Kcyganik small changes v03 #185
Conversation
|
The test will fail until SebChw/ART-Templates#33 is merged |
art/steps.py
Outdated
| @@ -298,6 +298,10 @@ def validate(self, trainer_kwargs: Dict): | |||
| Args: | |||
| trainer_kwargs (Dict): Arguments to be passed to the trainer for validating the model. | |||
| """ | |||
|
|
|||
| if "datamodule" not in trainer_kwargs.keys(): | |||
There was a problem hiding this comment.
How does this solves the issue of setup? I run EvaluateBaselines step from tutorial and setup was run 6 times for 3 baselines. So you run setup to be able to pass train dataloader. Next lightning runs it too.
There was a problem hiding this comment.
Actually this if inside validate and test seems not necessary. If "datamodule" is not in kwargs it means that a dataloader must be. If dataloader is inside trainer kwargs it means that setup must have been run already.
| @@ -395,6 +402,8 @@ def do(self, previous_states: Dict): | |||
| Args: | |||
| previous_states (Dict): previous states | |||
| """ | |||
| # Running setup for ml_train with pure train dataloader | |||
| self.datamodule.setup(stage=TrainingStage.TRAIN.value) | |||
There was a problem hiding this comment.
This doesn't scale well to new steps. My idea would be like this
add get_trainloader() Function to the base with try catch statement. If it suceeds it just return train_dataloader if it fails it runs datamodule setup function and then returns train_dataloader.
Something like this:
def get_trainloader():
try:
return self.datamodule.get_trainloader()
except :
self.datamodule.setip()
return self.datamodule.get_Trainloader()
But this is just my idea maybe it can be done in other way
No description provided.