Enhancing training with Lightning callbacks

Scalable AI Models with PyTorch Lightning

Sergiy Tkachuk

Director, GenAI Productivity

What are callbacks?

Functions executed at key stages of training
Add custom actions without cluttering code
Enhance flexibility and control

callbacks overview

What are callbacks?

from lightning.pytorch.callbacks import Callback

class MyPrintingCallback(Callback):

    def on_train_start(self, trainer, pl_module):
        print("Training is starting")

    def on_train_end(self, trainer, pl_module):
        print("Training is ending")

Adding custom actions at various stages of training

Lightning ModelCheckpoint callback

Automatically saves model at specified intervals
Choose metric to track
Keep only the best model

from lightning.pytorch.callbacks
import ModelCheckpoint

checkpoint_callback = ModelCheckpoint(

    monitor='val_loss',

    dirpath='my/path/',

    filename='{epoch}-{val_loss:.2f}',

    save_top_k=1,

    mode='min'

)

¹ https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.callbacks.ModelCheckpoint.html

Lightning EarlyStopping callback

Monitor a metric
Stop training when the metric stops improving

from lightning.pytorch.callbacks
import EarlyStopping

early_stopping_callback = EarlyStopping(

    monitor='val_loss',

    patience=3,

    mode='min'

)

¹ https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.callbacks.EarlyStopping.html

Customizing and using lightning callbacks

from lightning.pytorch import Trainer
from lightning.pytorch.callbacks import EarlyStopping, ModelCheckpoint

checkpoint = ModelCheckpoint(
    monitor='val_accuracy',
    save_top_k=2,
    mode='max')

early_stopping = EarlyStopping(
    monitor='val_accuracy',
    patience=5,
    mode='max')


trainer = Trainer(max_epochs=50, callbacks=[checkpoint, early_stopping])

¹ https://lightning.ai/docs/pytorch/stable/common/trainer.html

Let's practice!

Scalable AI Models with PyTorch Lightning