Using multiple trainers vs. single trainer if max_epochs needs to change.

[MarioIshac]

❓ Questions and Help

Before asking:

I've looked at Issue with running multiple models in PyTorch Lightning which talks about using multiple trainers for multiple models.

What is your question?

I want to clarify when I should be using multiple trainers vs. one trainer if I'm trying to train a bunch of models (all of BaseModel, a custom subclass of pytorch_lightning.LightningModule) which only differ among their hyperparameters (and there hyperparameters are provided as hparams in the instantiation of BaseModel). The catch is that the max epochs is a hyperparameter in my case.

Code

Given that train_model is being called for each model I want to train, do I go this route:

def train_model(model: BaseModel):
    trainer = pytorch_lightning.Trainer(..., max_epochs=model.hparams.max_epochs)
    trainer.fit(model)

or this route:

trainer = pytorch_lightning.Trainer(...)

def train_model(model: BaseModel):
    trainer.max_epochs = model.hparam.max_epochs
    trainer.fit(model)

Everything else about the trainer remains unchanged after initialization (such as gpus, logger, early_stop_callback, etc, this is all part of the ...). Only max_epochs needs to change.

What's your environment?

• OS: Ubuntu 20.04
• Packaging: pip
• Version: 0.8.5

[github-actions]

Hi! thanks for your contribution!, great first issue!

[awaelchli]

@MarioIshac I would choose the first option, because the Trainer saves some state (e.g. in the loggers) that you probably don't want to reuse in the following model that will train.
Besides, I don't recommend setting attributes on the Trainer directly (like trainer.max_epochs = ...). These attributes are not meant to be modified from outside.

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!

[drozzy]

I wish there was a bit more clarification on this in the docs, as I had to google for this.

[awaelchli]

I'm sorry that you had to google it. Where did you find your answer? Here in github?
Any idea where would be a good place to mention this?

[drozzy]

Yeah your answer above.
I think good place for it would be here: https://pytorch-lightning.readthedocs.io/en/stable/trainer.html#
Something like "don't use the same trainer multiple times" etc.