{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# PI0：保护训练与评估\n", "\n", "本 Notebook 只执行 protected recipe。训练、checkpoint 保存、耗时记录、14 条固定 seed 闭环评估、视频和动作序列图都在 Notebook kernel 内完成。评估严格绑定本次训练生成的 `TRAINED_POLICY_PATH`，不会回退到历史或预训练权重。\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from pathlib import Path\n", "import json\n", "import os\n", "import shlex\n", "import shutil\n", "import subprocess\n", "import sys\n", "\n", "try:\n", " from IPython.display import HTML, Markdown, display\n", "except Exception:\n", " class Markdown(str):\n", " pass\n", "\n", " class HTML(str):\n", " pass\n", "\n", " def display(obj):\n", " print(obj)\n", "\n", "\n", "def find_topic_root():\n", " cwd = Path.cwd().resolve()\n", " for candidate in [cwd, *cwd.parents]:\n", " if (candidate / \"assets\" / \"metrics_snapshot.json\").exists():\n", " return candidate\n", " raise RuntimeError(\"请从 AMD ROCm 专题目录或 notebooks 子目录启动 Jupyter。\")\n", "\n", "\n", "TOPIC_ROOT = find_topic_root()\n", "ASSET_DIR = TOPIC_ROOT / \"assets\"\n", "PROJECT_ROOT = Path(os.environ.get(\"PROJECT_ROOT\", \"/path/to/04mujoco复现ACT、Pi0、SmolVLA\"))\n", "DATA_ROOT = Path(os.environ.get(\"DATA_ROOT\", \"/path/to/datasets/every_embodied\"))\n", "MODEL_ROOT = Path(os.environ.get(\"MODEL_ROOT\", \"/path/to/model/checkpoints\"))\n", "OUTPUT_ROOT = Path(os.environ.get(\"OUTPUT_ROOT\", TOPIC_ROOT / 'outputs' / 'protected' / 'pi0'))\n", "\n", "# The AMD teaching workflow should be runnable from local datasets/checkpoints.\n", "# Avoid surprising network calls during class or when AUP/Radeon Cloud cannot\n", "# reach Hugging Face.\n", "os.environ.setdefault(\"HF_HUB_OFFLINE\", os.environ.get(\"NOTEBOOK_HF_OFFLINE\", \"1\"))\n", "os.environ.setdefault(\"TRANSFORMERS_OFFLINE\", os.environ.get(\"NOTEBOOK_HF_OFFLINE\", \"1\"))\n", "os.environ.setdefault(\"HF_DATASETS_OFFLINE\", os.environ.get(\"NOTEBOOK_HF_OFFLINE\", \"1\"))\n", "os.environ.setdefault(\"HF_HOME\", str(Path(os.environ.get(\"CACHE_ROOT\", OUTPUT_ROOT / \"cache\")) / \"huggingface\"))\n", "os.environ.setdefault(\"HF_DATASETS_CACHE\", str(Path(os.environ[\"HF_HOME\"]) / \"datasets\"))\n", "\n", "def public_path(path):\n", " path = Path(path)\n", " replacements = [\n", " (TOPIC_ROOT, \"$TOPIC_ROOT\"),\n", " (PROJECT_ROOT, \"$PROJECT_ROOT\"),\n", " (DATA_ROOT, \"$DATA_ROOT\"),\n", " (MODEL_ROOT, \"$MODEL_ROOT\"),\n", " (OUTPUT_ROOT, \"$OUTPUT_ROOT\"),\n", " ]\n", " value = str(path)\n", " for root, label in sorted(replacements, key=lambda item: len(str(item[0])), reverse=True):\n", " root_value = str(root)\n", " if root_value and value.startswith(root_value):\n", " return label + value[len(root_value):]\n", " return value\n", "\n", "\n", "print(\"TOPIC_ROOT = $TOPIC_ROOT\")\n", "print(\"PROJECT_ROOT =\", public_path(PROJECT_ROOT))\n", "print(\"DATA_ROOT =\", public_path(DATA_ROOT))\n", "print(\"MODEL_ROOT =\", public_path(MODEL_ROOT))\n", "print(\"OUTPUT_ROOT =\", public_path(OUTPUT_ROOT))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "def md_table(headers, rows):\n", " lines = [\"| \" + \" | \".join(headers) + \" |\", \"| \" + \" | \".join([\"---\"] * len(headers)) + \" |\"]\n", " for row in rows:\n", " lines.append(\"| \" + \" | \".join(public_path(x) if isinstance(x, (str, Path)) else str(x) for x in row) + \" |\")\n", " display(Markdown(\"\\n\".join(lines)))\n", "\n", "\n", "def show_json(path, max_chars=5000):\n", " path = Path(path)\n", " if not path.exists():\n", " print(\"文件不存在：\", public_path(path))\n", " return None\n", " data = json.loads(path.read_text(encoding=\"utf-8\"))\n", " text = json.dumps(data, ensure_ascii=False, indent=2)\n", " print(text[:max_chars] + (\"\\n...\" if len(text) > max_chars else \"\"))\n", " return data\n", "\n", "\n", "def show_video(filename, title):\n", " path = ASSET_DIR / filename\n", " display(Markdown(f\"**{title}**\"))\n", " if path.exists():\n", " rel = f\"../assets/{filename}\"\n", " display(HTML(f\"

\"))\n", " else:\n", " print(\"缺少视频素材：\", public_path(path))\n", "\n", "\n", "def show_image(filename, title, width=960):\n", " path = ASSET_DIR / filename\n", " display(Markdown(f\"**{title}**\"))\n", " if path.exists():\n", " rel = f\"../assets/{filename}\"\n", " display(HTML(f\"

\"))\n", " else:\n", " print(\"缺少图片素材：\", public_path(path))\n", "\n", "\n", "def run_cmd_preview(command, cwd=None):\n", " shown = [public_path(x) if isinstance(x, (str, Path)) else x for x in command]\n", " print(\"$\", shlex.join([str(x) for x in shown]))\n", " if cwd:\n", " print(\"cwd =\", public_path(cwd))\n", "\n", "\n", "def tail_log(log_path, lines=40):\n", " path = Path(log_path)\n", " if not path.exists():\n", " print(\"日志不存在：\", public_path(path))\n", " return\n", " content = path.read_text(encoding=\"utf-8\", errors=\"replace\").splitlines()\n", " print(\"\\n\".join(content[-lines:]))\n", "\n", "\n", "def env_flag(name, default=False):\n", " value = os.environ.get(name)\n", " if value is None:\n", " return default\n", " return value.strip().lower() in {\"1\", \"true\", \"yes\", \"y\", \"on\"}\n", "\n", "\n", "RUN_SMOKE = env_flag(\"RUN_SMOKE\")\n", "RUN_LONG_TRAIN = env_flag(\"RUN_LONG_TRAIN\", True)\n", "RUN_EVAL = env_flag(\"RUN_EVAL\")\n", "EVAL_SCRIPT = Path(os.environ.get(\"EVAL_SCRIPT\", PROJECT_ROOT / \"eval_policy_success.py\"))\n", "\n", "\n", "_XVFB_PROCESS = None\n", "\n", "\n", "def ensure_xvfb_display():\n", " \"\"\"Start a lightweight virtual display for headless MuJoCo evaluation.\"\"\"\n", " global _XVFB_PROCESS\n", " if os.environ.get(\"DISPLAY\"):\n", " print(\"DISPLAY =\", os.environ[\"DISPLAY\"])\n", " return None\n", " xvfb_bin = shutil.which(\"Xvfb\")\n", " if not xvfb_bin:\n", " print(\"没有发现 Xvfb；如遇 GLFW DISPLAY 报错，请先安装 xvfb。\")\n", " return None\n", " display_id = os.environ.get(\"NOTEBOOK_XVFB_DISPLAY\", \":99\")\n", " _XVFB_PROCESS = subprocess.Popen(\n", " [xvfb_bin, display_id, \"-screen\", \"0\", \"1280x1024x24\"],\n", " stdout=subprocess.DEVNULL,\n", " stderr=subprocess.DEVNULL,\n", " )\n", " os.environ[\"DISPLAY\"] = display_id\n", " print(\"已启动 Notebook 内部 Xvfb：DISPLAY =\", display_id)\n", " return _XVFB_PROCESS\n", "\n", "\n", "def ensure_project_layout():\n", " required = [PROJECT_ROOT / \"asset\" / \"example_scene_y2.xml\", PROJECT_ROOT / \"mujoco_env\"]\n", " missing = [path for path in required if not path.exists()]\n", " if missing:\n", " print(\"当前 PROJECT_ROOT 还不是可运行工程，缺少：\")\n", " for path in missing:\n", " print(\" -\", public_path(path))\n", " print(\"请先设置 PROJECT_ROOT，再运行训练或评估单元。\")\n", " return False\n", " return True\n", "\n", "\n", "def write_json_yaml(path, payload):\n", " path = Path(path)\n", " path.parent.mkdir(parents=True, exist_ok=True)\n", " try:\n", " import yaml\n", " text = yaml.safe_dump(payload, allow_unicode=True, sort_keys=False)\n", " except Exception:\n", " text = json.dumps(payload, ensure_ascii=False, indent=2) + \"\\n\"\n", " path.write_text(text, encoding=\"utf-8\")\n", " print(\"写出配置：\", public_path(path))\n", " return path\n", "\n", "\n", "def make_lerobot_train_config(policy_type, dataset_repo_id, dataset_root, output_dir, steps, batch_size, chunk_size, n_action_steps, seed=42):\n", " save_freq = int(os.environ.get(f\"{policy_type.upper()}_SAVE_FREQ\", os.environ.get(\"SAVE_FREQ\", str(steps))))\n", " return {\n", " \"dataset\": {\n", " \"repo_id\": dataset_repo_id,\n", " \"root\": str(dataset_root),\n", " \"use_imagenet_stats\": True,\n", " },\n", " \"policy\": {\n", " \"type\": policy_type,\n", " \"chunk_size\": int(chunk_size),\n", " \"n_action_steps\": int(n_action_steps),\n", " \"device\": \"cuda\",\n", " },\n", " \"output_dir\": str(output_dir),\n", " \"job_name\": Path(output_dir).name,\n", " \"batch_size\": int(batch_size),\n", " \"steps\": int(steps),\n", " \"save_freq\": max(1, save_freq),\n", " \"log_freq\": 20,\n", " \"num_workers\": 4,\n", " \"seed\": int(seed),\n", " \"resume\": False,\n", " \"eval_freq\": -1,\n", " \"save_checkpoint\": True,\n", " \"use_policy_training_preset\": True,\n", " \"wandb\": {\"enable\": False, \"disable_artifact\": True},\n", " }\n", "\n", "\n", "\n", "\n", "class _NotebookCompactProgress:\n", " def __init__(self, iterable, desc, total=None):\n", " self.iterable = iterable\n", " self.desc = desc\n", " self.total = int(total if total is not None else len(iterable))\n", " self.every = max(1, int(os.environ.get(\"NOTEBOOK_PROGRESS_EVERY\", \"1\")))\n", " self.postfix = \"\"\n", "\n", " def set_postfix(self, **kwargs):\n", " self.postfix = \", \".join(f\"{key}={value}\" for key, value in kwargs.items())\n", "\n", " def __iter__(self):\n", " for index, value in enumerate(self.iterable, start=1):\n", " yield value\n", " if index == 1 or index == self.total or index % self.every == 0:\n", " suffix = f\" | {self.postfix}\" if self.postfix else \"\"\n", " print(f\"{self.desc}: {index}/{self.total}{suffix}\", flush=True)\n", "\n", "def notebook_progress(iterable, desc, total=None):\n", " return _NotebookCompactProgress(iterable, desc, total=total)\n", "\n", "\n", "def train_lerobot_config_in_notebook(config_path, enabled=False, progress_name=\"train\"):\n", " \"\"\"Run LeRobot offline training directly inside the notebook kernel.\n", "\n", " The notebook cell owns dataset creation, policy creation, optimizer steps,\n", " checkpoint saving, tqdm progress, and metric JSONL writing.\n", " \"\"\"\n", " config_path = Path(config_path)\n", " print(\"config =\", public_path(config_path))\n", " if not enabled:\n", " print(\"未启动。设置 RUN_SMOKE=1 或 RUN_LONG_TRAIN=1 后，本单元会直接在 Notebook 内训练。\")\n", " return None\n", " if not ensure_project_layout():\n", " return None\n", "\n", " import time\n", " from datetime import datetime, timezone\n", " from contextlib import nullcontext\n", "\n", " import draccus\n", " import torch\n", " from torch.amp import GradScaler\n", "\n", " from lerobot.common.datasets.factory import make_dataset\n", " from lerobot.common.datasets.sampler import EpisodeAwareSampler\n", " from lerobot.common.datasets.utils import cycle\n", " from lerobot.common.optim.factory import make_optimizer_and_scheduler\n", " from lerobot.common.policies.factory import make_policy\n", " from lerobot.common.policies.utils import get_device_from_parameters\n", " from lerobot.common.utils.random_utils import set_seed\n", " from lerobot.common.utils.train_utils import get_step_checkpoint_dir, save_checkpoint, update_last_checkpoint\n", " from lerobot.common.utils.utils import get_safe_torch_device\n", " from lerobot.configs.train import TrainPipelineConfig\n", "\n", " class _CompactProgress:\n", " def __init__(self, iterable, desc, total=None):\n", " self.iterable = iterable\n", " self.desc = desc\n", " self.total = int(total if total is not None else len(iterable))\n", " self.every = max(1, int(os.environ.get(\"NOTEBOOK_PROGRESS_EVERY\", \"100\")))\n", " self.postfix = \"\"\n", "\n", " def set_postfix(self, **kwargs):\n", " self.postfix = \", \".join(f\"{key}={value}\" for key, value in kwargs.items())\n", "\n", " def __iter__(self):\n", " for index, value in enumerate(self.iterable, start=1):\n", " yield value\n", " if index == 1 or index == self.total or index % self.every == 0:\n", " suffix = f\" | {self.postfix}\" if self.postfix else \"\"\n", " print(f\"{self.desc}: {index}/{self.total}{suffix}\", flush=True)\n", "\n", " def notebook_progress(iterable, desc, total=None):\n", " return _CompactProgress(iterable, desc, total=total)\n", "\n", " run_started_at = datetime.now(timezone.utc).isoformat()\n", " cfg = draccus.parse(TrainPipelineConfig, config_path=config_path, args=[])\n", " cfg.validate()\n", " if cfg.seed is not None:\n", " set_seed(cfg.seed)\n", "\n", " device = get_safe_torch_device(cfg.policy.device, log=True)\n", " torch.backends.cudnn.benchmark = True\n", " torch.backends.cuda.matmul.allow_tf32 = True\n", "\n", " print(\"Creating dataset...\")\n", " dataset = make_dataset(cfg)\n", " print(\"Creating policy...\")\n", " pretrained_override = os.environ.get(f\"{cfg.policy.type.upper()}_PRETRAINED_PATH_OVERRIDE\") or os.environ.get(\"POLICY_PRETRAINED_PATH_OVERRIDE\")\n", " if pretrained_override and not cfg.resume:\n", " cfg.policy.pretrained_path = str(Path(pretrained_override))\n", " print(\"pretrained override =\", public_path(cfg.policy.pretrained_path))\n", " elif cfg.policy.type == \"pi0\" and not cfg.resume:\n", " cfg.policy.pretrained_path = \"lerobot/pi0\"\n", " elif cfg.policy.type == \"smolvla\" and not cfg.resume:\n", " smolvla_base_candidates = [\n", " os.environ.get(\"SMOLVLA_BASE_PATH\"),\n", " os.environ.get(\"SMOLVLA_PRETRAINED_BASE_PATH\"),\n", " str(MODEL_ROOT / \"smolvla_base\" / \"pretrained_model\"),\n", " str(MODEL_ROOT / \"lerobot_smolvla_base_legacy\"),\n", " str(MODEL_ROOT / \"lerobot_smolvla_base\"),\n", " ]\n", " local_smolvla_base = next((Path(p) for p in smolvla_base_candidates if p and Path(p).exists()), None)\n", " if local_smolvla_base is not None:\n", " cfg.policy.pretrained_path = str(local_smolvla_base)\n", " print(\"local smolvla base =\", public_path(cfg.policy.pretrained_path))\n", " else:\n", " cfg.policy.pretrained_path = \"lerobot/smolvla_base\"\n", " policy = make_policy(cfg=cfg.policy, ds_meta=dataset.meta)\n", "\n", " # Compatibility for newer Transformers: PaliGemmaForConditionalGeneration may expose\n", " # language_model as GemmaModel directly, while this LeRobot Pi0 code expects\n", " # language_model.model. Use a non-Module proxy so checkpoints/state_dict stay clean.\n", " if cfg.policy.type == \"pi0\":\n", " try:\n", " lm = policy.model.paligemma_with_expert.paligemma.language_model\n", " if not hasattr(lm, \"model\"):\n", " class _LanguageModelCoreProxy:\n", " def __init__(self, core):\n", " self._core = core\n", "\n", " def __getattr__(self, name):\n", " return getattr(self._core, name)\n", "\n", " object.__setattr__(lm, \"model\", _LanguageModelCoreProxy(lm))\n", " print(\"patched Pi0 PaliGemma language_model.model compatibility proxy\")\n", " except Exception as exc:\n", " print(f\"Pi0 PaliGemma compatibility patch skipped: {exc}\")\n", "\n", " policy.to(device)\n", " policy.train()\n", "\n", " optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)\n", " grad_scaler = GradScaler(device.type, enabled=cfg.policy.use_amp)\n", "\n", " def _dataset_column_values(name):\n", " hf_dataset = getattr(dataset, \"hf_dataset\", None)\n", " if hf_dataset is None or name not in getattr(hf_dataset, \"column_names\", []):\n", " return None\n", " values = hf_dataset[name]\n", " try:\n", " return list(values)\n", " except TypeError:\n", " return [values[i] for i in range(len(values))]\n", "\n", " def _task_name_map():\n", " meta = getattr(dataset, \"meta\", None)\n", " tasks = getattr(meta, \"tasks\", None)\n", " if tasks is None:\n", " return {}\n", " if isinstance(tasks, dict):\n", " return {int(k): str(v) for k, v in tasks.items()}\n", " try:\n", " return {int(k): str(v) for k, v in dict(tasks).items()}\n", " except Exception:\n", " return {}\n", "\n", " def _make_weighted_sampler(generator):\n", " mode = os.environ.get(\"NOTEBOOK_FRAME_WEIGHT_MODE\", \"\").strip().lower()\n", " if not mode or mode in {\"0\", \"none\", \"off\", \"false\"}:\n", " return None, {\"mode\": \"none\"}\n", " weights = torch.ones(len(dataset), dtype=torch.double)\n", " info = {\"mode\": mode, \"num_frames\": len(dataset)}\n", "\n", " if \"blue\" in mode:\n", " blue_weight = float(os.environ.get(\"NOTEBOOK_BLUE_WEIGHT\", \"2.0\"))\n", " mask = [False] * len(dataset)\n", " task_indices = _dataset_column_values(\"task_index\")\n", " task_names = _task_name_map()\n", " if task_indices is not None and task_names:\n", " for idx, task_index in enumerate(task_indices):\n", " task_text = task_names.get(int(task_index), \"\").lower()\n", " mask[idx] = (\"blue\" in task_text) or (\"蓝\" in task_text)\n", " else:\n", " for column in [\"task\", \"language_instruction\", \"instruction\"]:\n", " values = _dataset_column_values(column)\n", " if values is None:\n", " continue\n", " for idx, value in enumerate(values):\n", " text = str(value).lower()\n", " mask[idx] = (\"blue\" in text) or (\"蓝\" in text)\n", " break\n", " blue_count = int(sum(mask))\n", " if blue_count == 0:\n", " print(\"警告：NOTEBOOK_FRAME_WEIGHT_MODE=blue 但没有识别到 blue/蓝指令帧，采样退回均匀。\")\n", " else:\n", " for idx, is_blue in enumerate(mask):\n", " if is_blue:\n", " weights[idx] *= blue_weight\n", " info.update({\"blue_weight\": blue_weight, \"blue_frames\": blue_count})\n", "\n", " weight_file = os.environ.get(\"NOTEBOOK_FRAME_WEIGHT_JSON\")\n", " if weight_file:\n", " payload = json.loads(Path(weight_file).read_text(encoding=\"utf-8\"))\n", " for key, value in payload.items():\n", " weights[int(key)] *= float(value)\n", " info.update({\"weight_json\": public_path(weight_file), \"json_entries\": len(payload)})\n", "\n", " if float(weights.sum()) <= 0:\n", " raise ValueError(\"采样权重总和为 0。\")\n", " sampler = torch.utils.data.WeightedRandomSampler(\n", " weights=weights,\n", " num_samples=len(weights),\n", " replacement=True,\n", " generator=generator,\n", " )\n", " info.update(\n", " {\n", " \"weight_min\": float(weights.min()),\n", " \"weight_max\": float(weights.max()),\n", " \"weight_mean\": float(weights.mean()),\n", " }\n", " )\n", " return sampler, info\n", "\n", " generator = torch.Generator()\n", " if cfg.seed is not None:\n", " generator.manual_seed(int(cfg.seed))\n", "\n", " weighted_sampler, sampler_info = _make_weighted_sampler(generator)\n", " if weighted_sampler is not None:\n", " shuffle = False\n", " sampler = weighted_sampler\n", " print(\"Notebook weighted sampler =\", json.dumps(sampler_info, ensure_ascii=False))\n", " elif hasattr(cfg.policy, \"drop_n_last_frames\"):\n", " shuffle = False\n", " sampler = EpisodeAwareSampler(\n", " dataset.episode_data_index,\n", " drop_n_last_frames=cfg.policy.drop_n_last_frames,\n", " shuffle=True,\n", " )\n", " else:\n", " shuffle = True\n", " sampler = None\n", "\n", " dataloader = torch.utils.data.DataLoader(\n", " dataset,\n", " num_workers=cfg.num_workers,\n", " batch_size=cfg.batch_size,\n", " shuffle=shuffle,\n", " sampler=sampler,\n", " generator=generator if sampler is None else None,\n", " pin_memory=device.type != \"cpu\",\n", " drop_last=False,\n", " )\n", " dl_iter = cycle(dataloader)\n", "\n", " output_dir = Path(cfg.output_dir)\n", " output_dir.mkdir(parents=True, exist_ok=True)\n", " metrics_path = output_dir / \"notebook_train_metrics.jsonl\"\n", " num_learnable = sum(p.numel() for p in policy.parameters() if p.requires_grad)\n", " num_total = sum(p.numel() for p in policy.parameters())\n", " print(f\"output_dir = {public_path(output_dir)}\")\n", " print(f\"steps = {cfg.steps}, batch_size = {cfg.batch_size}, frames = {dataset.num_frames}, episodes = {dataset.num_episodes}\")\n", " print(f\"learnable_params = {num_learnable:,}, total_params = {num_total:,}\")\n", "\n", " last_metrics = None\n", " progress = notebook_progress(range(1, cfg.steps + 1), desc=progress_name, total=cfg.steps)\n", " start_all = time.perf_counter()\n", " for step in progress:\n", " load_start = time.perf_counter()\n", " batch = next(dl_iter)\n", " data_s = time.perf_counter() - load_start\n", " for key, value in batch.items():\n", " if isinstance(value, torch.Tensor):\n", " batch[key] = value.to(device, non_blocking=True)\n", "\n", " update_start = time.perf_counter()\n", " device_from_policy = get_device_from_parameters(policy)\n", " with torch.autocast(device_type=device_from_policy.type) if cfg.policy.use_amp else nullcontext():\n", " loss, output_dict = policy.forward(batch)\n", " grad_scaler.scale(loss).backward()\n", " grad_scaler.unscale_(optimizer)\n", " grad_norm = torch.nn.utils.clip_grad_norm_(\n", " policy.parameters(),\n", " cfg.optimizer.grad_clip_norm,\n", " error_if_nonfinite=False,\n", " )\n", " grad_scaler.step(optimizer)\n", " grad_scaler.update()\n", " optimizer.zero_grad()\n", " if lr_scheduler is not None:\n", " lr_scheduler.step()\n", " if hasattr(policy, \"update\"):\n", " policy.update()\n", " update_s = time.perf_counter() - update_start\n", "\n", " is_log_step = cfg.log_freq > 0 and (step % cfg.log_freq == 0 or step == 1 or step == cfg.steps)\n", " is_saving_step = cfg.save_checkpoint and (step % cfg.save_freq == 0 or step == cfg.steps)\n", " if is_log_step:\n", " last_metrics = {\n", " \"step\": step,\n", " \"loss\": float(loss.detach().cpu()),\n", " \"grad_norm\": float(grad_norm.detach().cpu()) if hasattr(grad_norm, \"detach\") else float(grad_norm),\n", " \"lr\": float(optimizer.param_groups[0][\"lr\"]),\n", " \"update_s\": float(update_s),\n", " \"data_s\": float(data_s),\n", " \"elapsed_s\": float(time.perf_counter() - start_all),\n", " }\n", " with metrics_path.open(\"a\", encoding=\"utf-8\") as f:\n", " f.write(json.dumps(last_metrics, ensure_ascii=False) + \"\\n\")\n", " progress.set_postfix(\n", " loss=f\"{last_metrics['loss']:.4f}\",\n", " lr=f\"{last_metrics['lr']:.1e}\",\n", " updt_s=f\"{last_metrics['update_s']:.3f}\",\n", " )\n", " if is_saving_step:\n", " checkpoint_dir = get_step_checkpoint_dir(cfg.output_dir, cfg.steps, step)\n", " print(f\"\\nSaving checkpoint at step {step}: {public_path(checkpoint_dir)}\")\n", " save_checkpoint(checkpoint_dir, step, cfg, policy, optimizer, lr_scheduler)\n", " update_last_checkpoint(checkpoint_dir)\n", "\n", " run_finished_at = datetime.now(timezone.utc).isoformat()\n", " duration_s = float(time.perf_counter() - start_all)\n", " checkpoint_paths = sorted(str(path) for path in output_dir.glob(\"checkpoints/*/pretrained_model\"))\n", " run_summary = {\n", " \"recipe\": str(cfg.policy.type),\n", " \"native_notebook_kernel\": True,\n", " \"started_at_utc\": run_started_at,\n", " \"finished_at_utc\": run_finished_at,\n", " \"duration_s\": duration_s,\n", " \"steps\": int(cfg.steps),\n", " \"batch_size\": int(cfg.batch_size),\n", " \"output_dir\": str(output_dir),\n", " \"metrics_path\": str(metrics_path),\n", " \"checkpoint_paths\": checkpoint_paths,\n", " \"last_metrics\": last_metrics,\n", " }\n", " summary_path = output_dir / \"training_run_summary.json\"\n", " summary_path.write_text(json.dumps(run_summary, ensure_ascii=False, indent=2), encoding=\"utf-8\")\n", " print(\"训练完成。metrics =\", public_path(metrics_path))\n", " print(\"训练耗时 =\", f\"{duration_s / 60.0:.2f} 分钟\")\n", " print(\"训练摘要 =\", public_path(summary_path))\n", " if last_metrics is not None:\n", " print(json.dumps(last_metrics, ensure_ascii=False, indent=2))\n", " return {\"output_dir\": output_dir, \"metrics_path\": metrics_path, \"summary_path\": summary_path, \"last_metrics\": last_metrics, \"duration_s\": duration_s}\n", "\n", "\n", "def load_eval_module():\n", " import importlib.util\n", "\n", " if not EVAL_SCRIPT.exists():\n", " raise FileNotFoundError(f\"评估脚本不存在：{public_path(EVAL_SCRIPT)}\")\n", " spec = importlib.util.spec_from_file_location(\"notebook_eval_policy_success\", EVAL_SCRIPT)\n", " module = importlib.util.module_from_spec(spec)\n", " assert spec.loader is not None\n", " spec.loader.exec_module(module)\n", " return module\n", "\n", "\n", "def run_eval_policy_in_notebook(\n", " policy_name,\n", " policy_path,\n", " result_path,\n", " episodes,\n", " seed_start,\n", " render=False,\n", " enabled=False,\n", " repo_id=None,\n", " dataset_root=None,\n", " video_dir=None,\n", " trace_dir=None,\n", " seed_list=None,\n", "):\n", " print(\"policy =\", policy_name)\n", " print(\"policy_path =\", public_path(policy_path))\n", " print(\"result =\", public_path(result_path))\n", " if not enabled:\n", " print(\"未启动。设置 RUN_EVAL=1 后，本单元会在 Notebook 内直接加载策略并闭环评估。\")\n", " return None\n", " if not ensure_project_layout():\n", " return None\n", "\n", " import argparse\n", " from contextlib import contextmanager\n", "\n", " @contextmanager\n", " def pushd(path):\n", " old = Path.cwd()\n", " os.chdir(path)\n", " try:\n", " yield\n", " finally:\n", " os.chdir(old)\n", "\n", " ensure_xvfb_display()\n", " module = load_eval_module()\n", " result_path = Path(result_path)\n", " result_path.parent.mkdir(parents=True, exist_ok=True)\n", " if result_path.exists():\n", " result_path.unlink()\n", "\n", " args = argparse.Namespace(\n", " policy=policy_name,\n", " episodes=int(episodes),\n", " seed_start=int(seed_start),\n", " max_action_steps=int(os.environ.get(\"EVAL_MAX_ACTION_STEPS\", \"400\")),\n", " hz=float(os.environ.get(\"EVAL_HZ\", \"20\")),\n", " render=bool(render),\n", " output_jsonl=result_path,\n", " video_dir=Path(video_dir) if video_dir else None,\n", " trace_dir=Path(trace_dir) if trace_dir else None,\n", " device=os.environ.get(\"EVAL_DEVICE\", \"cuda\"),\n", " reset_policy_each_action=env_flag(\"EVAL_RESET_POLICY_EACH_ACTION\", False),\n", " act_n_action_steps=None,\n", " act_force_dataset_gripper=False,\n", " act_clamp_timestamp=False,\n", " act_policy_path=Path(policy_path),\n", " act_repo_id=repo_id or \"datawhale_eai_pnp\",\n", " act_dataset_root=Path(dataset_root or \"./demo_data\"),\n", " act_episode_timestamp_offsets=\"\",\n", " act_episode_source_flags=\"\",\n", " physical_success=env_flag(\"EVAL_PHYSICAL_SUCCESS\", True),\n", " physical_min_lift=float(os.environ.get(\"EVAL_PHYSICAL_MIN_LIFT\", \"0.06\")),\n", " physical_min_lift_steps=int(os.environ.get(\"EVAL_PHYSICAL_MIN_LIFT_STEPS\", \"3\")),\n", " physical_final_upright_cos=float(os.environ.get(\"EVAL_PHYSICAL_FINAL_UPRIGHT_COS\", \"0.85\")),\n", " smolvla_policy_path=Path(policy_path),\n", " pi0_policy_path=Path(policy_path),\n", " pi0_repo_id=repo_id or os.environ.get(\"PI0_DATASET_REPO_ID\", \"datawhale_eai_pnp_language\"),\n", " pi0_dataset_root=Path(dataset_root or os.environ.get(\"PI0_DATASET_ROOT\", \"./demo_data_language\")),\n", " )\n", "\n", " with pushd(PROJECT_ROOT):\n", " selected_cases = set()\n", " selected_cases = set()\n", " selected_cases = set()\n", " if policy_name == \"act\":\n", " policy = module.make_act_policy(\n", " args.device,\n", " args.act_policy_path,\n", " args.act_repo_id,\n", " args.act_dataset_root,\n", " n_action_steps=args.act_n_action_steps,\n", " episode_timestamp_offsets=args.act_episode_timestamp_offsets,\n", " episode_source_flags=args.act_episode_source_flags,\n", " )\n", " rollout = module.rollout_act\n", " elif policy_name == \"smolvla\":\n", " policy = module.make_smolvla_policy(args.device, args.smolvla_policy_path)\n", " rollout = module.rollout_language_policy\n", " elif policy_name == \"pi0\":\n", " policy = module.make_pi0_policy(args.device, args.pi0_policy_path, args.pi0_repo_id, args.pi0_dataset_root)\n", " rollout = module.rollout_language_policy\n", " else:\n", " raise ValueError(policy_name)\n", "\n", " rows = []\n", " for offset in notebook_progress(range(args.episodes), desc=f\"{policy_name} eval\", total=args.episodes):\n", " seed = (list(seed_list)[offset] if seed_list is not None else args.seed_start + offset)\n", " row = rollout(args, policy, seed)\n", " row = module.finalize_case_media(args, row, selected_cases)\n", " rows.append(row)\n", " with result_path.open(\"a\", encoding=\"utf-8\") as f:\n", " f.write(json.dumps(row, ensure_ascii=False) + \"\\n\")\n", " print(json.dumps(row, ensure_ascii=False))\n", " summarize_jsonl(result_path)\n", " return rows\n", "\n", "\n", "def list_checkpoints(run_dir):\n", " run_dir = Path(run_dir)\n", " candidates = []\n", " for pattern in [\"checkpoints/*/pretrained_model\", \"checkpoint*/pretrained_model\", \"*/pretrained_model\", \"pretrained_model\"]:\n", " candidates.extend(run_dir.glob(pattern))\n", " unique = sorted(set(candidates))\n", " if not unique:\n", " print(\"尚未发现 checkpoint：\", public_path(run_dir))\n", " return []\n", " for path in unique:\n", " print(\" -\", public_path(path))\n", " return unique\n", "\n", "\n", "def require_trained_policy(run_dir, label=\"本轮训练\"):\n", " checkpoints = list_checkpoints(run_dir)\n", " if not checkpoints:\n", " raise RuntimeError(\n", " f\"{label}没有发现 checkpoint：{public_path(run_dir)}。\"\n", " \"请先在本 Notebook 执行训练单元格，不能回退到历史或预训练权重。\"\n", " )\n", " path = checkpoints[-1]\n", " if not path.exists():\n", " raise FileNotFoundError(f\"训练 checkpoint 不存在：{public_path(path)}\")\n", " print(f\"评估绑定到{label} checkpoint：\", public_path(path))\n", " return path\n", "\n", "\n", "def summarize_jsonl(path):\n", " path = Path(path)\n", " if not path.exists():\n", " print(\"结果 JSONL 尚不存在：\", public_path(path))\n", " return None\n", " rows = [json.loads(line) for line in path.read_text(encoding=\"utf-8\").splitlines() if line.strip()]\n", " total = len(rows)\n", " legacy = sum(bool(row.get(\"success\") or row.get(\"legacy_success\")) for row in rows)\n", " if rows and all(\"physical_success\" in row for row in rows):\n", " physical_count = sum(bool(row.get(\"physical_success\")) for row in rows)\n", " physical_text = str(physical_count) + \"/\" + str(total)\n", " else:\n", " physical_text = \"未记录\"\n", " md_table(\n", " [\"结果文件\", \"episodes\", \"legacy_success\", \"physical_success\"],\n", " [(public_path(path), total, f\"{legacy}/{total}\", physical_text)],\n", " )\n", " return rows\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 1. 运行控制\n", "\n", "本 Notebook 默认执行保护训练和评估。首次运行可先将 `RUN_PROTECTED_TRAIN=False` 做环境检查；正式复现时保持 `True`。训练进度每 100 步刷新一次。\n" ] }, { "cell_type": "code", "metadata": {}, "source": [ "# RUN_CONTROL_CELL\n", "RUN_SMOKE = False\n", "RUN_LONG_TRAIN = False\n", "RUN_EVAL = True\n", "RUN_PROTECTED_TRAIN = True\n", "os.environ[\"NOTEBOOK_PROGRESS_EVERY\"] = os.environ.get(\"NOTEBOOK_PROGRESS_EVERY\", \"100\")\n", "print({\"RUN_SMOKE\": RUN_SMOKE, \"RUN_LONG_TRAIN\": RUN_LONG_TRAIN, \"RUN_EVAL\": RUN_EVAL, \"RUN_PROTECTED_TRAIN\": RUN_PROTECTED_TRAIN})\n" ], "outputs": [], "execution_count": null }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 2. 保护训练配方\n", "\n", "保护初始化权重只用于训练初始化；最终评估只能使用本单元新生成的最终 checkpoint。\n" ] }, { "cell_type": "code", "metadata": {}, "source": [ "DATASET_REPO_ID = os.environ.get(\"PI0_DATASET_REPO_ID\", \"datawhale_eai_pnp_pi0_clean_oracle_y060z000_3100_3139_g8_rebuild_v1\")\n", "TRAIN_DATA_ROOT = Path(os.environ.get(\"PI0_TRAIN_DATA_ROOT\", DATA_ROOT / \"pi0_clean40_successonly\"))\n", "CONFIG_DIR = OUTPUT_ROOT / \"configs\"\n", "print(\"DATASET_REPO_ID =\", DATASET_REPO_ID)\n", "print(\"TRAIN_DATA_ROOT =\", public_path(TRAIN_DATA_ROOT))\n" ], "outputs": [], "execution_count": null }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 3. 执行 protected clean40 + blue2x 训练\n", "\n", "Pi0 会先在 Notebook 内训练 clean40 success-only parent，再从本次 parent 继续 blue2x 续训。不会跳过 parent，也不会加载外部保护结果作为评估模型。\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# PROTECTED_TRAIN_CELL\n", "protected_train_enabled = globals().get(\"RUN_PROTECTED_TRAIN\", env_flag(\"RUN_PROTECTED_TRAIN\", False))\n", "if not protected_train_enabled:\n", " print(\"未启动。设置 RUN_PROTECTED_TRAIN=1 后，本单元会原生训练 Pi0 protected recipe。\")\n", "else:\n", " DATASET_REPO_ID = globals().get(\"DATASET_REPO_ID\", \"datawhale_eai_pnp_pi0_clean_oracle_y060z000_3100_3139_g8_rebuild_v1\")\n", " TRAIN_DATA_ROOT = globals().get(\"TRAIN_DATA_ROOT\", Path(os.environ.get(\"PI0_TRAIN_DATA_ROOT\", DATA_ROOT / \"pi0_clean40_successonly\")))\n", " CONFIG_DIR = OUTPUT_ROOT / \"configs\"\n", " RUN_ROOT = OUTPUT_ROOT / \"runs\" / \"pi0_protected_recipe\"\n", " CLEAN_OUTPUT = RUN_ROOT / \"clean40_s7500\"\n", " BLUE_OUTPUT = RUN_ROOT / \"blue2x_s8500\"\n", " clean_config = make_lerobot_train_config(\n", " \"pi0\", DATASET_REPO_ID, TRAIN_DATA_ROOT, CLEAN_OUTPUT,\n", " steps=int(os.environ.get(\"PI0_CLEAN_STEPS\", \"7500\")),\n", " batch_size=int(os.environ.get(\"PI0_BATCH_SIZE\", \"2\")),\n", " chunk_size=50,\n", " n_action_steps=50,\n", " )\n", " blue_config = make_lerobot_train_config(\n", " \"pi0\", DATASET_REPO_ID, TRAIN_DATA_ROOT, BLUE_OUTPUT,\n", " steps=int(os.environ.get(\"PI0_BLUE_STEPS\", \"1000\")),\n", " batch_size=int(os.environ.get(\"PI0_BATCH_SIZE\", \"2\")),\n", " chunk_size=50,\n", " n_action_steps=50,\n", " )\n", " clean_path = write_json_yaml(CONFIG_DIR / \"pi0_protected_clean40_s7500.yaml\", clean_config)\n", " blue_path = write_json_yaml(CONFIG_DIR / \"pi0_protected_blue2x_s8500.yaml\", blue_config)\n", " clean_result = train_lerobot_config_in_notebook(clean_path, enabled=True, progress_name=\"Pi0 protected clean40\")\n", " clean_checkpoints = list_checkpoints(CLEAN_OUTPUT)\n", " if not clean_checkpoints:\n", " raise RuntimeError(\"Pi0 protected clean40 训练没有产生父 checkpoint。\")\n", " clean_ckpt = clean_checkpoints[-1]\n", " old_override = os.environ.get(\"PI0_PRETRAINED_PATH_OVERRIDE\")\n", " old_mode = os.environ.get(\"NOTEBOOK_FRAME_WEIGHT_MODE\")\n", " old_blue = os.environ.get(\"NOTEBOOK_BLUE_WEIGHT\")\n", " os.environ[\"PI0_PRETRAINED_PATH_OVERRIDE\"] = str(clean_ckpt)\n", " os.environ[\"NOTEBOOK_FRAME_WEIGHT_MODE\"] = \"blue\"\n", " os.environ[\"NOTEBOOK_BLUE_WEIGHT\"] = os.environ.get(\"PI0_BLUE_WEIGHT\", \"2.0\")\n", " try:\n", " train_lerobot_config_in_notebook(blue_path, enabled=True, progress_name=\"Pi0 protected blue2x\")\n", " finally:\n", " if old_override is None:\n", " os.environ.pop(\"PI0_PRETRAINED_PATH_OVERRIDE\", None)\n", " else:\n", " os.environ[\"PI0_PRETRAINED_PATH_OVERRIDE\"] = old_override\n", " if old_mode is None:\n", " os.environ.pop(\"NOTEBOOK_FRAME_WEIGHT_MODE\", None)\n", " else:\n", " os.environ[\"NOTEBOOK_FRAME_WEIGHT_MODE\"] = old_mode\n", " if old_blue is None:\n", " os.environ.pop(\"NOTEBOOK_BLUE_WEIGHT\", None)\n", " else:\n", " os.environ[\"NOTEBOOK_BLUE_WEIGHT\"] = old_blue\n", " print(\"protected candidate checkpoints:\")\n", " protected_candidates = list_checkpoints(BLUE_OUTPUT)\n", " if not protected_candidates:\n", " raise RuntimeError(\"Pi0 protected 续训结束后没有 checkpoint，禁止进入评估。\")\n", " TRAINED_POLICY_PATH = protected_candidates[-1]\n", " TRAINING_SUMMARY_PATHS = [\n", " CLEAN_OUTPUT / \"training_run_summary.json\",\n", " BLUE_OUTPUT / \"training_run_summary.json\",\n", " ]\n", " print(\"本 Notebook protected checkpoint =\", public_path(TRAINED_POLICY_PATH))\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 4. 训练耗时与 checkpoint\n", "\n", "下面只读取本次训练写入的 `training_run_summary.json`，显示开始时间、结束时间、总耗时、步数和输出目录。\n" ] }, { "cell_type": "code", "metadata": {}, "source": [ "summary_paths = [Path(path) for path in globals().get(\"TRAINING_SUMMARY_PATHS\", [])]\n", "if not summary_paths:\n", " print(\"尚未执行 protected 训练，暂无训练耗时。\")\n", "else:\n", " summaries = []\n", " for summary_path in summary_paths:\n", " if summary_path.exists():\n", " summaries.append(json.loads(summary_path.read_text(encoding=\"utf-8\")))\n", " else:\n", " print(\"尚未生成：\", public_path(summary_path))\n", " if summaries:\n", " md_table(\n", " [\"recipe\", \"steps\", \"duration_min\", \"started_at_utc\", \"finished_at_utc\", \"output_dir\"],\n", " [(item.get(\"recipe\"), item.get(\"steps\"), f\"{float(item.get('duration_s', 0.0))/60.0:.2f}\", item.get(\"started_at_utc\"), item.get(\"finished_at_utc\"), item.get(\"output_dir\")) for item in summaries],\n", " )\n" ], "outputs": [], "execution_count": null }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 5. Pi0 原生 14 条评估协议\n", "\n", "Pi0 使用 EEF-absolute、median reducer 和物理成功判定。\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# PI0_NATIVE_FORMAL_EVAL_CELL\n", "import gc\n", "import random\n", "import sys\n", "import time\n", "from collections import deque\n", "\n", "import numpy as np\n", "import torch\n", "from PIL import Image\n", "from torchvision import transforms\n", "\n", "_PI0_NATIVE_TO_TENSOR = transforms.ToTensor()\n", "\n", "def _pi0_native_to_tensor_image(image):\n", " return _PI0_NATIVE_TO_TENSOR(Image.fromarray(image).resize((256, 256)))\n", "\n", "\n", "def _pi0_native_make_policy(policy_path, repo_id, dataset_root, device):\n", " from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata\n", " from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy\n", "\n", " metadata = LeRobotDatasetMetadata(repo_id, root=dataset_root)\n", " policy = PI0Policy.from_pretrained(policy_path, dataset_stats=metadata.stats)\n", " policy.to(device)\n", " policy.eval()\n", " return policy\n", "\n", "\n", "def _pi0_native_expected_state_dim(policy):\n", " feature = policy.config.input_features[\"observation.state\"]\n", " return int(tuple(feature.shape)[0])\n", "\n", "\n", "def _pi0_native_gripper(env):\n", " try:\n", " return float(np.clip(float(env.env.get_qpos_joint(\"rh_r1\")[0]), 0.0, 1.0))\n", " except Exception:\n", " return 0.0\n", "\n", "\n", "def _pi0_native_state(env, policy, action_step, hz=20.0):\n", " base = np.asarray(env.get_joint_state()[:6], dtype=np.float32).reshape(-1)\n", " expected = _pi0_native_expected_state_dim(policy)\n", " if expected == 6:\n", " return base\n", " if expected == 8:\n", " # Match the protected evaluator exactly: timestamp is the action\n", " # schedule time, not MuJoCo wall/simulation time after the 25 physics steps.\n", " timestamp = float(action_step) / float(hz)\n", " return np.concatenate(\n", " [base, np.asarray([_pi0_native_gripper(env), timestamp], dtype=np.float32)]\n", " )\n", " raise ValueError(\n", " f\"Pi0 checkpoint expects observation.state={expected}D; \"\n", " \"formal raw VLA protocol only supplies 6D or 8D state.\"\n", " )\n", "\n", "\n", "def _pi0_native_debug(env, tracker):\n", " target_body = env.obj_target\n", " target_pos = np.asarray(env.env.get_p_body(target_body), dtype=np.float32)\n", " plate_pos = np.asarray(env.env.get_p_body(\"body_obj_plate_11\"), dtype=np.float32)\n", " tcp_pos = np.asarray(env.env.get_p_body(\"tcp_link\"), dtype=np.float32)\n", " try:\n", " upright_cos = float(env.env.get_R_body(target_body)[2, 2])\n", " except Exception:\n", " upright_cos = float(\"nan\")\n", " terminal = bool(env.check_success())\n", " lifted = int(tracker[\"lifted_steps\"]) >= 3\n", " upright = bool(np.isfinite(upright_cos) and upright_cos >= 0.7)\n", " return {\n", " \"xy_dist\": float(np.linalg.norm(target_pos[:2] - plate_pos[:2])),\n", " \"target_pos\": [round(float(x), 4) for x in target_pos],\n", " \"plate_pos\": [round(float(x), 4) for x in plate_pos],\n", " \"tcp_pos\": [round(float(x), 4) for x in tcp_pos],\n", " \"gripper\": _pi0_native_gripper(env),\n", " \"success\": terminal,\n", " \"physical_success\": bool(terminal and lifted and upright),\n", " \"physical_lifted_enough\": bool(lifted),\n", " \"physical_final_upright\": bool(upright),\n", " \"physical_min_lift\": 0.03,\n", " \"physical_min_lift_steps\": 3,\n", " \"physical_final_upright_cos_threshold\": 0.7,\n", " \"target_body\": target_body,\n", " \"initial_target_pos\": tracker[\"initial_target_pos\"],\n", " \"initial_plate_pos\": tracker[\"initial_plate_pos\"],\n", " \"final_target_pos\": [round(float(x), 4) for x in target_pos],\n", " \"max_target_lift\": float(tracker[\"max_target_lift\"]),\n", " \"lifted_steps\": int(tracker[\"lifted_steps\"]),\n", " \"final_target_upright_cos\": upright_cos,\n", " }\n", "\n", "\n", "@torch.inference_mode()\n", "def _pi0_native_sample_chunk(policy, batch, samples=4, exec_chunk_steps=5):\n", " # This follows the configured Pi0 EEF-absolute sampling path,\n", " # while keeping the implementation visible in the Notebook.\n", " from lerobot.common.constants import OBS_STATE\n", "\n", " normalized = policy.normalize_inputs(batch)\n", " images, image_masks = policy.prepare_images(normalized)\n", " state = policy.prepare_state(normalized)\n", " language_tokens, language_masks = policy.prepare_language(normalized)\n", " chunks = []\n", " for _ in range(int(samples)):\n", " actions = policy.model.sample_actions(\n", " images, image_masks, language_tokens, language_masks, state, noise=None\n", " )\n", " action_dim = policy.config.action_feature.shape[0]\n", " actions = actions[:, :, :action_dim]\n", " actions = policy.unnormalize_outputs({\"action\": actions})[\"action\"]\n", " if getattr(policy.config, \"adapt_to_pi_aloha\", False):\n", " actions = policy._pi_aloha_encode_actions(actions)\n", " chunks.append(actions)\n", " selected = torch.stack(chunks, dim=0).median(dim=0).values\n", " # Keep a coherent short suffix from one predicted chunk before replanning.\n", " return deque(selected[0, : int(exec_chunk_steps)].unbind(0))\n", "\n", "\n", "def _pi0_native_configure_env(env):\n", " # The historical protected evaluator sets this reference pose once per episode.\n", " # Recomputing env.p0 on every action changes the MuJoCo eef_pose controller.\n", " from mujoco_env.transforms import rpy2r\n", "\n", " env.action_type = \"eef_pose\"\n", " env.p0, _ = env.env.get_pR_body(body_name=\"tcp_link\")\n", " env.R0 = rpy2r(np.deg2rad([90.0, 0.0, 90.0]))\n", "\n", "\n", "def _pi0_native_eef_abs_to_env(action, env, max_step=0.004):\n", " current_tcp = np.asarray(env.env.get_p_body(\"tcp_link\")[:3], dtype=np.float32)\n", " target_tcp = np.asarray(action[:3], dtype=np.float32)\n", " env_action = np.zeros(7, dtype=np.float32)\n", " env_action[:3] = np.clip(target_tcp - current_tcp, -float(max_step), float(max_step))\n", " env_action[6] = float(np.clip(action[6], 0.0, 1.0))\n", " return env_action\n", "\n", "\n", "def _pi0_native_close_env(env):\n", " try:\n", " env.env.close_viewer()\n", " except Exception:\n", " pass\n", " gc.collect()\n", "\n", "\n", "def run_pi0_eval_native_in_notebook(\n", " policy_path,\n", " result_path,\n", " seeds,\n", " repo_id,\n", " dataset_root,\n", " enabled=False,\n", " render=False,\n", " video_dir=None,\n", " trace_dir=None,\n", "):\n", " print(\"Pi0 formal checkpoint =\", public_path(policy_path))\n", " print(\"Pi0 formal seeds =\", list(seeds))\n", " print(\"Pi0 protocol = eef_abs / samples=4 / reducer=median / execute=5 / state=8D / physics=25\")\n", " if not enabled:\n", " print(\"未启动。设置 RUN_EVAL=1 后运行 Pi0 正式保护协议。\")\n", " return None\n", " if not ensure_project_layout():\n", " return None\n", "\n", " from contextlib import contextmanager\n", " import argparse\n", " media_module = load_eval_module()\n", "\n", " project_path = str(Path(PROJECT_ROOT).resolve())\n", " if project_path not in sys.path:\n", " sys.path.insert(0, project_path)\n", " from mujoco_env.y_env2 import SimpleEnv2\n", "\n", " @contextmanager\n", " def pushd(path):\n", " old = Path.cwd()\n", " os.chdir(path)\n", " try:\n", " yield\n", " finally:\n", " os.chdir(old)\n", "\n", " ensure_xvfb_display()\n", " result_path = Path(result_path)\n", " result_path.parent.mkdir(parents=True, exist_ok=True)\n", " existing_rows = []\n", " completed_seeds = set()\n", " if result_path.exists():\n", " # Resume after an interrupted long Notebook run without duplicating seeds.\n", " for line in result_path.read_text(encoding=\"utf-8\").splitlines():\n", " line = line.strip()\n", " if not line:\n", " continue\n", " row = json.loads(line)\n", " existing_rows.append(row)\n", " if row.get(\"seed\") is not None:\n", " completed_seeds.add(int(row[\"seed\"]))\n", " print(\"Pi0 resume rows =\", len(existing_rows), \"completed seeds =\", sorted(completed_seeds))\n", " run_seeds = [int(seed) for seed in seeds if int(seed) not in completed_seeds]\n", " if not run_seeds:\n", " print(\"Pi0 formal result already complete; no seed will be repeated.\")\n", "\n", " device = os.environ.get(\"EVAL_DEVICE\", \"cuda\")\n", " with pushd(PROJECT_ROOT):\n", " policy = _pi0_native_make_policy(Path(policy_path), repo_id, Path(dataset_root), device)\n", " expected_dim = _pi0_native_expected_state_dim(policy)\n", " print(\"Pi0 observation.state dimension =\", expected_dim)\n", " if expected_dim not in (6, 8):\n", " raise ValueError(f\"Unsupported Pi0 state dimension: {expected_dim}\")\n", "\n", " rows = list(existing_rows)\n", " media_args = argparse.Namespace(policy=\"pi0\", video_dir=Path(video_dir) if video_dir else None, trace_dir=Path(trace_dir) if trace_dir else None)\n", " selected_cases = set()\n", " for seed in notebook_progress(run_seeds, desc=\"pi0 formal eval\", total=len(run_seeds)):\n", " env = SimpleEnv2(\"./asset/example_scene_y2.xml\", action_type=\"joint_angle\")\n", " try:\n", " try:\n", " env.env.reset(step=False)\n", " except TypeError:\n", " env.env.reset()\n", " env.reset(seed=int(seed))\n", " random.seed(int(seed))\n", " np.random.seed(int(seed))\n", " torch.manual_seed(int(seed))\n", " if torch.cuda.is_available():\n", " torch.cuda.manual_seed_all(int(seed))\n", " _pi0_native_configure_env(env)\n", " policy.reset()\n", " tracker = {\n", " \"initial_target_pos\": [float(x) for x in env.env.get_p_body(env.obj_target)],\n", " \"initial_plate_pos\": [float(x) for x in env.env.get_p_body(\"body_obj_plate_11\")],\n", " \"initial_target_z\": float(env.env.get_p_body(env.obj_target)[2]),\n", " \"max_target_lift\": 0.0,\n", " \"lifted_steps\": 0,\n", " }\n", " action_queue = deque()\n", " action_steps = 0\n", " sim_steps = 0\n", " success_ever = False\n", " physical_success_ever = False\n", " first_success_step = None\n", " first_physical_success_step = None\n", " start = time.time()\n", " frames = [] if video_dir is not None else None\n", " trace = []\n", " while action_steps < int(os.environ.get(\"PI0_EVAL_MAX_ACTION_STEPS\", \"700\")) and env.env.is_viewer_alive():\n", " for _ in range(25):\n", " env.step_env()\n", " sim_steps += 1\n", " target_z = float(env.env.get_p_body(env.obj_target)[2])\n", " lift = target_z - tracker[\"initial_target_z\"]\n", " tracker[\"max_target_lift\"] = max(tracker[\"max_target_lift\"], lift)\n", " if lift >= 0.03:\n", " tracker[\"lifted_steps\"] += 1\n", " debug = _pi0_native_debug(env, tracker)\n", " if debug[\"success\"]:\n", " success_ever = True\n", " first_success_step = action_steps if first_success_step is None else first_success_step\n", " if debug[\"physical_success\"]:\n", " physical_success_ever = True\n", " first_physical_success_step = action_steps if first_physical_success_step is None else first_physical_success_step\n", " break\n", "\n", " state = _pi0_native_state(env, policy, action_steps)\n", " image, wrist_image = env.grab_image()\n", " media_module._capture_panel(frames, image, wrist_image)\n", " batch = {\n", " \"observation.state\": torch.tensor([state], dtype=torch.float32, device=device),\n", " \"observation.image\": _pi0_native_to_tensor_image(image).unsqueeze(0).to(device),\n", " \"observation.wrist_image\": _pi0_native_to_tensor_image(wrist_image).unsqueeze(0).to(device),\n", " \"task\": [env.instruction],\n", " }\n", " if not action_queue:\n", " action_queue = _pi0_native_sample_chunk(policy, batch, samples=4, exec_chunk_steps=5)\n", " action = action_queue.popleft().detach().float().cpu().numpy()\n", " env.step(_pi0_native_eef_abs_to_env(action, env))\n", " action_steps += 1\n", " trace.append({\"step\": int(action_steps), \"action\": [float(x) for x in action]})\n", " trace.append({\"step\": int(action_steps), \"action\": [float(x) for x in action]})\n", " trace.append({\"step\": int(action_steps), \"action\": [float(x) for x in action]})\n", " if render:\n", " env.render()\n", "\n", " debug = _pi0_native_debug(env, tracker)\n", " if debug[\"success\"]:\n", " success_ever = True\n", " first_success_step = action_steps if first_success_step is None else first_success_step\n", " if debug[\"physical_success\"]:\n", " physical_success_ever = True\n", " first_physical_success_step = action_steps if first_physical_success_step is None else first_physical_success_step\n", " break\n", "\n", " debug = _pi0_native_debug(env, tracker)\n", " row = {\n", " \"policy\": \"pi0\",\n", " \"seed\": int(seed),\n", " \"success\": bool(success_ever or debug[\"success\"]),\n", " \"physical_success\": bool(physical_success_ever or debug[\"physical_success\"]),\n", " \"action_steps\": int(action_steps),\n", " \"sim_steps\": int(sim_steps),\n", " \"elapsed_s\": round(time.time() - start, 3),\n", " \"instruction\": getattr(env, \"instruction\", None),\n", " \"debug\": debug,\n", " }\n", " row[\"_media\"] = media_module._write_episode_media(media_args, seed, frames, trace)\n", " row = media_module.finalize_case_media(media_args, row, selected_cases)\n", " row[\"_media\"] = media_module._write_episode_media(media_args, seed, frames, trace)\n", " row = media_module.finalize_case_media(media_args, row, selected_cases)\n", " row[\"_media\"] = media_module._write_episode_media(media_args, seed, frames, trace)\n", " row = media_module.finalize_case_media(media_args, row, selected_cases)\n", " rows.append(row)\n", " with result_path.open(\"a\", encoding=\"utf-8\") as f:\n", " f.write(json.dumps(row, ensure_ascii=False) + \"\\n\")\n", " print(json.dumps(row, ensure_ascii=False))\n", " finally:\n", " _pi0_native_close_env(env)\n", " if torch.cuda.is_available():\n", " torch.cuda.empty_cache()\n", "\n", " summarize_jsonl(result_path)\n", " success_count = sum(bool(row[\"success\"]) for row in rows)\n", " physical_count = sum(bool(row[\"physical_success\"]) for row in rows)\n", " print(f\"Pi0 formal result: success={success_count}/{len(rows)}, physical_success={physical_count}/{len(rows)}\")\n", " return rows\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# FULL_14_EVAL_WITH_MEDIA\n", "FULL_EVAL_EPISODES = 14\n", "FULL_EVAL_SEEDS = [3007, 3002, 3004, 3006, 3010, 3011, 3012, 3013, 3014, 3015, 3016, 3017, 3018, 3019]\n", "FULL_EVAL_ROOT = OUTPUT_ROOT / \"full_eval_14\"\n", "FULL_EVAL_RESULT = FULL_EVAL_ROOT / \"result.jsonl\"\n", "if not globals().get(\"RUN_EVAL\", False):\n", " print(\"未启动评估。请将 RUN_EVAL=True 后执行本单元。\")\n", "else:\n", " if globals().get(\"TRAINED_POLICY_PATH\") is None:\n", " raise RuntimeError(\"本 Notebook 没有当前训练 checkpoint；请先执行本 Notebook 的训练单元格。\")\n", " FULL_EVAL_POLICY = Path(TRAINED_POLICY_PATH)\n", " if not FULL_EVAL_POLICY.exists():\n", " raise FileNotFoundError(f\"当前训练 checkpoint 不存在：{public_path(FULL_EVAL_POLICY)}\")\n", " print(\"评估模型（来自本 Notebook 当前训练）：\", public_path(FULL_EVAL_POLICY))\n", " FULL_EVAL_ROWS = run_pi0_eval_native_in_notebook(\n", " FULL_EVAL_POLICY, FULL_EVAL_RESULT, FULL_EVAL_SEEDS,\n", " repo_id=DATASET_REPO_ID, dataset_root=TRAIN_DATA_ROOT, render=True, enabled=True,\n", " video_dir=FULL_EVAL_ROOT / \"videos\", trace_dir=FULL_EVAL_ROOT / \"traces\",\n", " )\n", "FULL_EVAL_ROWS = run_pi0_eval_native_in_notebook(\n", " FULL_EVAL_POLICY,\n", " FULL_EVAL_RESULT,\n", " FULL_EVAL_SEEDS,\n", " repo_id=DATASET_REPO_ID,\n", " dataset_root=TRAIN_DATA_ROOT,\n", " render=True,\n", " enabled=RUN_EVAL,\n", " video_dir=FULL_EVAL_ROOT / \"videos\",\n", " trace_dir=FULL_EVAL_ROOT / \"traces\",\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# FULL_14_MEDIA_REVIEW\n", "from IPython.display import Video\n", "import matplotlib.pyplot as plt\n", "\n", "def display_full_eval_media(result_path, title=\"pi0 - 14 episode evaluation\"):\n", " result_path = Path(result_path)\n", " if not result_path.exists():\n", " print(\"Evaluation has not run yet:\", public_path(result_path))\n", " return\n", " rows = [json.loads(line) for line in result_path.read_text(encoding=\"utf-8\").splitlines() if line.strip()]\n", " success = sum(bool(row.get(\"physical_success\", row.get(\"success\"))) for row in rows)\n", " print(f\"{title}: physical_success={success}/{len(rows)}\")\n", " md_table([[\"metric\", \"value\"]][0], [[\"physical success\", f\"{success}/{len(rows)}\"]])\n", " for kind in (\"success\", \"failure\"):\n", " row = next((item for item in rows if bool(item.get(\"physical_success\", item.get(\"success\"))) == (kind == \"success\") and item.get(\"video_path\")), None)\n", " if row is None:\n", " print(f\"No {kind} episode was found.\")\n", " continue\n", " print(f\"{kind} case: seed={row['seed']}\")\n", " video_path = Path(row[\"video_path\"])\n", " if video_path.exists():\n", " display(Video(filename=str(video_path), embed=False, width=960))\n", " trace_path = Path(row.get(\"trace_path\", \"\"))\n", " if not trace_path.exists():\n", " continue\n", " trace = json.loads(trace_path.read_text(encoding=\"utf-8\"))\n", " actions = np.asarray([item.get(\"action\", []) for item in trace], dtype=np.float32)\n", " if actions.ndim != 2 or not len(actions):\n", " continue\n", " fig, ax = plt.subplots(figsize=(12, 4))\n", " channels = min(actions.shape[1], 7)\n", " for channel in range(channels):\n", " label = \"gripper\" if channel == 6 else f\"action[{channel}]\"\n", " ax.plot(actions[:, channel], label=label, linewidth=1.2)\n", " ax.set_title(f\"{title} - {kind} sequence, seed={row['seed']}\")\n", " ax.set_xlabel(\"policy step\")\n", " ax.set_ylabel(\"normalized action / command\")\n", " ax.grid(alpha=0.25)\n", " ax.legend(ncol=4, fontsize=8)\n", " plt.tight_layout()\n", " display(fig)\n", " plt.close(fig)\n", "\n", "display_full_eval_media(FULL_EVAL_RESULT)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 6. 复现检查\n", "\n", "评估结果必须来自本 Notebook 当前训练生成的 `TRAINED_POLICY_PATH`。没有当前 checkpoint 时，评估单元会直接报错。\n" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3" } }, "nbformat": 4, "nbformat_minor": 5 }