code-HFD_lzk / scripts /infer /render_third_party_scale.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
bcdb5ec verified
Raw History Blame Contribute Delete
8.99 kB
#!/usr/bin/env python3
"""CPU-only final scale previews, with separate train and held-out indexes."""
from __future__ import annotations
import argparse
from concurrent.futures import ThreadPoolExecutor, as_completed
import os
from pathlib import Path
import sys
import time
import wave
PROJECT = Path(__file__).resolve().parents[2]
sys.path[:0] = [str(PROJECT), str(PROJECT / "src")]
from scripts.infer.render_third_party_review import (
ASSETS, LABELS, contained, digest, read, render_one, write)
def selected_rows(selection, report, step):
if (selection.get("scaleup") is not True or selection.get("split") not in ("train", "validation")
or report.get("status") != "complete" or report.get("step") != step
or report.get("evaluation_split") != selection["split"]):
raise ValueError("Final prediction split/step differs from its frozen selection")
by_id = {r["session_id"]: r for r in selection["sessions"]}
samples = {r["session_id"]: r for r in report["samples"]}
if len(samples) != report["count"] or len(samples) != len(report["samples"]):
raise ValueError("Duplicate/incomplete final predictions")
ids = selection["video_session_ids"]
if not ids or len(ids) != len(set(ids)) or not set(ids) <= samples.keys() & by_id.keys():
raise ValueError("Missing predetermined video samples")
rows = []
for index, sid in enumerate(ids, 1):
row, chosen = samples[sid], by_id[sid]
if (row["profile"] != chosen["session_type"] or row["split"] != selection["split"]
or row["background_language"] != chosen["background_language"]):
raise ValueError("Prediction labels differ from frozen selection")
rows.append({**row, "number": f"{index:02d}", "background_high": chosen["background_high"]})
return rows
def prepare(run, manifest):
import numpy as np
run = contained(run, PROJECT / "runs")
result, identity = read(run / "result.json"), read(run / "scale_identity.json")
if result["status"] != "complete":
raise ValueError("Training/evaluation has not completed")
rows, hashes = [], {}
for split in ("train", "validation"):
path = manifest if split == "train" else manifest.parent / "validation.json"
key = "manifest_sha256" if split == "train" else "validation_sha256"
if digest(path) != identity[key]:
raise ValueError("Frozen video selection changed")
selection = read(path)
report_path = run / "evaluations/final" / split / "report.json"
report = read(report_path); hashes[split] = digest(report_path)
for row in selected_rows(selection, report, result["step"]):
source = contained(row["prediction_directory"], report_path.parent / "samples")
meta, recipe = read(source / "infer_meta.json"), read(source / "session_recipe.json")
if (meta["session_id"] != row["session_id"] or meta["global_step"] != result["step"]
or meta.get("evaluation_split") != split or meta["artifact_role"] != "model_prediction_not_GT"
or recipe["session_id"] != row["session_id"] or recipe["split"] != split):
raise ValueError("Prediction provenance mismatch")
audio = contained(meta["frozen_audio_path"], Path(selection["production_root"]) / "audio" / split)
if digest(audio) != meta["frozen_audio_sha256"] or digest(audio) != recipe["audio"]["sha256"]:
raise ValueError("Original mixed audio changed")
with wave.open(str(audio), "rb") as wav:
if (wav.getframerate(), wav.getnchannels(), wav.getsampwidth()) != (16000, 1, 2):
raise ValueError("Expected original mono 16kHz PCM16")
seconds = wav.getnframes() / 16000
frames = round(seconds * 50)
motion = np.load(source / "motion.npy", allow_pickle=False)
if motion.shape != (frames, 36) or not np.isfinite(motion).all() or digest(source / "motion.npy") != row["motion_sha256"]:
raise ValueError("Prediction trajectory changed or incomplete")
for name, width in (("body_pos.csv", 3), ("body_quat.csv", 4), ("joint_pos.csv", 29)):
values = np.loadtxt(source / name, delimiter=",", skiprows=1, ndmin=2)
if values.shape != (frames, width) or not np.isfinite(values).all():
raise ValueError(f"Invalid predicted render CSV: {name}")
rows.append({**row, "source_directory": str(source), "audio_path": str(audio),
"audio_sha256": meta["frozen_audio_sha256"], "duration_seconds": seconds,
"source_files": {name: digest(source / name) for name in ASSETS}})
return {"schema": 1, "run_dir": str(run), "step": result["step"], "count": len(rows), "samples": rows,
"report_hashes": hashes, "gpu_used": False, "inference_repeated": False,
"selection_policy": "predetermined one per available class/language/loudness cell, never by model score",
"artifact_role": "model_prediction_not_GT", "render": {"workers": 12, "fps": 25, "size": [320, 240]}}
def publish(output, plan, completed):
done = {(r["split"], r["number"]) for r in completed}
lines = ["# 全量训练最终预测视频(不是 GT)", "",
f"新阶段 checkpoint {plan['step']};完成 {len(completed)}/{plan['count']}。",
"复用完整预测;原混音不改变音量和时间。CPU 并行渲染,320×240、25fps;原预测仍为50Hz。",
"样例在训练前按六类×中英文背景×普通/较高音量固定随机选定,不按效果挑选。", ""]
for split, label in (("train", "训练集(拟合)"), ("validation", "独立验证集(未参与优化)")):
lines += [f"## {label}", "", "| 视频序号 | 类型 | 背景语言 | 背景音量 | 控制准确率 | 背景误动率 |",
"|---|---|---|---|---:|---:|"]
for row in plan["samples"]:
if row["split"] != split:
continue
number = row["number"]
link = f"[{number}]({split}/{number}.mp4)" if (split, number) in done else f"{number} 待生成"
metrics = row["metrics"]; bg = metrics.get("background_only_false_motion_rate")
bg_text = "无评分区间" if bg is None else f"{bg:.1%}"
lines.append(f"| {link} | {LABELS[row['profile']]} | {'中文' if row['background_language']=='zh' else '英文'} | "
f"{'较高' if row['background_high'] else '普通'} | {metrics['control_accuracy']:.1%} | {bg_text} |")
lines += ["", f"[本组原始逐条指标](../evaluations/final/{split}/report.json)", ""]
lines += ["单条视频指标用于定位问题,不代替全部验证结果,也不代表机器人实际执行成功。", "",
"[渲染状态](status.json) · [完整抽样溯源](plan.json)", ""]
(output / "README.md").write_text("\n".join(lines))
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--run-dir", type=Path, required=True)
parser.add_argument("--manifest", type=Path, required=True)
args = parser.parse_args()
os.environ.update(CUDA_VISIBLE_DEVICES="", OMP_NUM_THREADS="1", MKL_NUM_THREADS="1", PYTHONDONTWRITEBYTECODE="1")
from scripts.infer.third_party_checkpoint_probe import check_renderer
plan = prepare(args.run_dir, args.manifest.resolve())
output = Path(plan["run_dir"]) / "prediction_videos"
output.mkdir(exist_ok=False)
write(output / "plan.json", plan)
for split in ("train", "validation"):
(output / split).mkdir()
completed, failures = [], []
def status(phase):
write(output / "status.json", {"phase": phase, "time": time.time(), "pid": os.getpid(),
"completed": len(completed), "total": plan["count"], "failures": failures,
"gpu_used": False, "inference_repeated": False})
publish(output, plan, completed)
status("starting")
try:
check_renderer()
with ThreadPoolExecutor(max_workers=12) as pool:
jobs = {pool.submit(render_one, row, output / row["split"]): row for row in plan["samples"]}
for future in as_completed(jobs):
row = jobs[future]
try:
completed.append(future.result())
print(f"{row['split']}/{row['number']} completed", flush=True)
except Exception as error:
failures.append({"split": row["split"], "number": row["number"], "error": repr(error)})
status("rendering")
if failures:
raise RuntimeError("Some final videos failed; inspect status.json and per-sample render.log")
status("completed")
except Exception as error:
failures.append({"error": repr(error)}); status("failed"); raise
if __name__ == "__main__":
main()