes=["side-panel"]):
gr.HTML(stats_html)
gr.HTML('<div class="episodes-title">Episodes</div>')
episode_radio = gr.Radio(
choices=[format_episode_label(i) for i in episode_indices],
value=default_label,
label="Episodes",
elem_classes=["episode-list"],
)
with gr.Column(scale=2, min_width=640, elem_classes=["main-panel"]):
instruction_box = gr.HTML(
format_instruction_html(default_instruction),
label="Language Instruction",
)
with gr.Column(elem_classes=["video-card"]):
gr.HTML('<div class="video-title">RGB</div>')
video = gr.Video(
height=360,
value=default_video,
elem_classes=["video-panel"],
show_label=False,
show_download_button=False,
)
download_button = gr.DownloadButton(
label="Download",
value=default_video,
elem_classes=["download-button"],
)
plot_outputs = []
gr.Markdown("### Joint trajectories", elem_classes=["plots-title"])
with gr.Column(elem_classes=["plots-wrap"]):
for row in PLOT_GRID:
with gr.Row():
for metric in row:
plot = gr.HTML(value=default_figs[metric], elem_classes=["plot-html"])
plot_outputs.append(plot)
outputs = [instruction_box, video, download_button] + plot_outputs
def load_episode_payload(label: str):
idx = parse_episode_label(label)
data = load_episode(DEFAULT_DATASET_ID, idx, revision)
video_path = str(data["rgb_path"])
figs = [build_plot_html(data, metric) for metric in PLOT_ORDER]
return [
format_instruction_html(data["instruction"]),
video_path,
gr.DownloadButton.update(value=video_path),
*figs,
]
episode_radio.change(fn=load_episode_payload, inputs=episode_radio, outputs=outputs)
return demo
def main():
demo = build_interface()
demo.queue().launch(show_api=False)
if __name__ == "__main__":
main()