fix(config): --set 拦下不存在的键,并支持续训时重开学习率曲线

用户反馈"聚类配置里改不了 lr/weight_decay",实测是两件事:

1. --set 的键打错字母会凭空造出一个谁也不读的键,训练照跑但改动无效。
   实例:simclr.optmizer.weight_decay=0.0(optimizer 少个 i)被静默接受。
   现在 apply_overrides 默认只允许覆盖已存在的键,报错时给出同级可用键
   与最接近的拼写("是不是想写 'optimizer'?")。需要新增键时可 strict=False。

2. lr 其实生效了,但从 2229/2500 续训时 cosine 只剩满值的 3%,
   新设的 1e-3 落到 2.87e-5。新增 restart_schedule 开关:按剩余轮数重开曲线,
   新 lr 从满值开始;SimCLR 与检测训练都支持。续训日志补上调度进度,
   并在 lr 不足 initial_lr 的 10% 时警告并给出两种解法。

顺带:严格校验立刻暴露出 configs/cluster/film.yaml 里根本没有 scatter 段
(代码 cfg.get("scatter", {}) 兜底走默认值,所以功能正常但配置项形同虚设),已补上。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Shaun Zhou
2026-09-07 19:30:17 +08:00
co-authored by Claude Opus 5
parent 506fde0290
commit 92c88b785c
8 changed files with 165 additions and 7 deletions
+36 -1
View File
@@ -37,9 +37,13 @@ def test_parse_override_accepts_yaml_style_scalars():
def test_apply_overrides_creates_missing_nodes():
cfg = apply_overrides({}, ["a.b.c=1"])
"""非严格模式下仍可新建嵌套节点;严格模式(load_config 默认)会拦下未知键。"""
cfg = apply_overrides({}, ["a.b.c=1"], strict=False)
assert cfg == {"a": {"b": {"c": 1}}}
with pytest.raises(KeyError, match="不存在"):
apply_overrides({}, ["a.b.c=1"])
def test_load_config_with_base(tmp_path):
(tmp_path / "base.yaml").write_text("train:\n lr: 0.1\n epochs: 10\n", encoding="utf-8")
@@ -152,3 +156,34 @@ def test_eval_config_matches_its_training_backbone():
("configs/train/resnet18.yaml", "configs/eval/resnet18.yaml"),
]:
assert load_config(train_cfg)["model"] == load_config(eval_cfg)["model"], eval_cfg
@pytest.mark.parametrize(
"override,expected_hint",
[
("simclr.optmizer.weight_decay=0.0", "optimizer"), # 少了个 i
("simclr.optimizer.lrr=1e-3", "lr"), # 多了个 r
("num_clusers=4", "num_clusters"), # 少了个 t
],
)
def test_set_rejects_typos_with_suggestion(override, expected_hint):
"""--set 打错字母会凭空造出没人读的键,训练照跑但改动无效——必须拦下并给出建议。"""
with pytest.raises(KeyError) as excinfo:
load_config("configs/cluster/film_simclr_resnet50.yaml", [override])
assert expected_hint in str(excinfo.value)
def test_set_accepts_existing_keys():
cfg = load_config(
"configs/cluster/film_simclr_resnet50.yaml",
["simclr.optimizer.lr=0.001", "simclr.optimizer.weight_decay=0.0", "num_clusters=7"],
)
assert cfg["simclr"]["optimizer"]["lr"] == 0.001
assert cfg["simclr"]["optimizer"]["weight_decay"] == 0.0
assert cfg["num_clusters"] == 7
def test_apply_overrides_can_add_keys_when_not_strict():
from utils.config import apply_overrides
assert apply_overrides({"a": {"b": 1}}, ["a.c=2"], strict=False) == {"a": {"b": 1, "c": 2}}
+33
View File
@@ -292,6 +292,39 @@ def test_resume_with_a_longer_schedule_rebuilds_the_lr_curve(tiny_image_dir, tmp
assert lr == pytest.approx(1e-3 * (1 + math.cos(math.pi * 10 / 30)) / 2, rel=1e-3)
def test_restart_schedule_gives_the_new_lr_full_value(tiny_image_dir, tmp_path):
"""续训到调度末段时,cosine 会把新 lr 乘上接近 0 的系数;
restart_schedule 按剩余轮数重开曲线,让新 lr 从满值开始。"""
import math
work_dir = tmp_path / "run"
_make_trainer(
tiny_image_dir, work_dir, epochs=9,
optimizer={"name": "adamw", "lr": 1e-3},
).fit()
def resumed(restart):
return SimCLRTrainer(
BACKBONES.build({"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]}),
_make_trainer(tiny_image_dir, work_dir, epochs=1).dataloader,
{"epochs": 10, "amp": False, "log_interval": 999, "save_interval": 1,
"scheduler": {"name": "cosine"}, "optimizer": {"name": "adamw", "lr": 1e-3},
"projection": {"hidden_channels": 16, "out_channels": 8},
"swanlab": {"enabled": False}, "restart_schedule": restart},
torch.device("cpu"),
work_dir,
resume=work_dir / "simclr_last.pth",
)
# 默认:沿用整条曲线,9/10 处只剩很小的系数
default_lr = resumed(False).optimizer.param_groups[0]["lr"]
assert default_lr == pytest.approx(1e-3 * (1 + math.cos(math.pi * 9 / 10)) / 2)
assert default_lr < 1e-3 * 0.1
# 打开开关:新 lr 从满值开始
assert resumed(True).optimizer.param_groups[0]["lr"] == pytest.approx(1e-3)
def test_resume_takes_hyperparams_from_config_not_checkpoint(tiny_image_dir, tmp_path):
"""续训时 lr / weight_decay / 总轮数一律以 yaml 为准,存档只提供进度。