Files
superspider-rl/scripts/superspider_blind_to_teacher.py

57 lines
2.9 KiB
Python

"""把【盲走】ckpt 扩成【teacher(带高度图)】的形状,扫描那部分权重置零。
.venv/bin/python scripts/superspider_blind_to_teacher.py <blind.pt> <out.pt>
为什么可以这么搬:teacher 的观测**前 65 维与盲走完全一致**,height_scan 是追加的
(actor 在末尾 [65:318];critic 在中间 [65:318],特权项挪到 [318:393])。
把对应列拷过去、扫描列置零 ⇒ **扩出来的网络与原网络逐点等价**(零权重的输入不贡献),
teacher 一开始就会走路,全部训练量用在"学会看高度图"上,不用从零学走路。
(实测:从零训的 teacher 到 2400 迭代平地只能走 0.58m。)
扫描部分的归一化统计置为 mean 0 / var 1 —— 不能沿用不存在的旧统计。
"""
import sys
import torch
BLIND, OUT = sys.argv[1], sys.argv[2]
A_BASE, SCAN, C_PRIV = 65, 253, 75 # 本体维 / 扫描维 / critic 特权维
ck = torch.load(BLIND, map_location="cpu", weights_only=False)
def widen(sd, old_dim, new_dim, insert_at, n_ins):
"""在第一层权重与归一化统计的 `insert_at` 处插入 n_ins 个零通道。"""
out = dict(sd)
w = sd["mlp.0.weight"]
assert w.shape[1] == old_dim, (w.shape, old_dim)
nw = torch.zeros((w.shape[0], new_dim), dtype=w.dtype)
nw[:, :insert_at] = w[:, :insert_at]
nw[:, insert_at + n_ins:] = w[:, insert_at:]
out["mlp.0.weight"] = nw
for key, fill in (("_mean", 0.0), ("_var", 1.0), ("_std", 1.0)):
k = "obs_normalizer." + key
if k not in sd: continue
v = sd[k]
nv = torch.full((1, new_dim), fill, dtype=v.dtype)
nv[:, :insert_at] = v[:, :insert_at]
nv[:, insert_at + n_ins:] = v[:, insert_at:]
out[k] = nv
return out
# actor: 扫描接在末尾
ck["actor_state_dict"] = widen(ck["actor_state_dict"], A_BASE, A_BASE + SCAN, A_BASE, SCAN)
# critic: 扫描插在本体项之后、特权项之前
ck["critic_state_dict"] = widen(ck["critic_state_dict"], A_BASE + C_PRIV,
A_BASE + SCAN + C_PRIV, A_BASE, SCAN)
# 优化器的 Adam 动量缓存还是旧形状 (512,65),直接沿用会在第一次 step 就炸:
# RuntimeError: size of tensor a (65) must match tensor b (318)
# train.py 不传 load_cfg 会连优化器一起加载,所以这里把逐参数状态清空,
# 让 Adam 按新形状惰性重建(param_groups 里的超参保留)。
if "optimizer_state_dict" in ck and isinstance(ck["optimizer_state_dict"], dict):
ck["optimizer_state_dict"]["state"] = {}
ck["iter"] = 0 # 计数归零,免得 max-iterations / 监控条件被偏移坑到
torch.save(ck, OUT)
print("actor mlp.0.weight ->", tuple(ck["actor_state_dict"]["mlp.0.weight"].shape))
print("critic mlp.0.weight ->", tuple(ck["critic_state_dict"]["mlp.0.weight"].shape))
print("扫描列权重全零:", bool(ck["actor_state_dict"]["mlp.0.weight"][:, A_BASE:].abs().max() == 0))
print("->", OUT)