What is the best practice to define a model?

[RafailFridman]

I want to try many architectures and it is not convenient to copy all the loading/train/val/test logic from file to file. Currently, I create a "container", where I define all the common logic, and then I inherit that class. It looks like this:

class ModelContainer(pl.LightningModule):
    
    def __init__(self,hparams):

        super().__init__()

    def prepare_data(self):
    ....
    def configure_optimizers(self):
    ....
    def train_dataloader(self):
    ....

And in another file I create a model, where I redefine only a forward method:

class MyCoolNet(ModelContainer):
    def __init__(self, hparams):
        super().__init__(hparams)

    def forward(self, x):
    ...

Am I doing it the right way?

[williamFalcon]

Overall:

class Root(LightningModule):
  def __init__
  def forward
  def training_step
  def train_dataloader
  def val_dataloader

class ModelA(Root):
   def training_step
   def validation_step

class ModelB(Root):
   def training_step
   def validation_step

In ModelA, B you only want to change training_step (ie the inner loop).

https://pytorch-lightning.readthedocs.io/en/0.7.1/introduction_guide.html#child-modules

I recommend to read this full guide.