diag: H 段（老师 oracle 走 rollout / 前缀消融 / 子目标越过顶点规则 / 分轴与分类别指标）；训练：EMA、选模不落在 p_neq 预热、闭环分数选模、24000 步余弦、p_neq 0.5 + p_edge 0.1、子目标头切断抄输入、前缀/dq 增广；仿真：last 记真正提交的候选、碰撞与到达按每题时限、回放门禁读 pass+指纹并走 rollout；数据：动作标签按 1.25·STEP 归一化（act_scale 随 meta/ckpt/ONNX 传递，旧模型 1.0）、收回亏空结转、DAgger 带 qPrev 与起步速度、打包签名变更自动重打；ik-policy.js 同步（actScale、子目标规则、DLS 时清 lastPlan）；本机执行清单与 README 更新
This commit is contained in:
Hark
+26
-20
@@ -6,7 +6,10 @@
|
||||
2026-10-18：展开长度不再定长 60 次推理（3.6 s，会截掉老师时长 > 3.6 s 的绕远题），改为覆盖本批 exec_s p90×1.2，上限 plans_max；
|
||||
"到达"仍是每 seg_plans 段末的距离项（无到达即停），属有意设计：段末项给出持续的接近梯度。
|
||||
margin_train 3 mm：与训练数据过滤线一致（认证最小间隙 < 3 mm 的老师轨迹不进训练集），碰撞项不与 BC 标签对抗；
|
||||
运行时投影护栏独立保持 5 mm。端点与中点两个铰链取均值，与只查端点同量纲，w_col 无需重标。"""
|
||||
运行时投影护栏独立保持 5 mm。端点与中点两个铰链取均值，与只查端点同量纲，w_col 无需重标。
|
||||
2026-10-19：起点取 EMA 权重（ckpt.load_policy）；动作反归一化乘 act_scale（与数据同一尺度，不一致报错），产物存 act_scale；
|
||||
子目标推进改用 closed_loop.sub_advance（默认「越过顶点」，与标签同口径）；子目标头 BC 项用 sub 特征置零的输入（同 train.py）；
|
||||
val 只取主数据目录（第一个 --samples）。"""
|
||||
import argparse, json, math, os, random, sys, time
|
||||
import torch
|
||||
|
||||
@@ -15,7 +18,7 @@ from ml.common.features import features, STEP, H, wrap180
|
||||
from ml.data import Samples, load_tasks
|
||||
from ml.model.policy import build, wta_loss
|
||||
from ml.model.meanflow import meanflow_loss
|
||||
from ml.sim.closed_loop import evaluate, tasks_to_tensors, goal_err
|
||||
from ml.sim.closed_loop import evaluate, tasks_to_tensors, goal_err, sub_advance
|
||||
from ml.common.ckpt import Saver, atomic_save, rng_state, set_rng_state
|
||||
from ml.common import vram
|
||||
|
||||
@@ -41,18 +44,23 @@ def main():
|
||||
dev = a.device; torch.manual_seed(a.seed); random.seed(a.seed)
|
||||
vram.limit()
|
||||
os.makedirs(a.out, exist_ok=True)
|
||||
ck = torch.load(a.init, map_location=dev)
|
||||
mcfg = ck['cfg']
|
||||
model = build(mcfg).to(dev); model.load_state_dict(ck['model'])
|
||||
from ml.common.ckpt import load_policy
|
||||
model, ck = load_policy(a.init, dev); mcfg = ck['cfg']
|
||||
kin = Kin(device=dev)
|
||||
data = Samples(a.samples, device=dev, split='train')
|
||||
data = Samples(a.samples, device=dev, split='train', val_dirs=a.samples[:1])
|
||||
act_scale = float(ck.get('act_scale', 1.0))
|
||||
if abs(act_scale - data.act_scale) > 1e-9:
|
||||
raise SystemExit(f'起点 act_scale={act_scale} 与样本 act_scale={data.act_scale} 不一致（BC 项标签尺度不同），换同一版数据训练的模型')
|
||||
model.act_scale = act_scale
|
||||
all_tasks = load_tasks(a.samples)
|
||||
train_tasks = [t for t in all_tasks if t['split'] == 'train'] + load_pool(a.pool)
|
||||
val_tasks = [t for t in all_tasks if t['split'] == 'val'][:C['eval_tasks']]
|
||||
main_tid = {t['tid'] for t in load_tasks(a.samples[:1]) if t['split'] == 'val'}
|
||||
train_tasks = [t for t in all_tasks if t['tid'] not in main_tid] + load_pool(a.pool)
|
||||
val_tasks = [t for t in load_tasks(a.samples[:1]) if t['split'] == 'val'][:C['eval_tasks']]
|
||||
w = [C['cat_weights'].get(t.get('cat', 'A'), 1.0) * (2.0 if t.get('hard') else 1.0) for t in train_tasks]
|
||||
opt = torch.optim.AdamW(model.parameters(), lr=C['lr'], betas=(0.9, 0.95), weight_decay=0.0)
|
||||
amp = dev.startswith('cuda') and C.get('amp') == 'bf16'
|
||||
step = torch.tensor(STEP, device=dev)
|
||||
astep = step * act_scale # 动作 → 度
|
||||
it0 = 0
|
||||
if a.resume and os.path.exists(a.resume):
|
||||
rk = torch.load(a.resume, map_location=dev); model.load_state_dict(rk['model'])
|
||||
@@ -68,11 +76,11 @@ def main():
|
||||
base, _ = evaluate(model.eval(), kin, val_tasks, dev)
|
||||
print('EVAL0', json.dumps(base), flush=True); log.write(json.dumps({'eval': base, 'step': 0}) + '\n')
|
||||
best = base['success']
|
||||
atomic_save({'model': model.state_dict(), 'cfg': mcfg, 'step': 0, 'fp': kin.fp, 'eval': base}, os.path.join(a.out, 'best.pt'))
|
||||
atomic_save({'model': model.state_dict(), 'cfg': mcfg, 'step': 0, 'fp': kin.fp, 'eval': base, 'act_scale': act_scale}, os.path.join(a.out, 'best.pt'))
|
||||
M = C['margin_train_mm']; t0 = time.time()
|
||||
saver = Saver(C.get('ckpt_min', 10)) # 2026-10-18：每 10 分钟 + 收到停止请求时存 last.pt（不再只在每 500 次验证时存）
|
||||
def save_last(n, ev=None):
|
||||
atomic_save({'model': model.state_dict(), 'opt': opt.state_dict(), 'cfg': mcfg, 'step': n, 'fp': kin.fp, 'eval': ev, 'rng': rng_state()}, os.path.join(a.out, 'last.pt')); saver.mark()
|
||||
atomic_save({'model': model.state_dict(), 'opt': opt.state_dict(), 'cfg': mcfg, 'step': n, 'fp': kin.fp, 'eval': ev, 'rng': rng_state(), 'act_scale': act_scale}, os.path.join(a.out, 'last.pt')); saver.mark()
|
||||
for it in range(it0, C['steps']):
|
||||
model.train()
|
||||
batch = random.choices(train_tasks, weights=w, k=C['batch'])
|
||||
@@ -81,7 +89,7 @@ def main():
|
||||
z = torch.randn(B, H, 6, device=dev)
|
||||
q, qp = q0.clone(), q0.clone()
|
||||
queue = torch.zeros(B, 0, 6, device=dev)
|
||||
subs = torch.zeros(B, 2, 6, device=dev); nsub = torch.zeros(B, dtype=torch.long, device=dev); sub_t = torch.zeros(B, device=dev)
|
||||
subs = torch.zeros(B, 2, 6, device=dev); nsub = torch.zeros(B, dtype=torch.long, device=dev); sub_t = torch.zeros(B, device=dev); sub_prev = q0.clone()
|
||||
opt.zero_grad(set_to_none=True)
|
||||
stats = {'col': 0.0, 'tgt': 0.0, 'smooth': 0.0}
|
||||
# 展开长度随本批任务时长：覆盖本批 exec_s 的 p90 × 1.2（无老师参考题按 6 s），夹在 [plans, plans_max]
|
||||
@@ -95,11 +103,7 @@ def main():
|
||||
pre = torch.zeros(B, 3, 6, device=dev); pre[:, :r] = queue[:, :r]
|
||||
qe = q + queue.sum(1)
|
||||
with torch.no_grad():
|
||||
if (nsub > 0).any():
|
||||
cur = subs[:, 0]; dd = torch.where(kin.cont, wrap180(cur - qe), cur - qe).abs().max(-1).values
|
||||
adv = (nsub > 0) & ((dd < 5) | (tnow - sub_t > 1.5))
|
||||
subs = torch.where(adv[:, None, None], torch.cat([subs[:, 1:], subs[:, 1:]], 1), subs)
|
||||
nsub = torch.where(adv, nsub - 1, nsub); sub_t = torch.where(adv, torch.full_like(sub_t, tnow), sub_t)
|
||||
subs, nsub, sub_prev, sub_t = sub_advance(kin, subs, nsub, sub_prev, sub_t, qe.detach(), tnow)
|
||||
feat = features(kin, q, qp, T, R, pre, torch.full((B,), r, device=dev), subs[:, 0], (nsub > 0).float())
|
||||
with torch.autocast('cuda', dtype=torch.bfloat16, enabled=amp):
|
||||
act, val, sub = model.sample_k(feat, 1, z)
|
||||
@@ -108,8 +112,8 @@ def main():
|
||||
with torch.no_grad():
|
||||
need = sub[:, 12].float() > 0.5
|
||||
subs = q.detach()[:, None] + sub[:, :12].float().view(B, 2, 6) * 180
|
||||
nsub = torch.where(need, torch.full_like(nsub, 2), torch.zeros_like(nsub))
|
||||
inc = act.clamp(-1, 1) * step # [B,16,6]
|
||||
nsub = torch.where(need, torch.full_like(nsub, 2), torch.zeros_like(nsub)); sub_prev = q.detach().clone()
|
||||
inc = act.clamp(-1, 1) * astep # [B,16,6]
|
||||
commit = inc[:, r:r + 3]
|
||||
if r == 0: # 冷启动：先提交一块，下一轮才开始执行
|
||||
queue = commit; continue
|
||||
@@ -141,7 +145,9 @@ def main():
|
||||
with torch.autocast('cuda', dtype=torch.bfloat16, enabled=amp):
|
||||
mem = model.encode(X)
|
||||
l_bc = wta_loss(model, mem, Y)[0] if model.head == 'wta' else meanflow_loss(model, mem, Y)[0]
|
||||
v, s = model.heads(mem)
|
||||
v, _ = model.heads(mem)
|
||||
Xs = X.clone(); Xs[:, 68:75] = 0
|
||||
_, s = model.heads(model.encode(Xs))
|
||||
l_bc = l_bc + 0.1 * (v.float() - V).abs().mean() + 0.1 * (s.float() - SG).abs().mean()
|
||||
(C['w_bc'] * l_bc).backward()
|
||||
gn = torch.nn.utils.clip_grad_norm_(model.parameters(), C['grad_clip'])
|
||||
@@ -160,7 +166,7 @@ def main():
|
||||
res, _ = evaluate(model.eval(), kin, val_tasks, dev)
|
||||
res['it'] = it + 1; res['t'] = time.time()
|
||||
print('EVAL', json.dumps(res), flush=True); log.write(json.dumps({'eval': res}) + '\n'); log.flush()
|
||||
sd = {'model': model.state_dict(), 'cfg': mcfg, 'step': it + 1, 'fp': kin.fp, 'eval': res}
|
||||
sd = {'model': model.state_dict(), 'cfg': mcfg, 'step': it + 1, 'fp': kin.fp, 'eval': res, 'act_scale': act_scale}
|
||||
save_last(it + 1, res) # last 带优化器状态供续跑；best 不带（导出/门禁用）
|
||||
if res['success'] > best + 1e-4:
|
||||
best = res['success']; atomic_save(sd, os.path.join(a.out, 'best.pt'))
|
||||
|
||||
Reference in New Issue
Block a user