"""ACT args for the pinned Psi0 scripts/train.py (config hb_act_config), on the Psi0 views $G/psi0-data/. Same 20-epoch plan / probe decision as Psi-Zero (hb_psi0/hb_args.py: plan, decide, receipt); only the args differ. hb_act_args.py plan hb_act_args.py args (--decision D | --batch B --steps S --save V --warmup W) --out DIR --timestamp T hb_act_args.py decide """ import argparse import json import os import sys from pathlib import Path H = Path(__file__).resolve().parent sys.path.insert(0, str(H.parent / "hb_psi0")) import hb_args as HA # noqa: E402 (TASKS, DATA, receipt, plan, decide; EPOCHS 20, save at 10, warmup 2.5 %) CHUNK = 16 # native ACT G1 simulation recipe (train_act_g1_simple.sh) def build_args(task, batch, steps, save, warmup, out, ts, workers, resume=True): r = HA.receipt(task) cams = r["image_keys"] SK, AK, SD, AD, NB = r["state_key"], r["action_key"], int(r["state_dim"]), int(r["action_dim"]), int(r["body_dims"]) assert SD in (44, 78), f"raw state must be 44 / 78 (no joint velocity), got {SD}" stats = str(HA.DATA / task / "meta" / "stats_psi0.json") return [ "--seed=292285", f"--exp=hb-act-{task}", f"--timestamp={ts}", *(["--train.resume_from_checkpoint=latest"] if resume else []), "--train.name=act-g1", f"--train.output_dir={out}", "--train.data_parallel=ddp", "--train.mixed_precision=bf16", f"--train.train_batch_size={batch}", f"--train.val_batch_size={batch}", "--train.max_checkpoints_to_keep=3", "--train.preserve_checkpoint_every=None", "--train.gradient_accumulation_steps=1", "--train.learning_rate=1e-4", f"--train.max_training_steps={steps}", "--train.warmup_ratio=None", f"--train.warmup_steps={warmup}", f"--train.checkpointing_steps={save}", "--train.validation_steps=0", "--train.max_grad_norm=1.0", "--train.lr_scheduler_type=cosine", "--train.lr_scheduler_kwargs.weight_decay=1e-6", "--train.lr_scheduler_kwargs.betas", "0.95", "0.999", f"--data.root_dir={HA.DATA}", "--data.train_repo_ids", task, "--data.val_repo_ids", task, f"--data.num_workers={workers}", "--data.prefetch_factor=4", "--data.transform.repack.image-keys", *cams, "--data.transform.repack.state-keys", SK, "--data.transform.repack.action-keys", AK, f"--data.transform.repack.dataset-name=hb_{task}", f"--data.transform.repack.pad-action-dim={AD}", f"--data.transform.repack.pad-state-dim={SD}", "--data.transform.repack.instruction-key=task", f"--data.transform.repack.action-chunk-size={CHUNK}", "--data.transform.repack.state-temporal-jitter=0", "--data.transform.repack.hb-delay-prob=0.15", "--data.transform.repack.hb-delay-max-frames=2", f"--data.transform.repack.hb-body-dims={NB}", "--data.transform.repack.hb-bias-seed=292285", f"--data.transform.field.stat-path={stats}", "--data.transform.field.state-noise-std=0.0", "--data.transform.field.stat-action-keys", AK, "--data.transform.field.stat-state-keys", SK, "--data.transform.field.action_norm_type=bounds", "--data.transform.field.no-use-norm-mask", "--data.transform.field.normalize-state", f"--data.transform.field.pad-action-dim={AD}", f"--data.transform.field.pad-state-dim={SD}", "--data.transform.model.no-img-aug", "--data.transform.model.hb-aug", f"--model.chunk-size={CHUNK}", f"--model.n-action-steps={CHUNK}", f"--model.action-dim={AD}", f"--model.state-dim={SD}", "--model.use-vae", "--model.kl-weight=10.0", ] def main(): ap = argparse.ArgumentParser() sub = ap.add_subparsers(dest="cmd", required=True) p = sub.add_parser("plan"); p.add_argument("task"); p.add_argument("batch", type=int) p = sub.add_parser("args"); p.add_argument("task") p.add_argument("--decision"); p.add_argument("--batch", type=int); p.add_argument("--steps", type=int) p.add_argument("--save", type=int); p.add_argument("--warmup", type=int) p.add_argument("--out", required=True); p.add_argument("--timestamp", required=True) p.add_argument("--workers", type=int, default=int(os.environ.get("HB_NUM_WORKERS", "8"))) p.add_argument("--no-resume", action="store_true") p = sub.add_parser("decide"); p.add_argument("task"); p.add_argument("probe_dir"); p.add_argument("decision") p.add_argument("cands", nargs="+") a = ap.parse_args() if a.cmd != "decide" and a.task not in HA.TASKS: sys.exit(f"unknown task {a.task}; known {HA.TASKS}") if a.cmd == "plan": print(json.dumps(HA.plan(HA.receipt(a.task)["frames"], a.batch), indent=2)) elif a.cmd == "args": if a.decision: d = json.loads(Path(a.decision).read_text()) assert d["task"] == a.task and d["frames"] == HA.receipt(a.task)["frames"], "stale decision file" b, s, sv, w = d["per_gpu_batch"], d["steps"], d["save_steps"], d["warmup_steps"] else: b, s, sv, w = a.batch, a.steps, a.save, a.warmup assert None not in (b, s, sv, w), "--batch/--steps/--save/--warmup or --decision" print("\n".join(build_args(a.task, b, s, sv, w, a.out, a.timestamp, a.workers, not a.no_resume))) else: sys.exit(HA.decide(a.task, a.probe_dir, a.decision, a.cands)) if __name__ == "__main__": main()