"""Timing pipeline check: model inputs equal the action pipeline's, and the model's timing labels (computed from the normalized pipeline actions) equal calibrate_timing_anchor.py's from the raw parquet (same dataset indices).""" import os import sys os.environ.setdefault("JAX_PLATFORMS", "cpu") sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) # adaspe_train/ import configs # noqa: E402 if __name__ == "__main__": import jax import jax.numpy as jnp import numpy as np import openpi.training.data_loader as dl from adaspe import labels configs.register() act_cfg, tim_cfg = configs.INPUT_REF, configs.TIMINGS[0] def pipeline(cfg): dc = cfg.data.create(cfg.assets_dirs, cfg.model) ds = dl.create_torch_dataset(dc, cfg.model.action_horizon, cfg.model) return dl.transform_dataset(ds, dc), dc act_ds, _ = pipeline(act_cfg) tim_ds, tdc = pipeline(tim_cfg) print("timing sequence keys:", tdc.action_sequence_keys) import tasks ref = np.load(tasks.timing_labels_path()) rng = np.random.default_rng(0) idx = np.concatenate([rng.choice(len(act_ds), 200, replace=False), [len(act_ds) - 1, len(act_ds) - 5]]) mismatch = 0 labs = [] for i in idx: a, t = act_ds[int(i)], tim_ds[int(i)] for k in ("state", "tokenized_prompt", "tokenized_prompt_mask"): assert np.array_equal(a[k], t[k]), (i, k) for k in a["image"]: assert np.array_equal(a["image"][k], t["image"][k]), (i, k) assert np.array_equal(a["actions"][:, :7], t["actions"][:, :7]), i assert np.all(a["actions"][:, 7:] == 0) and np.all(t["actions"][:, 14:] == 0), i # The exact label the model trains on (adaspe/labels.py), from the normalized pipeline actions. lab = float(labels.timing_target(jnp.asarray(t["actions"][None]), tim_cfg.model)[0, 0]) labs.append(lab) mismatch += abs(lab - ref[i]) > 1e-4 print(f"checked {len(idx)} samples: inputs identical to action pipeline; label mismatches vs calibration: {mismatch}") print("label sample:", np.round(labs[:12], 2))