scripts/patch_openpi_config.py

"""Apply the G1 Dex3 pick-and-place configs to an openpi checkout, idempotently.

    uv run python scripts/patch_openpi_config.py \
        --openpi /home/ubuntu/openpi --repo-id <YOUR_HF_USER>/<DATASET_NAME>

Replaces the manual copy-paste of `openpi_config_block.py`. Does three things:

1. copies `g1_dex3_policy.py` into `src/openpi/policies/`
2. adds the two imports to `src/openpi/training/config.py`
3. inserts `LeRobotG1Dex3DataConfig` and the two `TrainConfig` entries

Safe to rerun: each edit is skipped if already present. `config.py` is backed up once to
`config.py.orig` before the first modification.
"""
import argparse
import ast
import os
import pathlib
import shutil
import subprocess
import sys

PATCH_VERSION = 3
MARKER = "# G1DEX3-PATCH-VERSION:"

IMPORTS = """import openpi.shared.nnx_utils as nnx_utils
from openpi.policies import g1_dex3_policy
"""

DATA_CONFIG = '''
@dataclasses.dataclass(frozen=True)
class LeRobotG1Dex3DataConfig(DataConfigFactory):
    """G1 + Dex3 upper-body pick-and-place. 43-dim whole-body -> 28 trained dims."""

    @override
    def create(self, assets_dirs: pathlib.Path, model_config: _model.BaseModelConfig) -> DataConfig:
        repack = _transforms.Group(
            inputs=[
                _transforms.RepackTransform(
                    {
                        "images": {
                            "base_0_rgb": "observation.images.ego_view",
                            "left_wrist_0_rgb": "observation.images.ego_left",
                            "right_wrist_0_rgb": "observation.images.ego_right",
                        },
                        "state": "observation.state",
                        "actions": "action",
                        "prompt": "prompt",
                    }
                )
            ]
        )

        data_transforms = _transforms.Group(
            inputs=[g1_dex3_policy.G1Dex3Inputs(action_dim=model_config.action_dim)],
            outputs=[g1_dex3_policy.G1Dex3Outputs()],
        )

        # The 28 upper-body dims are ordered [L_arm 7, L_hand 7, R_arm 7, R_hand 7].
        # Delta on the arms, absolute on the hands -> interleaved. make_bool_mask(14, -14)
        # would leave the RIGHT arm (the one doing the task) on absolute actions.
        delta_mask = _transforms.make_bool_mask(7, -7, 7, -7)
        data_transforms = data_transforms.push(
            inputs=[_transforms.DeltaActions(delta_mask)],
            outputs=[_transforms.AbsoluteActions(delta_mask)],
        )

        return dataclasses.replace(
            self.create_base_config(assets_dirs, model_config),
            repack_transforms=repack,
            data_transforms=data_transforms,
            model_transforms=ModelTransformFactory()(model_config),
            # This dataset's action column is `action` (singular). action_sequence_keys is fed
            # straight into LeRobotDataset's delta_timestamps -- i.e. it queries the raw hf_dataset
            # column BEFORE the repack renames action -> actions. The DataConfig default is
            # ("actions",), which raises
            #   KeyError: Column actions not in the dataset
            # deep inside the dataloader worker. LeRobotAlohaDataConfig overrides it the same way.
            action_sequence_keys=("action",),
        )

'''

TRAIN_CONFIGS = '''    #
    # G1 Dex3 pick-and-place (pi0.5).
    #
    TrainConfig(
        name="pi05_g1_pickplace",
        # max_token_len is deliberately unset. For pi05, discrete_state_input defaults to True
        # and embed_suffix() skips the continuous state token, so the ONLY proprioception path
        # is the discretised state string in the prompt -- measured 120-142 tokens here. The pi0
        # default of 48 would truncate after 9 of 32 state values, keeping the left arm and
        # dropping the entire right arm and hand. Unset => the pi05 default of 200.
        model=pi0_config.Pi0Config(pi05=True, action_dim=32, action_horizon=50),
        data=LeRobotG1Dex3DataConfig(
            repo_id="{REPO_ID}",
            base_config=DataConfig(prompt_from_task=True),
        ),
        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
        lr_schedule=_optimizer.CosineDecaySchedule(
            warmup_steps=1_000, peak_lr=5e-5, decay_steps=1_000_000, decay_lr=5e-5
        ),
        optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
        ema_decay=0.999,
        wandb_enabled=False,
        # openpi defaults num_workers to 2. Every sample decodes 3 random-access h264 frames, so
        # 2 workers deliver a batch of 32 roughly every 6.7 s -- against ~1.5 s of A100 compute,
        # that leaves the GPU ~78% idle and turns a 12 h run into 56 h. Also throttles
        # compute_norm_stats.py, which decodes all three videos per sample only to discard them.
        num_workers={NUM_WORKERS},
        num_train_steps=30_000,
        batch_size=32,
        save_interval=2_500,
        keep_period=2_500,
        log_interval=50,
        seed=0,
    ),
    # GR00T-matched: trains the action expert + all projections, freezes the PaliGemma language
    # model AND the SigLIP vision tower. get_freeze_filter() only ever freezes ".*llm.*", and the
    # tree is PaliGemma = nnx.Dict(llm=..., img=...), so a "gemma_2b_lora" variant would leave the
    # vision tower fully trainable -- hence the explicit filter.
    TrainConfig(
        name="pi05_g1_pickplace_frozen_vlm",
        model=pi0_config.Pi0Config(pi05=True, action_dim=32, action_horizon=50),
        data=LeRobotG1Dex3DataConfig(
            repo_id="{REPO_ID}",
            base_config=DataConfig(prompt_from_task=True),
        ),
        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
        freeze_filter=nnx.All(
            nnx_utils.PathRegex(".*(llm|img).*"),
            nnx.Not(nnx_utils.PathRegex(".*llm.*_1.*")),
        ),
        ema_decay=None,
        lr_schedule=_optimizer.CosineDecaySchedule(
            warmup_steps=1_000, peak_lr=5e-5, decay_steps=1_000_000, decay_lr=5e-5
        ),
        optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
        wandb_enabled=False,
        num_workers={NUM_WORKERS},
        num_train_steps=30_000,
        batch_size=32,
        save_interval=2_500,
        keep_period=2_500,
        log_interval=50,
        seed=0,
    ),
'''


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--openpi", required=True, help="path to the openpi checkout")
    ap.add_argument("--repo-id", required=True, help="HF dataset repo id, e.g. simpk/single_pickplace")
    ap.add_argument("--num-workers", type=int, default=None,
                    help="dataloader workers. Default: max(2, nproc - 2). openpi's default of 2 "
                         "starves the GPU -- each sample needs 3 random-access h264 decodes.")
    ap.add_argument("--dry-run", action="store_true")
    ap.add_argument("--no-verify", action="store_true",
                    help="skip the get_config() check (use if the openpi env is not synced yet)")
    a = ap.parse_args()

    if a.num_workers is None:
        a.num_workers = max(2, (os.cpu_count() or 4) - 2)

    here = pathlib.Path(__file__).resolve().parent.parent
    openpi = pathlib.Path(a.openpi).expanduser().resolve()
    cfg = openpi / "src/openpi/training/config.py"
    pol_dst = openpi / "src/openpi/policies/g1_dex3_policy.py"
    pol_src = here / "g1_dex3_policy.py"

    for p, what in ((cfg, "config.py"), (pol_src, "g1_dex3_policy.py")):
        if not p.exists():
            print(f"FAIL: {what} not found at {p}")
            return 1

    src = cfg.read_text()
    backup = cfg.with_suffix(".py.orig")
    actions = []

    # If an older revision of this patch is already in place, revert to the pristine backup and
    # re-apply from scratch. Editing in place would need a diff for every past version.
    found = None
    for line in src.splitlines():
        if line.startswith(MARKER):
            found = int(line.split(":")[1].strip())
            break
    if found is None and "class LeRobotG1Dex3DataConfig" in src:
        found = 1  # pre-marker revision of this patch
    if found is not None and found < PATCH_VERSION:
        if not backup.exists():
            print(f"FAIL: config.py carries patch v{found} but {backup.name} is missing.")
            print("      Restore config.py from git (`git checkout -- src/openpi/training/config.py`) and rerun.")
            return 1
        print(f"found patch v{found}, upgrading to v{PATCH_VERSION}: reverting to {backup.name} first")
        src = backup.read_text()
        actions.append(f"revert to pristine config.py and re-apply at v{PATCH_VERSION}")
    elif found == PATCH_VERSION:
        pass  # up to date; the checks below will all no-op

    # 1. the policy module
    need_policy = (not pol_dst.exists()) or pol_dst.read_text() != pol_src.read_text()
    if need_policy:
        actions.append(f"copy g1_dex3_policy.py -> {pol_dst.relative_to(openpi)}")

    # 2. imports, appended to the existing openpi import block
    new = src
    if "from openpi.policies import g1_dex3_policy" not in new:
        anchor = "import openpi.transforms as _transforms\n"
        if anchor not in new:
            print("FAIL: could not find the import anchor in config.py")
            return 1
        new = new.replace(anchor, anchor + f"{MARKER} {PATCH_VERSION}\n" + IMPORTS, 1)
        actions.append("add imports (nnx_utils, g1_dex3_policy)")

    # 3. the data config class, immediately before TrainConfig
    if "class LeRobotG1Dex3DataConfig" not in new:
        anchor = "@dataclasses.dataclass(frozen=True)\nclass TrainConfig:"
        if anchor not in new:
            print("FAIL: could not find the TrainConfig anchor in config.py")
            return 1
        new = new.replace(anchor, DATA_CONFIG.lstrip("\n") + "\n" + anchor, 1)
        actions.append("insert LeRobotG1Dex3DataConfig")

    # 4. the two TrainConfig entries, at the head of _CONFIGS
    if 'name="pi05_g1_pickplace"' not in new:
        anchor = "_CONFIGS = [\n"
        if anchor not in new:
            print("FAIL: could not find the _CONFIGS anchor in config.py")
            return 1
        block = TRAIN_CONFIGS.replace("{REPO_ID}", a.repo_id).replace("{NUM_WORKERS}", str(a.num_workers))
        new = new.replace(anchor, anchor + block, 1)
        actions.append(f'insert TrainConfigs (repo_id="{a.repo_id}", num_workers={a.num_workers})')

    if not actions:
        print("Already patched -- nothing to do.")
    else:
        print("planned:")
        for x in actions:
            print("  -", x)

    if a.dry_run:
        print("\n--dry-run: nothing written.")
        return 0

    if new != cfg.read_text():
        try:
            ast.parse(new)
        except SyntaxError as e:
            print(f"FAIL: patched config.py would not parse: {e}")
            return 1
        if not backup.exists():
            shutil.copy2(cfg, backup)
            print(f"backed up -> {backup.relative_to(openpi)}")
        cfg.write_text(new)
    if need_policy:
        shutil.copy2(pol_src, pol_dst)

    if a.no_verify:
        print("\npatched. Skipping verification (--no-verify). Check it yourself with:")
        print("  uv run python -c \"from openpi.training import config; "
              "print(config.get_config('pi05_g1_pickplace').name)\"")
        return 0

    # verify for real, in the openpi environment
    print("\nverifying...")
    code = (
        "from openpi.training import config\n"
        "for n in ('pi05_g1_pickplace', 'pi05_g1_pickplace_frozen_vlm'):\n"
        "    c = config.get_config(n); m = c.model\n"
        "    assert m.max_token_len >= 200, f'{n}: max_token_len={m.max_token_len} too small'\n"
        "    assert m.discrete_state_input, f'{n}: discrete_state_input is False'\n"
        "    assert m.action_dim == 32 and m.action_horizon == 50\n"
    "    d = c.data.create(c.assets_dirs, c.model)\n"
    "    assert tuple(d.action_sequence_keys) == ('action',), \\\n"
    "        f'{n}: action_sequence_keys={d.action_sequence_keys}, must be (\\'action\\',) for this dataset'\n"
        "    print(f'  {n}: action_dim={m.action_dim} horizon={m.action_horizon} "
        "max_token_len={m.max_token_len} discrete_state={m.discrete_state_input} "
        "batch={c.batch_size} ema={c.ema_decay} workers={c.num_workers} repo={c.data.repo_id}')\n"
        "print('OK')\n"
    )
    r = subprocess.run(["uv", "run", "python", "-c", code], cwd=openpi,
                       capture_output=True, text=True)
    print(r.stdout.strip() or r.stderr.strip()[-2000:])
    return r.returncode


if __name__ == "__main__":
    sys.exit(main())