it's better to make optimizer as a class.

[xpqiu]

fastNLP/fastNLP/action/optimizor.py

Line 1 in 7c2f260

from torch import optim

[FengZiYjun]

We do not implement optimizers, only borrow them from pytorch. Currently a single function as an interface is enough. Making optimizer as a class is reasonable if it provides more functionalities.

[xpqiu]

Yes, but it's more convenient for users to choose optimizer by "selecting" the enumerated options instead of "typing" the name.

[xpqiu]

there is also a potential plan to implement some customized optimizers.

[choosewhatulike]

Rewrite a optimizer class may be redundant. In Tester or Trainer, we actually use optimizers in Pytorch. And further implemented optimizer can inherit class Optimizer from Pytorch.

But I agree that choose optimizer should avoid tying names, maybe we can change optimizer.py to import all of optimizers and implemented customized optimizer.

from torch.optim import *
from XXoptim import XXoptimizer
...

And users can "selecting" different optimizer now.

[xpqiu]

agree

[FengZiYjun]

all right