pytorch lightning 기초 형태 정리

https://www.kaggle.com/code/shivanandmn/beginners-guide-to-pytorch-lightning/notebook

Beginners Guide to PyTorch-Lightning

 

 

This is basic look of pytorch-lighning¶

There are many reserved methods in the lighningmodules called hooks:

• configure_optimizers - this should return optimizer(Adam/SGD)
• training_step - training loop, takes batch and batch_idx as parameters
• validation_step-validation loop, takes batch and batch_idx as parameters
• testing_step- testing loop, takes batch and batch_idx as parameters

In [1]:

import pytorch_lightning as pl

In [2]:

class LightningModel(pl.LightningModule):
    def __init__(self):
        super().__init__()
    
    def forward(self, x):
        pass
  
    def configure_optimizers(self):
        pass
  
    def loss_fn(self, output, target):
        pass 
  
    def training_step(self):
        pass
  
    def validation_step(self):
        pass

• Pytorch-lightning has data module extension that structures your data preprocessing.
• This structure helps to read and understand code easily for everyone.
• It helps to reuse data across multiple projects even with complex data transform and multiple-GPU handling

Hooks:¶

• train_dataloader()
• val_dataloader()
• test_dataloader()

Above methods in lightning datamodule are dataloaders

• prepare_data(): Download and tokenize or do preprocessing on complete dataset, because this is called on single gpu if your using mulitple gpu, data here is not shared accross gpus.
• setup(): splitting or transformations etc. setup takes stage argument None by default or fit or test for training and testing respectively.

In [3]:

class LightningDataset(pl.LightningDataModule):
    def __init__(self):
        super().__init__()
  
    def prepare_data(self):
        pass
  
    def setup(self, stage=None):
        pass
  
    def train_dataloader(self):
        pass
  
    def val_dataloader(self):
        pass
  
    def test_dataloader(self):
        pass

• There are mainly two classes required to train model and test it on test data.
• Now, we can use pl.Trainer for instantiating trainer with many parameters like max_epochs, tpu_cores,num_gpus,cpu(default),logger etc
• trainer.fit, fitting the data on model

In [ ]:

dataset = LightningDataset()
model = LightningModel()
trainer = pl.Trainer(max_epochs=10)
trainer.fit(model=model, datamodule=dataset)