Issue in fitting model and finding optimal learning rate parameter

[geeksouvik]

following is the error: NotImplementedError: `val_dataloader` must be implemented to be used with the Lightning Trainer

LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
<ipython-input-11-263e8be26564> in <module>()
      3     tft,
      4     train_dataloader=train_dataloader,
----> 5     val_dataloaders=val_dataloader,
      6 )

11 frames
/usr/local/lib/python3.7/dist-packages/pytorch_lightning/core/hooks.py in val_dataloader(self)
    590             will have an argument ``dataloader_idx`` which matches the order here.
    591         """
--> 592         raise NotImplementedError("`val_dataloader` must be implemented to be used with the Lightning Trainer")
    593 
    594     def predict_dataloader(self) -> EVAL_DATALOADERS:

NotImplementedError: `val_dataloader` must be implemented to be used with the Lightning Trainer

trainer.fit(
    tft,
    train_dataloader=train_dataloader,
    val_dataloaders=val_dataloader,
)

[furbyhaxx]

I think it's a problem with pytorch-lightning==1.5.0.
Had this problem too with code that worked before recreating my venv and the difference was version 1.5.0 release on 2. Nov.
Switched to 1.4.9 and it worked again.
Checked 1.5.0rc1 and it did not work either.

[paapu88]

Thanks!!! So typical with pytorch-lightning. I dislike it and its video documentations but its best what you can get...

[geeksouvik]

Yea.. right it's the problem with the 1.5.0 version. downgraded the package and it's working.

[ilkerkesen]

I realized that you cannot use self.x_dataloader() methods inside the LightningModule implementation anymore. I need to call self.val_dataloader() method to post-process the model predictions. I guess this has been changed with the new version. Right now, you can access the datamodule object with datamodule field within the lightning module implementation (e.g. self.datamodule, self.datamodule.val_dataloader() or self.datamodule.val_data).

[gsoykan]

as of version 1.7.7, here is what I'm doing similar to your solution =>
self.trainer.datamodule.train_dataloader()

[ananthsub]

@geeksouvik @paapu88 @ilkerkesen :

the self.x_dataloader() are optional methods defined in the LightningModule, in case your module and data are not separable. The default implementation of these methods was changed from logging a warning message that these funtions should be implemented to raising a NotImplementedError to be explicit in case you calling something which you haven't defined.

Though the naming is similar, this is very different from accessing an already instantiated dataloader which you've passed to the trainer outside of the LightningModule's definition. If your model depends on the data, I'd recommend making it an input to your LightningModule constructor, and then pass that to the Trainer for fit and test

[sherandoresearch]

I am new to lightning and trying to get up to speed. I am using it with pytorch-forecasting and attempting to run this example: https://pytorch-forecasting.readthedocs.io/en/latest/tutorials/stallion.html

At step 8 I get the Not Implemented Error for the val_dataloader. Sorry to ask for a more explicit explanation, but where are you suggesting I include the loaders? at pl.Trainer?