[Need help] How to init an object of a class with position parameters in yaml config file? Also how to call a func to init an object in yaml file?

[shenmishajing]

🚀 Feature request

We can init an object with keyword parameters already, which is very helpful. An example:

  encoders:
      class_path: torch.nn.ModuleDict
      init_args:
          modules:
              table:
                class_path: torch.nn.TransformerEncoder
                init_args:
                    encoder_layer:
                        class_path: torch.nn.TransformerEncoderLayer
                        init_args:
                            d_model: 128
                            nhead: 128
                            activation: relu
                            batch_first: true
                    num_layers: 2

But, what should I do if I want to use a torch.nn.Sequential or other class requiring position parameters instead of ModuleDict? The init func of nn.Sequential:

    @overload
    def __init__(self, *args: Module) -> None:
        ...

    @overload
    def __init__(self, arg: 'OrderedDict[str, Module]') -> None:
        ...

    def __init__(self, *args):
        super(Sequential, self).__init__()
        if len(args) == 1 and isinstance(args[0], OrderedDict):
            for key, module in args[0].items():
                self.add_module(key, module)
        else:
            for idx, module in enumerate(args):
                self.add_module(str(idx), module)

And, how to use a func to init an object? For example, I have a func like this:

# encoder.py
def init_encoder(dim):
  return nn.Sequential(nn.Linear(dim, dim), nn.Relu())

What should I do if I want something like:

  encoders:
      class_path: torch.nn.ModuleDict
      init_args:
          modules:
              table:
                method_path: encoder.init_encoder
                init_args:
                    dim: 128

[mauvilsa]

what should I do if I want to use a torch.nn.Sequential

Variable positional arguments *args currently is not possible to specify in a config file. However, you can easily implement your Sequential class to support it, like:

import torch.nn

class Sequential(torch.nn.Sequential):
    def __init__(self, modules: list[Module]):
        super().__init__(*modules)

Then in a config file:

class_path: your.module.Sequential
init_args:
  modules:
  - class_path: module1
  - ...

how to use a func to init an object?

This has been requested in many issues. Right now I don't remember where else it has been requested to reference here. But it is not yet supported, and it is not a simple feature to implement. For the time being, this is only possible by using classes-from-functions.

[shenmishajing]

@mauvilsa Ok, got it. Thanks for your reply very much. But, is there any plan to support those two features, the features about init objects with position parameters and objects from function? Thanks for you gays' great efforts for this useful lib again.

[mauvilsa]

All features are potential future additions, unless stated otherwise. But note that the list of possible features is large, and I don't mean only the ones publicly seen in issues. There is also a private list of things that can be added/improved. What gets implemented and when depends on many factors.

[shenmishajing]

Thanks for your explaination and all you guys' great work. I can not wait until the day when I get the news that those features are gotten supported. :)

[mauvilsa]

An update. Providing to class_path the import path to a function or method was added in #513, and is already part of the latest releases. But it is only supported if the function must have the class as return type annotation.

Support for functions without return type annotation will not be added. Doing so would open the door to executing arbitrary functions from a config file. If you want to us a function in some public package that doesn't have a return annotation, consider opening a pull request in that project.

Regarding *args no update for now.

[mauvilsa]

Actually an update for *args. There is another issue requesting this exclusively, see #514. Since this issue mixes two topics I prefer to track this feature there. Thus, I will be closing this.