{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": "# 14 SmolVLA 端到端：训练、评估、视频与诊断\n\n 这个 Notebook 把 SmolVLA 的完整学习路径放在一个文件里：先看已经复现成功的权重和视频，再检查数据，最后给出 smoke、长训、严格评估和日志追踪命令。\n\n 运行方式建议：第一次教学演示只开 `RUN_SMOKE=1 RUN_EVAL=1`；完整复现实验再开 `RUN_LONG_TRAIN=1`。长训输出会真实写回 Notebook 单元格，而不是粘贴静态日志。\n" }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": "TOPIC_ROOT = $TOPIC_ROOT\nPROJECT_ROOT = $PROJECT_ROOT\nDATA_ROOT = $DATA_ROOT\nMODEL_ROOT = $MODEL_ROOT\nOUTPUT_ROOT = $OUTPUT_ROOT\n" } ], "source": [ "from pathlib import Path\n", "import json\n", "import os\n", "import shlex\n", "import shutil\n", "import subprocess\n", "import sys\n", "\n", "try:\n", " from IPython.display import HTML, Markdown, display\n", "except Exception:\n", " class Markdown(str):\n", " pass\n", "\n", " class HTML(str):\n", " pass\n", "\n", " def display(obj):\n", " print(obj)\n", "\n", "\n", "def find_topic_root():\n", " cwd = Path.cwd().resolve()\n", " for candidate in [cwd, *cwd.parents]:\n", " if (candidate / \"assets\" / \"metrics_snapshot.json\").exists():\n", " return candidate\n", " raise RuntimeError(\"请从 AMD ROCm 专题目录或 notebooks 子目录启动 Jupyter。\")\n", "\n", "\n", "TOPIC_ROOT = find_topic_root()\n", "ASSET_DIR = TOPIC_ROOT / \"assets\"\n", "PROJECT_ROOT = Path(os.environ.get(\"PROJECT_ROOT\", \"/path/to/04mujoco复现ACT、Pi0、SmolVLA\"))\n", "DATA_ROOT = Path(os.environ.get(\"DATA_ROOT\", \"/path/to/datasets/every_embodied\"))\n", "MODEL_ROOT = Path(os.environ.get(\"MODEL_ROOT\", \"/path/to/model/checkpoints\"))\n", "OUTPUT_ROOT = Path(os.environ.get(\"OUTPUT_ROOT\", TOPIC_ROOT / \"outputs\"))\n", "\n", "# The AMD teaching workflow should be runnable from local datasets/checkpoints.\n", "# Avoid surprising network calls during class or when AUP/Radeon Cloud cannot\n", "# reach Hugging Face.\n", "os.environ.setdefault(\"HF_HUB_OFFLINE\", os.environ.get(\"NOTEBOOK_HF_OFFLINE\", \"1\"))\n", "os.environ.setdefault(\"TRANSFORMERS_OFFLINE\", os.environ.get(\"NOTEBOOK_HF_OFFLINE\", \"1\"))\n", "os.environ.setdefault(\"HF_DATASETS_OFFLINE\", os.environ.get(\"NOTEBOOK_HF_OFFLINE\", \"1\"))\n", "os.environ.setdefault(\"HF_HOME\", str(Path(os.environ.get(\"CACHE_ROOT\", OUTPUT_ROOT / \"cache\")) / \"huggingface\"))\n", "os.environ.setdefault(\"HF_DATASETS_CACHE\", str(Path(os.environ[\"HF_HOME\"]) / \"datasets\"))\n", "\n", "def public_path(path):\n", " path = Path(path)\n", " replacements = [\n", " (TOPIC_ROOT, \"$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", "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 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", " 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", " if cfg.policy.type == \"smolvla\":\n", " local_vlm_model = os.environ.get(\"SMOLVLA_VLM_MODEL_PATH\")\n", " if local_vlm_model and Path(local_vlm_model).exists():\n", " cfg.policy.vlm_model_name = str(Path(local_vlm_model))\n", " cfg.policy.load_vlm_weights = False\n", " print(\"local smolvlm processor/config =\", public_path(cfg.policy.vlm_model_name))\n", "\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", " print(\"训练完成。metrics =\", public_path(metrics_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, \"last_metrics\": last_metrics}\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", "):\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", " 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", " 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 = args.seed_start + offset\n", " row = rollout(args, policy, seed)\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 resolve_eval_policy(default_path, trained_run_dir=None, env_name=None):\n", " if env_name and os.environ.get(env_name):\n", " path = Path(os.environ[env_name])\n", " print(\"评估使用环境变量指定权重：\", public_path(path))\n", " return path\n", " if trained_run_dir is not None and env_flag(\"EVAL_USE_LONG_TRAIN\"):\n", " checkpoints = list_checkpoints(trained_run_dir)\n", " if checkpoints:\n", " path = checkpoints[-1]\n", " print(\"评估使用本次长训最新 checkpoint：\", public_path(path))\n", " return path\n", " print(\"未找到本次长训 checkpoint，回退到保护权重。\")\n", " path = Path(default_path)\n", " print(\"评估使用保护/预训练权重：\", 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": "## 运行控制台：先在这里选择本次实验\n\n下面这个单元格是三个模型通用的开关。默认全部关闭，避免打开 Notebook 后误启动数小时训练。\n\n建议顺序：\n\n1. 第一次把 `RUN_SMOKE` 设为 `True`，确认环境、数据、模型和一次反向传播都正常。\n2. 确认无误后把 `RUN_LONG_TRAIN` 设为 `True`，执行后面的真实训练单元格，Notebook 会显示实时 `tqdm`、loss、耗时和 checkpoint。\n3. 训练完成后把 `RUN_EVAL` 设为 `True`，执行严格评估单元格，结果会汇总为 `x/y` 成功率，并在有渲染条件时保存视频。\n4. 如果要复现正式保护结果，再把 `RUN_PROTECTED_TRAIN` 设为 `True`，执行对应 protected recipe 单元格。它和普通教学长训是两条明确标注的训练谱系。\n\n训练循环和评估循环都在 Notebook Python kernel 内执行，不是 `cat` 静态日志，也不是把训练交给外部 shell 脚本。\n" }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# RUN_CONTROL_CELL\n", "# 只修改下面四个布尔值，然后按顺序执行后面的单元格。\n", "RUN_SMOKE = False\n", "RUN_LONG_TRAIN = True\n", "RUN_EVAL = False\n", "RUN_PROTECTED_TRAIN = False\n", "\n", "print({\n", " 'RUN_SMOKE': RUN_SMOKE,\n", " 'RUN_LONG_TRAIN': RUN_LONG_TRAIN,\n", " 'RUN_EVAL': RUN_EVAL,\n", " 'RUN_PROTECTED_TRAIN': RUN_PROTECTED_TRAIN,\n", "})\n", "print('长训通常需要几十分钟到数小时；评估会逐 episode 闭环运行，时间更长。')\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 1：先确认这一版为什么作为主线\n" }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 结果对照\n", "\n", "| 项目 | 结果 | 说明 |\n", "| --- | --- | --- |\n", "| 历史教程记录 | 53/60 | SmolVLA weighted step500，红蓝杯平衡较好 |\n", "| 当前重建结果 | 57/60 | red 27/30，blue 30/30 |\n", "| 发布建议 | 主推权重 | 适合作为零训练预览和默认 Notebook 案例 |\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 2：显示严格成功与失败视频\n" }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 1.5：保护权重的训练配方，不和课堂轻量训练混用\n\n 上一个单元给出的是已经保护的 `57/60` 结果。这里说明它是怎么训练出来的。课堂默认长训只是让学习者体验完整流程；要复现保护权重，需要切到下面的 protected recipe，并跑完整训练与 60 episode strict eval。\n" }, { "cell_type": "markdown", "metadata": {}, "source": "### 结果口径对齐：本轮小面板与正式保护评估\n\n Notebook 后面的 eval 单元会跑一个便宜的小面板，用来确认本轮 checkpoint 能闭环执行并产出视频；正式保护评估使用更大的固定面板，才是发布权重和教程报告采用的口径。\n\n | 口径 | 成功率 | 评估范围 | 教学解释 |\n | --- | --- | --- | --- |\n | 本轮 Notebook 小面板 | `3/4` | post-long eval seed0-3 | 用于课堂展示和视频验收，不替代正式分数 |\n | 正式保护评估 | `57/60` | red30 + blue30 strict physical success | 红 `27/30`、蓝 `30/30`，作为默认发布权重 |\n" }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": "{\n \"TEACHING_RECIPE\": \"protected\",\n \"TRAIN_DATA_ROOT\": \"$TRAIN_DATA_ROOT\",\n \"SMOLVLA_STEPS\": \"5000 + weighted-blue continuation\",\n \"SMOLVLA_EVAL_EPISODES\": \"60\",\n \"SMOLVLA_POLICY_PATH\": \"$MODEL_ROOT/smolvla_weighted_000500/pretrained_model\"\n}\n注意：weighted-blue 采样逻辑必须和 README_04/README_06 中的 Weighted sampler 一致；普通课堂长训不能冒充 protected 结果。\n" }, { "output_type": "display_data", "data": { "text/markdown": "| 模式 | 数据 | 采样/数据策略 | 训练步数 | 评估面板 | 教学解释 |\n| --- | --- | --- | --- | --- | --- |\n| 教学默认 | demo_data_language | 普通 EpisodeAwareSampler | SMOLVLA_STEPS=5000 | 本轮小面板 3/4 | 用于课堂跑通和视频展示，不保证得到 57/60 |\n| 保护配方 parent | demo_data_language | 基础 SmolVLA 长训 | 5000 steps | 作为加权续训父权重 | 先得到可用 parent checkpoint |\n| 保护配方 weighted-blue | demo_data_language | 蓝杯 frame/episode 加权，不复制原始 parquet | 续训 1000 steps，选择 step500 | red30 + blue30 strict | 当前重建 57/60；红 27/30，蓝 30/30 |" }, "metadata": {} } ], "source": [ "# PROTECTED_RECIPE_CELL\n", "rows = [\n", " (\"教学默认\", \"demo_data_language\", \"普通 EpisodeAwareSampler\", \"SMOLVLA_STEPS=5000\", \"本轮小面板 3/4\", \"用于课堂跑通和视频展示，不保证得到 57/60\"),\n", " (\"保护配方 parent\", \"demo_data_language\", \"基础 SmolVLA 长训\", \"5000 steps\", \"作为加权续训父权重\", \"先得到可用 parent checkpoint\"),\n", " (\"保护配方 weighted-blue\", \"demo_data_language\", \"蓝杯 frame/episode 加权，不复制原始 parquet\", \"续训 1000 steps，选择 step500\", \"red30 + blue30 strict\", \"当前重建 57/60；红 27/30，蓝 30/30\"),\n", "]\n", "md_table([\"模式\", \"数据\", \"采样/数据策略\", \"训练步数\", \"评估面板\", \"教学解释\"], rows)\n", "\n", "protected_env = {\n", " \"TEACHING_RECIPE\": \"protected\",\n", " \"TRAIN_DATA_ROOT\": str(DATA_ROOT / \"demo_data_language\"),\n", " \"SMOLVLA_STEPS\": \"5000 + weighted-blue continuation\",\n", " \"SMOLVLA_EVAL_EPISODES\": \"60\",\n", " \"SMOLVLA_POLICY_PATH\": str(MODEL_ROOT / \"smolvla_weighted_000500\" / \"pretrained_model\"),\n", "}\n", "print(json.dumps(protected_env, ensure_ascii=False, indent=2))\n", "print(\"注意：weighted-blue 采样逻辑必须和 README_04/README_06 中的 Weighted sampler 一致；普通课堂长训不能冒充 protected 结果。\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "### 可执行 protected 训练：parent 5000 + weighted-blue 续训\n\n 设置 `RUN_PROTECTED_TRAIN=1` 后，这一格会在 Notebook 内部真实训练两个阶段：先训练 parent，再把 parent checkpoint 作为初始化继续做 blue 加权续训。默认关闭是为了避免课堂一打开就长训。\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 SmolVLA protected recipe。\")\n", "else:\n", " DATASET_REPO_ID = globals().get(\"DATASET_REPO_ID\", os.environ.get(\"DATASET_REPO_ID\", \"datawhale_eai_pnp_language\"))\n", " TRAIN_DATA_ROOT = globals().get(\"TRAIN_DATA_ROOT\", Path(os.environ.get(\"TRAIN_DATA_ROOT\", DATA_ROOT / \"demo_data_language\")))\n", " CONFIG_DIR = OUTPUT_ROOT / \"configs\"\n", " RUN_ROOT = OUTPUT_ROOT / \"runs\" / \"smolvla_protected_recipe\"\n", " PARENT_OUTPUT = RUN_ROOT / \"parent_5000\"\n", " WEIGHTED_OUTPUT = RUN_ROOT / \"weighted_blue2_step1000\"\n", " parent_config = make_lerobot_train_config(\n", " \"smolvla\", DATASET_REPO_ID, TRAIN_DATA_ROOT, PARENT_OUTPUT,\n", " steps=int(os.environ.get(\"SMOLVLA_PARENT_STEPS\", \"5000\")),\n", " batch_size=int(os.environ.get(\"SMOLVLA_BATCH_SIZE\", \"4\")),\n", " chunk_size=50,\n", " n_action_steps=50,\n", " )\n", " weighted_config = make_lerobot_train_config(\n", " \"smolvla\", DATASET_REPO_ID, TRAIN_DATA_ROOT, WEIGHTED_OUTPUT,\n", " steps=int(os.environ.get(\"SMOLVLA_WEIGHTED_STEPS\", \"1000\")),\n", " batch_size=int(os.environ.get(\"SMOLVLA_BATCH_SIZE\", \"4\")),\n", " chunk_size=50,\n", " n_action_steps=50,\n", " )\n", " parent_path = write_json_yaml(CONFIG_DIR / \"smolvla_protected_parent_5000.yaml\", parent_config)\n", " weighted_path = write_json_yaml(CONFIG_DIR / \"smolvla_protected_weighted_blue2.yaml\", weighted_config)\n", " train_lerobot_config_in_notebook(parent_path, enabled=True, progress_name=\"SmolVLA protected parent\")\n", " parent_ckpt = list_checkpoints(PARENT_OUTPUT)[-1]\n", " old_override = os.environ.get(\"SMOLVLA_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[\"SMOLVLA_PRETRAINED_PATH_OVERRIDE\"] = str(parent_ckpt)\n", " os.environ[\"NOTEBOOK_FRAME_WEIGHT_MODE\"] = \"blue\"\n", " os.environ[\"NOTEBOOK_BLUE_WEIGHT\"] = os.environ.get(\"SMOLVLA_BLUE_WEIGHT\", \"2.0\")\n", " try:\n", " train_lerobot_config_in_notebook(weighted_path, enabled=True, progress_name=\"SmolVLA protected weighted-blue\")\n", " finally:\n", " if old_override is None:\n", " os.environ.pop(\"SMOLVLA_PRETRAINED_PATH_OVERRIDE\", None)\n", " else:\n", " os.environ[\"SMOLVLA_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", " list_checkpoints(WEIGHTED_OUTPUT)\n" ] }, { "cell_type": "code", "execution_count": 6, "metadata": {}, "outputs": [ { "output_type": "display_data", "data": { "text/markdown": "**红杯成功回放：weighted500 seed0**" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/html": "

" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/markdown": "**蓝杯成功回放：weighted500 seed0**" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/html": "

" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/markdown": "**红杯失败回放：用于观察 upright/release 问题**" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/html": "

" }, "metadata": {} } ], "source": [ "show_video(\"smolvla_weighted500_red_success_seed0.mp4\", \"红杯成功回放：weighted500 seed0\")\n", "show_video(\"smolvla_weighted500_blue_success_seed0.mp4\", \"蓝杯成功回放：weighted500 seed0\")\n", "show_video(\"smolvla_weighted500_red_failure_seed8.mp4\", \"红杯失败回放：用于观察 upright/release 问题\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 3：检查数据和权重路径\n\n 预计耗时：几秒。这里不会训练，只确认数据、权重和输出目录是否指向正确位置。\n" }, { "cell_type": "code", "execution_count": 7, "metadata": {}, "outputs": [ { "output_type": "display_data", "data": { "text/markdown": "| 变量 | 当前值 |\n| --- | --- |\n| DATASET_REPO_ID | datawhale_eai_pnp_language |\n| TRAIN_DATA_ROOT | $DATA_ROOT/demo_data_language |\n| SMOLVLA_POLICY_PATH | $OUTPUT_ROOT/runs/smolvla_protected_recipe/weighted_blue2_step1000/checkpoints/001000/pretrained_model |\n| 数据 meta | $DATA_ROOT/demo_data_language/meta/info.json |" }, "metadata": {} } ], "source": [ "DATASET_REPO_ID = os.environ.get(\"DATASET_REPO_ID\", \"datawhale_eai_pnp_language\")\n", "TRAIN_DATA_ROOT = Path(os.environ.get(\"TRAIN_DATA_ROOT\", DATA_ROOT / \"demo_data_language\"))\n", "PRETRAINED_POLICY = Path(os.environ.get(\"SMOLVLA_POLICY_PATH\", MODEL_ROOT / \"smolvla_weighted_000500\" / \"pretrained_model\"))\n", "\n", "rows = [\n", " (\"DATASET_REPO_ID\", DATASET_REPO_ID),\n", " (\"TRAIN_DATA_ROOT\", TRAIN_DATA_ROOT),\n", " (\"SMOLVLA_POLICY_PATH\", PRETRAINED_POLICY),\n", " (\"数据 meta\", TRAIN_DATA_ROOT / \"meta\" / \"info.json\"),\n", "]\n", "md_table([\"变量\", \"当前值\"], rows)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 4：生成配置并真实启动 smoke / 长训\n\n 预计耗时：smoke 约 1-5 分钟；`SMOLVLA_STEPS=5000` 的长训通常需要几十分钟到数小时，取决于 ROCm、batch size 和数据盘速度。 \n 这一格是真实训练入口：设置 `RUN_SMOKE=1` 或 `RUN_LONG_TRAIN=1` 后执行，会在 Notebook kernel 内直接创建 dataset、policy、optimizer 和训练循环，并显示 tqdm 进度。\n" }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "CONFIG_DIR = OUTPUT_ROOT / \"configs\"\n", "LOG_DIR = OUTPUT_ROOT / \"logs\"\n", "RUN_ROOT = OUTPUT_ROOT / \"runs\" / \"smolvla_weighted_repro\"\n", "SMOKE_OUTPUT = RUN_ROOT / \"smoke\"\n", "LONG_OUTPUT = RUN_ROOT / \"weighted_full\"\n", "\n", "smoke_config = make_lerobot_train_config(\n", " \"smolvla\", DATASET_REPO_ID, TRAIN_DATA_ROOT, SMOKE_OUTPUT,\n", " steps=2, batch_size=2, chunk_size=50, n_action_steps=50,\n", ")\n", "long_config = make_lerobot_train_config(\n", " \"smolvla\", DATASET_REPO_ID, TRAIN_DATA_ROOT, LONG_OUTPUT,\n", " steps=int(os.environ.get(\"SMOLVLA_STEPS\", \"5000\")),\n", " batch_size=int(os.environ.get(\"SMOLVLA_BATCH_SIZE\", \"4\")),\n", " chunk_size=50,\n", " n_action_steps=50,\n", ")\n", "smoke_config_path = write_json_yaml(CONFIG_DIR / \"smolvla_smoke.yaml\", smoke_config)\n", "long_config_path = write_json_yaml(CONFIG_DIR / \"smolvla_weighted_full.yaml\", long_config)\n", "\n", "train_lerobot_config_in_notebook(smoke_config_path, enabled=RUN_SMOKE, progress_name=\"SmolVLA smoke\")\n", "train_lerobot_config_in_notebook(long_config_path, enabled=RUN_LONG_TRAIN, progress_name=\"SmolVLA long train\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 5：实时查看训练日志和 checkpoint\n\n 预计耗时：几秒。长训进行中可以反复执行本格，查看 Notebook 训练写出的 metrics 和已经落盘的 checkpoint。\n" }, { "cell_type": "code", "execution_count": 9, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": "smoke metrics:\n日志不存在： $OUTPUT_ROOT/runs/smolvla_weighted_repro/smoke/notebook_train_metrics.jsonl\n\nlong train metrics:\n日志不存在： $OUTPUT_ROOT/runs/smolvla_weighted_repro/weighted_full/notebook_train_metrics.jsonl\n\ncheckpoints:\n尚未发现 checkpoint： $OUTPUT_ROOT/runs/smolvla_weighted_repro/weighted_full\n" } ], "source": [ "print(\"smoke metrics:\")\n", "tail_log(SMOKE_OUTPUT / \"notebook_train_metrics.jsonl\", lines=20)\n", "print(\"\\nlong train metrics:\")\n", "tail_log(LONG_OUTPUT / \"notebook_train_metrics.jsonl\", lines=40)\n", "print(\"\\ncheckpoints:\")\n", "list_checkpoints(LONG_OUTPUT)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 6：已完成长训的实测对照\n\n 这是我们已经在 AMD 设备上跑完并复核过的视频/指标对照，用来帮助学习者先看到“正确跑起来是什么样”。它不替代本次 Notebook 的训练输出。\n" }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 结果对照\n", "\n", "| 阶段 | 训练/评估进度 | 结论 |\n", "| --- | --- | --- |\n", "| parent | 5000/5000 steps | 基础 SmolVLA 收敛到可用 checkpoint |\n", "| weighted-blue | selected step 500/1000 | step500 比 step1000 更平衡 |\n", "| strict gate | 57/60 | legacy 60/60，physical 57/60 |\n" ] }, { "cell_type": "code", "execution_count": 11, "metadata": {}, "outputs": [ { "output_type": "display_data", "data": { "text/markdown": "**历史训练进度与闭环结果**" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/html": "

" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/markdown": "**红杯/蓝杯分指令对比**" }, "metadata": {} }, { "output_type": "display_data", "data": { "text/html": "

" }, "metadata": {} } ], "source": [ "show_image(\"training_progress_overview.png\", \"历史训练进度与闭环结果\")\n", "show_image(\"smolvla_red_blue_success.png\", \"红杯/蓝杯分指令对比\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 7：Notebook 内严格评估\n\n 预计耗时：10 个 episode 通常 10-30 分钟；`SMOLVLA_EVAL_EPISODES=60` 会更久。 \n 本单元会在 Notebook kernel 内加载策略并逐个 seed 闭环 rollout；如无显示器，会自动尝试启动 Xvfb。\n" }, { "cell_type": "code", "execution_count": 12, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": "评估使用环境变量指定权重： $OUTPUT_ROOT/runs/smolvla_protected_recipe/weighted_blue2_step1000/checkpoints/001000/pretrained_model\npolicy = smolvla\npolicy_path = $OUTPUT_ROOT/runs/smolvla_protected_recipe/weighted_blue2_step1000/checkpoints/001000/pretrained_model\nresult = $OUTPUT_ROOT/smolvla_eval_seed0_59.jsonl\n没有发现 Xvfb；如遇 GLFW DISPLAY 报错，请先安装 xvfb。\n" }, { "ename": "FileNotFoundError", "evalue": "评估脚本不存在：$PROJECT_ROOT/eval_policy_success.py", "output_type": "error", "traceback": [ "Traceback (most recent call last):", " File \"/workspace/every-embodied-amd-topic/code/execute_tutorial_notebooks.py\", line 175, in execute_notebook", " exec(compile(source, f\"{path.name}:cell-{cell_index}\", \"exec\"), namespace)", " File \"14_smolvla_end_to_end.ipynb:cell-22\", line 3, in ", " File \"14_smolvla_end_to_end.ipynb:cell-2\", line 468, in run_eval_policy_in_notebook", " File \"14_smolvla_end_to_end.ipynb:cell-2\", line 426, in load_eval_module", "FileNotFoundError: 评估脚本不存在：$PROJECT_ROOT/eval_policy_success.py" ] } ], "source": [ "eval_episodes = os.environ.get(\"SMOLVLA_EVAL_EPISODES\", \"10\")\n", "eval_policy = resolve_eval_policy(PRETRAINED_POLICY, LONG_OUTPUT, \"SMOLVLA_EVAL_POLICY_PATH\")\n", "run_eval_policy_in_notebook(\n", " \"smolvla\",\n", " eval_policy,\n", " OUTPUT_ROOT / f\"smolvla_eval_seed0_{int(eval_episodes)-1}.jsonl\",\n", " episodes=eval_episodes,\n", " seed_start=0,\n", " render=env_flag(\"RENDER_EVAL\"),\n", " enabled=RUN_EVAL,\n", " repo_id=DATASET_REPO_ID,\n", " dataset_root=TRAIN_DATA_ROOT,\n", ")\n", "summarize_jsonl(OUTPUT_ROOT / f\"smolvla_eval_seed0_{int(eval_episodes)-1}.jsonl\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": "## Checkpoint 8：怎么读结果\n" }, { "cell_type": "code", "execution_count": 13, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": "{\n \"candidate\": \"weighted_000500\",\n \"episodes\": 60,\n \"physical_success_count\": 57,\n \"legacy_success_count\": 60,\n \"by_color\": {\n \"red\": \"27/30\",\n \"blue\": \"30/30\"\n },\n \"release_decision\": \"作为教程默认预训练权重\"\n}\n" } ], "source": [ "summary = {\n", " \"candidate\": \"weighted_000500\",\n", " \"episodes\": 60,\n", " \"physical_success_count\": 57,\n", " \"legacy_success_count\": 60,\n", " \"by_color\": {\"red\": \"27/30\", \"blue\": \"30/30\"},\n", " \"release_decision\": \"作为教程默认预训练权重\",\n", "}\n", "print(json.dumps(summary, ensure_ascii=False, indent=2))\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" }, "amd_rocm_tutorial_execution": { "executor": "code/execute_tutorial_notebooks.py", "python": "3.12.3", "code_cells_executed": 13, "errors": 1, "record_markers": [], "prepare_markers": [], "long_tasks_enabled": false } }, "nbformat": 4, "nbformat_minor": 5 }