"""Train an FM-DiT, DP (DDPM-DiT) or ACT baseline on a task's LeRobot v2.1 expert dataset. /bin/python -I evaluation/baselines/train.py --policy fm_dit --contract h2_sharpa \\ --dataset data/h2_pour_water/h2_pour_water_r1/lerobot --out --steps 60000 The recipe is DextraVerse's shared IL recipe (configs/algorithm/il/train.yaml; runners/train_dex_flow_matching.py): AdamW (lr 1e-4, betas 0.95/0.999, weight decay 1e-6), cosine decay (here after a 500-step linear warm-up, as Diffusion Policy), whole-policy EMA (Diffusion Policy's warm-up schedule, max decay 0.9999) whose weights are the saved checkpoint, batch 64, one observation frame and 16-row chunks. Added for the humanoid views: frames resized to 168x224 (the 480x640 aspect). bf16 autocast on CUDA. Augmentation (2026-11-16): the Humanoid Benchmark spec v2 (docs/augmentation_v2.md, evaluation/augmentation), the same for every model of the benchmark, is on by default (--augmentation hb-v2): per training item one hb_aug_v2 call on the cached 168x224 frames of all cameras (colour shared, geometry / light patches / white balance per camera, one of the 15 effects boosted 1.5 x) and the per-sample joint bias (+-2 deg) + noise (0.25 deg, clipped 3 sigma) on the body joint angles of the raw state (G1 29, H2 31), before the policy's normalizers; actions untouched. The native random shift is off (--shift 0; a shift with hb-v2 is refused). Before the first step a preflight reads training items through the augmented and the plain path and checks the state width (G1 44 / H2 78, no velocity), that only the body joints move and by at most 2.75 deg, that the action rows are bit-identical, that the images change, and that the offline-check items are deterministic; it is stored in run.json (--preflight-only stops after it). --augmentation none is an explicit exception to the spec, recorded in run.json and the checkpoints. The ResNet18 starts from torchvision's ImageNet-1k weights (--backbone-weights IMAGENET1K_V1, downloaded once to the torch hub cache) and is fine-tuned at 0.1 x the learning rate (--backbone-lr-scale, the DETR practice) with its BatchNorm frozen; --backbone-weights none trains it from scratch (DextraVerse's GroupNorm encoder for FM-DiT and DP). Writes into --out: run.json (arguments), train_log.jsonl, checkpoints/step-NNNNNN/ every --save-every steps and at the end (runtime.py format), checkpoints/final -> the last one, and train_state.pt (optimizer, schedule, EMA, step) that --resume continues from. Each save also measures the EMA policy offline on fixed frames (held-out episodes with --val-episodes, else training frames): normalized-action MSE overall and per action group. """ from __future__ import annotations import argparse import copy from datetime import datetime import json import math import os from pathlib import Path import subprocess import sys import time import numpy as np HERE = Path(__file__).resolve().parent REPO = HERE.parents[1] if str(HERE.parent) not in sys.path: sys.path.insert(0, str(HERE.parent)) from baselines.data import (CONTRACTS, ChunkDataset, LeRobotDataset, augmentation_module, frame_cache, # noqa: E402 load_lowdim, parse_episodes, statistics) from baselines.policies import POLICIES, PolicyConfig, build_policy # noqa: E402 from baselines.runtime import save_checkpoint # noqa: E402 DEFAULTS = dict(lr=1e-4, betas=(0.95, 0.999), weight_decay=1e-6, warmup=500, ema=0.9999, batch_size=64, backbone_weights='IMAGENET1K_V1', backbone_lr_scale=0.1) def git_commit(): try: return subprocess.run(['git', '-C', str(REPO), 'rev-parse', 'HEAD'], capture_output=True, text=True, check=True).stdout.strip() except Exception: # noqa: BLE001 - outside a checkout return None def parse_hw(text): h, w = text.lower().split('x') return int(h), int(w) def random_shift(images, pad): """DrQ random shift: replicate-pad by `pad` pixels and crop back at a random offset per sample. images (B, N, 3, H, W) float.""" import torch import torch.nn.functional as F if pad <= 0: return images B, N, C, H, W = images.shape x = F.pad(images.reshape(B * N, C, H, W), (pad,) * 4, mode='replicate') oy = torch.randint(0, 2 * pad + 1, (B * N,), device=images.device) ox = torch.randint(0, 2 * pad + 1, (B * N,), device=images.device) rows = (oy[:, None] + torch.arange(H, device=images.device)[None])[:, None, :, None].expand(-1, C, -1, W + 2 * pad) x = x.gather(2, rows) cols = (ox[:, None] + torch.arange(W, device=images.device)[None])[:, None, None, :].expand(-1, C, H, -1) return x.gather(3, cols).reshape(B, N, C, H, W) class EMA: """Diffusion Policy's EMAModel schedule: decay = min(max_decay, 1 - (1 + step) ** -0.75), parameters averaged, buffers (normalizers) copied.""" def __init__(self, model, max_decay): self.model = copy.deepcopy(model).eval().requires_grad_(False) self.max_decay = float(max_decay) def update(self, model, step): import torch decay = 0.0 if step <= 0 else min(self.max_decay, max(0.0, 1 - (1 + step) ** -0.75)) with torch.no_grad(): for e, p in zip(self.model.parameters(), model.parameters()): e.mul_(decay).add_(p.detach(), alpha=1 - decay) for e, b in zip(self.model.buffers(), model.buffers()): e.copy_(b) return decay def preflight(train_set, eval_set, contract, samples=64, seed=0) -> dict: """Training items of `train_set` (augmented unless its body_dims is 0) against the same frames read without augmentation, and two reads of the offline-check items; raises on any violation of docs/augmentation_v2.md.""" import torch plain = ChunkDataset(train_set.lowdim, train_set.paths, train_set.cameras, train_set.hw, train_set.horizon, train_set.positions) # the same frames without augmentation picks = np.random.default_rng(seed).choice(len(plain), size=min(samples, len(plain)), replace=False) nb, sd = train_set.body_dims, contract.state_dim limit = augmentation_module().max_body_offset_rad() * (1 + 1e-5) images_changed = body_moved = 0 worst = 0.0 for k in picks: (im, st, ac), (im0, st0, ac0) = train_set[int(k)], plain[int(k)] if st.shape != (sd,) or st0.shape != (sd,) or ac.shape != ac0.shape: raise SystemExit(f'preflight: state {tuple(st.shape)} / action {tuple(ac.shape)}, expected ({sd},)') if not torch.equal(ac, ac0): raise SystemExit('preflight: an action label differs from the recorded data') if not torch.equal(st[nb:], st0[nb:]): raise SystemExit(f'preflight: a state value past the {nb} body joints changed') delta = (st[:nb].double() - st0[:nb].double()).abs() if nb: worst = max(worst, float(delta.max())) if worst > limit: raise SystemExit(f'preflight: a body joint moved {math.degrees(worst):.3f} deg (> 2.75)') body_moved += int((delta > 0).sum()) images_changed += int(not torch.equal(im, im0)) n = len(picks) moved = body_moved / max(1, n * nb) if nb and (images_changed < 0.9 * n or moved < 0.95): raise SystemExit(f'preflight: augmentation inactive ({images_changed}/{n} images changed, {moved:.3f} of the ' 'body joint values moved)') if not nb and (images_changed or worst): raise SystemExit('preflight: items changed with the augmentation off') eval_picks = np.random.default_rng(seed + 1).choice(len(eval_set), size=min(8, len(eval_set)), replace=False) for k in eval_picks: a, b = eval_set[int(k)], eval_set[int(k)] if eval_set.body_dims or not all(torch.equal(x, y) for x, y in zip(a, b)): raise SystemExit('preflight: the offline-check items are not deterministic') return dict(samples=n, state_dim=sd, body_dims=nb, images_changed=images_changed, body_values_moved=round(moved, 4), max_body_offset_deg=round(math.degrees(worst), 4), actions_bit_identical=True, non_body_state_bit_identical=True, eval_items_deterministic=len(eval_picks)) def offline_eval(policy, dataset, samples, device, groups): """Normalized-action MSE of policy.predict on fixed dataset items (seeded), overall and per action group.""" import torch policy.eval() errs = [] generator = torch.Generator(device=device) generator.manual_seed(0) for start in range(0, len(samples), 32): batch = [dataset[int(k)] for k in samples[start:start + 32]] images = torch.stack([b[0] for b in batch]).to(device).permute(0, 1, 4, 2, 3).float().div_(255.) state = torch.stack([b[1] for b in batch]).to(device) target = torch.stack([b[2] for b in batch]).to(device) _, raw = policy.predict(images, state, generator=generator) errs.append((raw.float() - policy.action_norm.normalize(target)).pow(2).cpu().numpy()) err = np.concatenate(errs) # (S, horizon, action_dim) out = dict(mse=float(err.mean()), mse_row0=float(err[:, 0].mean())) offset = 0 for column, start, end in groups: out[f'mse_{column.split(".")[-1]}'] = float(err[..., offset:offset + end - start].mean()) offset += end - start return out def main(): p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) p.add_argument('--policy', choices=sorted(POLICIES), required=True) p.add_argument('--dataset', type=Path, required=True, help='LeRobot v2.1 dataset directory') p.add_argument('--contract', choices=sorted(CONTRACTS), required=True) p.add_argument('--cameras', help='comma-separated cameras (default: the contract\'s)') p.add_argument('--episodes', help="episodes to use, e.g. '0:100' (default: all)") p.add_argument('--val-episodes', type=int, default=0, help='hold out the last N of them for the offline check') p.add_argument('--out', type=Path, required=True) p.add_argument('--steps', type=int, default=60000) p.add_argument('--batch-size', type=int, default=DEFAULTS['batch_size']) p.add_argument('--lr', type=float, default=DEFAULTS['lr']) p.add_argument('--weight-decay', type=float, default=DEFAULTS['weight_decay']) p.add_argument('--warmup', type=int, default=DEFAULTS['warmup']) p.add_argument('--ema', type=float, default=DEFAULTS['ema'], help='max EMA decay; 0 saves the raw weights') p.add_argument('--horizon', type=int, default=16) p.add_argument('--image-size', default='168x224', help='HxW of the policy images') p.add_argument('--augmentation', choices=['hb-v2', 'none'], default='hb-v2', help='training augmentation: the benchmark spec v2 (default; docs/augmentation_v2.md) or none ' '(an explicit exception, recorded in run.json)') p.add_argument('--shift', type=int, default=0, help='native random-shift augmentation in pixels (default 0: off; only with --augmentation none)') p.add_argument('--preflight-only', action='store_true', help='run the data preflight, write run.json and stop') p.add_argument('--backbone-weights', default=DEFAULTS['backbone_weights'], help="torchvision weights the ResNet18 starts from (default %(default)s); 'none': from scratch") p.add_argument('--backbone-lr-scale', type=float, default=DEFAULTS['backbone_lr_scale'], help='learning rate of the image backbone relative to --lr (default %(default)s)') p.add_argument('--model', action='append', default=[], metavar='KEY=VALUE', help='override a PolicyConfig field (JSON value), e.g. num_layers=6; repeatable') p.add_argument('--cache-dir', type=Path, default=Path('~/hb_models/baselines/cache')) p.add_argument('--workers', type=int, default=32, help='DataLoader workers (and frame-cache decoders); the v2 augmentation costs ~17 ms of one CPU ' 'core per item, so 8 workers feed only ~6 steps/s at batch 64') p.add_argument('--prefetch', type=int, default=2, help='batches prefetched per worker') p.add_argument('--save-every', type=int, default=10000) p.add_argument('--log-every', type=int, default=100) p.add_argument('--eval-samples', type=int, default=256) p.add_argument('--seed', type=int, default=0) p.add_argument('--device', default='cuda:0') p.add_argument('--no-amp', action='store_true') p.add_argument('--resume', action='store_true', help='continue from /train_state.pt') args = p.parse_args() out = args.out.expanduser().resolve() out.mkdir(parents=True, exist_ok=True) contract = CONTRACTS[args.contract] aug = augmentation_module() aug.check_state_layout(contract.robot, contract.state_dim, [column for column, _, _ in contract.state]) hb_aug = args.augmentation == 'hb-v2' if hb_aug and args.shift: raise SystemExit('--shift is the native random shift, which the augmentation spec v2 turns off ' '(docs/augmentation_v2.md); use it only with --augmentation none') if contract.body_dims != aug.BODY_JOINTS[contract.robot]: raise SystemExit(f'contract {contract.name}: body_dims {contract.body_dims}, the spec perturbs ' f'{aug.BODY_JOINTS[contract.robot]} body joints') augmentation = (aug.describe(contract.robot, contract.body_dims) if hb_aug else dict(version=None, exception='--augmentation none (not the benchmark default)')) cameras = tuple(args.cameras.split(',')) if args.cameras else contract.cameras hw = parse_hw(args.image_size) dataset = LeRobotDataset(args.dataset) dataset.check_contract(contract, cameras) episodes = parse_episodes(args.episodes, dataset.lengths) if not 0 <= args.val_episodes < len(episodes): raise SystemExit('--val-episodes must leave training episodes') lowdim = load_lowdim(dataset, contract, episodes) if len(lowdim['tasks']) != 1: raise SystemExit(f"the baselines take no language: one task per dataset, found {lowdim['tasks']}") n_train = len(episodes) - args.val_episodes train_pos, val_pos = list(range(n_train)), list(range(n_train, len(episodes))) train_rows = np.concatenate([np.arange(s, s + n) for s, n in zip(lowdim['starts'][:n_train], lowdim['lengths'][:n_train])]) stats = statistics(lowdim, train_rows) frames = frame_cache(dataset, lowdim, cameras, hw, args.cache_dir, workers=args.workers) train_set = ChunkDataset(lowdim, frames, cameras, hw, args.horizon, train_pos, body_dims=contract.body_dims if hb_aug else 0) eval_set = ChunkDataset(lowdim, frames, cameras, hw, args.horizon, val_pos or train_pos) # never augmented checks = preflight(train_set, eval_set, contract, seed=args.seed) print(json.dumps(dict(event='preflight', augmentation=augmentation.get('version'), **checks)), flush=True) if args.preflight_only: (out / 'run.json').write_text(json.dumps(dict( argv=sys.argv, started=datetime.now().isoformat(timespec='seconds'), host=os.uname().nodename, contract=contract.to_dict(cameras), image_hw=list(hw), augmentation=augmentation, preflight=checks, shift_px=args.shift, code_commit=git_commit()), indent=1, default=str)) return import torch torch.manual_seed(args.seed) np.random.seed(args.seed) device = torch.device(args.device) overrides = {k: json.loads(v) for k, v in (item.split('=', 1) for item in args.model)} weights = None if str(args.backbone_weights).lower() in ('none', '') else args.backbone_weights pcfg = PolicyConfig.from_dict({**dict(policy=args.policy, n_cameras=len(cameras), state_dim=contract.state_dim, action_dim=contract.action_dim, horizon=args.horizon, backbone_weights=weights), **overrides}) policy = build_policy(pcfg) policy.fit_normalizers(stats) policy.to(device) ema = EMA(policy, args.ema) if args.ema > 0 else None named = list(policy.named_parameters()) backbone = [p for n, p in named if n.startswith(policy.backbone_prefix)] rest = [p for n, p in named if not n.startswith(policy.backbone_prefix)] if not backbone: raise SystemExit(f'no backbone parameters under {policy.backbone_prefix!r}') optimizer = torch.optim.AdamW([dict(params=rest), dict(params=backbone, lr=args.lr * args.backbone_lr_scale)], lr=args.lr, betas=DEFAULTS['betas'], weight_decay=args.weight_decay) def lr_lambda(step): if step < args.warmup: return (step + 1) / max(1, args.warmup) return 0.5 * (1 + math.cos(math.pi * min(1.0, (step - args.warmup) / max(1, args.steps - args.warmup)))) scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda) step = 0 state_path = out / 'train_state.pt' if args.resume and state_path.exists(): saved = torch.load(state_path, map_location='cpu', weights_only=False) if saved['config']['policy'] != pcfg.to_dict(): raise SystemExit('train_state.pt was written for another policy configuration') before = saved['config'].get('training', {}) if (before.get('augmentation') or {}).get('version') != augmentation.get('version') \ or before.get('shift_px') != args.shift: raise SystemExit('train_state.pt was trained with another augmentation (version / --shift); resuming ' 'would change it mid-run') policy.load_state_dict(saved['policy']) optimizer.load_state_dict(saved['optimizer']) scheduler.load_state_dict(saved['scheduler']) if ema is not None: ema.model.load_state_dict(saved['ema']) step = int(saved['step']) print(json.dumps(dict(event='resumed', step=step)), flush=True) samples = np.random.default_rng(1).choice(len(eval_set), size=min(args.eval_samples, len(eval_set)), replace=False) config = dict( policy=pcfg.to_dict(), contract=contract.to_dict(cameras), image_hw=list(hw), task_text=lowdim['tasks'][0], dataset=dict(name=dataset.name, root=str(dataset.root), robot_type=dataset.info.get('robot_type'), fps=dataset.fps, episodes=episodes, train_episodes=episodes[:n_train], val_episodes=episodes[n_train:], train_frames=int(len(train_set)), git_commit=dataset.info.get('script_config', {}).get('git_commit')), training=dict(steps=args.steps, batch_size=args.batch_size, lr=args.lr, betas=list(DEFAULTS['betas']), weight_decay=args.weight_decay, warmup=args.warmup, schedule='linear warm-up + cosine', backbone_lr_scale=args.backbone_lr_scale, ema=args.ema, shift_px=args.shift, amp=not args.no_amp, seed=args.seed, augmentation=augmentation, preflight=checks), stats=stats, code_commit=git_commit()) run = dict(argv=sys.argv, started=datetime.now().isoformat(timespec='seconds'), host=os.uname().nodename, torch=torch.__version__, **config) (out / 'run.json').write_text(json.dumps(run, indent=1, default=str)) loader = torch.utils.data.DataLoader(train_set, batch_size=args.batch_size, shuffle=True, drop_last=True, num_workers=args.workers, pin_memory=True, persistent_workers=True, prefetch_factor=args.prefetch) params = sum(x.numel() for x in policy.parameters()) print(json.dumps(dict(event='start', policy=args.policy, params=params, backbone_params=sum(x.numel() for x in backbone), backbone_weights=pcfg.backbone_weights, backbone_norm=pcfg.backbone_norm, train_frames=len(train_set), steps=args.steps, from_step=step, cameras=list(cameras), hw=list(hw))), flush=True) log = (out / 'train_log.jsonl').open('a') amp = (not args.no_amp) and device.type == 'cuda' def save(step): folder = out / f'checkpoints/step-{step:06d}' model = ema.model if ema is not None else policy metrics = offline_eval(model, eval_set, samples, device, contract.action) save_checkpoint(folder, model, dict(config, step=step, weights='ema' if ema is not None else 'raw', offline_eval=dict(metrics, samples=len(samples), frames='val' if val_pos else 'train'))) final = out / 'checkpoints/final' if final.is_symlink() or final.exists(): final.unlink() final.symlink_to(folder.name) torch.save(dict(policy=policy.state_dict(), optimizer=optimizer.state_dict(), scheduler=scheduler.state_dict(), ema=None if ema is None else ema.model.state_dict(), step=step, config=config), out / 'train_state.pt.tmp') (out / 'train_state.pt.tmp').replace(state_path) record = dict(event='save', step=step, checkpoint=str(folder), **metrics) print(json.dumps(record), flush=True) log.write(json.dumps(record) + '\n') log.flush() policy.train() policy.train() t0, data_s, acc = time.monotonic(), 0.0, {} while step < args.steps: tick = time.monotonic() for images, state, action in loader: data_s += time.monotonic() - tick images = images.to(device, non_blocking=True).permute(0, 1, 4, 2, 3).float().div_(255.) images = random_shift(images, args.shift) state, action = state.to(device, non_blocking=True), action.to(device, non_blocking=True) with torch.autocast(device.type, dtype=torch.bfloat16, enabled=amp): loss, metrics = policy.compute_loss(images, state, action) optimizer.zero_grad(set_to_none=True) loss.backward() optimizer.step() scheduler.step() step += 1 decay = ema.update(policy, step) if ema is not None else 0.0 for k, v in dict(metrics, loss=float(loss.detach())).items(): acc[k] = acc.get(k, 0.0) + v if step % args.log_every == 0: elapsed = time.monotonic() - t0 record = dict(event='train', step=step, lr=scheduler.get_last_lr()[0], backbone_lr=scheduler.get_last_lr()[1], ema_decay=round(decay, 6), it_s=round(args.log_every / elapsed, 2), data_frac=round(data_s / elapsed, 3), **{k: v / args.log_every for k, v in acc.items()}) print(json.dumps(record), flush=True) log.write(json.dumps(record) + '\n') log.flush() t0, data_s, acc = time.monotonic(), 0.0, {} if step % args.save_every == 0 or step == args.steps: save(step) if step >= args.steps: break tick = time.monotonic() log.close() print(json.dumps(dict(event='done', step=step, checkpoint=str(out / 'checkpoints/final'))), flush=True) if __name__ == '__main__': main()