[Docs] after resume training, change param_scheduler is no effective

[zhangchao-s]

📚 The doc issue

in my work , I already train 48 epochs before, and I want to use the epoch_48.pth to continue my work to 56 epochs. but when I change param_scheduler in config, I see the config lr did not change. what problem is it ? How can i solve it

before change config:
param_scheduler= [
dict(
type='LinearLR',
start_factor=1.0 / 3,
by_epoch=False,
begin=0,
end=500),
dict(
type='MultiStepLR',
begin=0,
end=48,
by_epoch=True,
milestones=[32, 44],
gamma=0.1),
]

after change config(I want):
param_scheduler= [
dict(
type='LinearLR',
start_factor=1.0 / 3,
by_epoch=False,
begin=0,
end=500),
dict(
type='MultiStepLR',
begin=0,
end=56,
by_epoch=True,
milestones=[32, 44, 56],
gamma=0.1),
]

In epoch 48, the learning rate was 1e-5, but the configuration was changed and the training continued, and the rest of the epoch was still 1e-5, with no change.

Suggest a potential alternative/fix

No response

[zhangchao-s]

How do I make sure that when I resume training, I can reset the learning rate scheduler to take effect with the new configuration?