Link_arguments on instantiate

implementation help
1

Hey there,

I am currently trying to set up a training pipeline. I am using config files for the parameters like:

data:
  class_path: "vv.data.CIFAR10Loader"
  init_args:
    train_val_split: 0.8
    seed: 42
    batch_size: 1024
    workers: 10
    verbose: True

and

model:
  class_path: "vv.models.Autoencoder"
  init_args:
    encoder:
      class_path: "vv.models.Encoder"
      init_args:
        hidden_layer_sizes: [1000, 500, 300]
        dropout_prob: 0.1
        batch_norm: True
    decoder:
      class_path: "vv.models.Decoder"
      init_args:
        hidden_layer_sizes: [300, 500, 1000]
        dropout_prob: 0.1
        batch_norm: True
    learning_rate: 0.001

i would like to link a property of my dataloader (data_size) that is created on instantiation to my model. In addition to the paramaters I provide in the config file for the model there also should be model.init_args.encoder.input_size and similar for the decoder. I would like these parameters to be infered from the data and I followed the Expert tutorial for extending the lighting CLI:

class CustomCLI(LightningCLI):
    def add_arguments_to_parser(self, parser: LightningArgumentParser) -> None:
        # Add early stopping as seperate argument
        parser.add_lightning_class_args(EarlyStopping, "early_stopping")
        parser.set_defaults({"early_stopping.monitor": "val_loss",
                             "early_stopping.patience": 5,
                             "early_stopping.verbose": False,
                             "early_stopping.mode": "min"})
        
        # Link the image size to the layer size of the model
        #print all class arguments of the parser
        input("Press Enter to continue...")
        parser.link_arguments("data.data_size",
                              "model.init_args.encoder.input_size",
                              apply_on="instantiate")
        parser.link_arguments("data.data_size",
                              "model.init_args.decoder.output_size",
                              apply_on="instantiate")

Unfortunately my encoder never receives the data_size but instead gets “None” (I checked in the init method of my model). The tutorial does not specify anything further and I am unable to find anything in this regard online. Could you help me out?

best regards,
Nikita