Repetitive setting for model, dataset,...etc

[REIGN12]

Repetitive setting for model, dataset,...etc

[ZhiyuanChen]

Thank you for your recognition~
Config has puzzled me for years and I didn't come up with the idea behind CHANfiG until very recently.
This is also a really tricky question and will possibly take months and even years before I can find the best practice.

I intentionally let items() or ** only return value in current level, so one solution is to put common attributes at another level / higher level.
For example, your model has an encoder and a decoder, you would like to let them shares the same droupout setting but different model size.

class ModelConfig(Config):
    def __init__(self):
        self.encoder.num_layers = 4
        self.decoder.num_layers = 2
        self.dropout = 0.1

class Model(nn.Module):
    def __init__(config):
        self.encoder = (**config.model.encoder, **config.model)

If there is only a few common attributes, something like Optimizer(abc=config.b.x, **config.optimizer) should work just fine.

[ZhiyuanChen]

Today we released the v0.0.34, which added a Variable class that wraps a immutable object, e.g. int, float, str to a mutable object.
With this new feature, the same variable that is used in different position will have alternation together.

[ZhiyuanChen]

With the recent release of v0.0.36, we implemented get and set method of Variable and make OrderedDict calls set for Variable object. By doing so, we can have the following outputs:

class Config(chanfig.Config):
    def init(self):
        emb = chanfig.Variable(384)
        self.model.emb = emb
        self.data.emb = emb

c = Config()
print(c)

Config(
  (model): Config(
    (emb): 384
  )
  (data): Config(
    (emb): 384
  )
)

c.model.emb = 256
print(d)

DC(
  (model): Config(
    (emb): 256
  )
  (data): Config(
    (emb): 256
  )
)

So, in your case, you could create a variable and use it everywhere.
CHANfiG will sync them automatically.