使用 Lightning 回调优化训练

使用 PyTorch Lightning 构建可扩展 AI 模型

Sergiy Tkachuk

Director, GenAI Productivity

什么是回调？

在训练关键阶段执行的函数
添加自定义操作而不使代码混乱
提升灵活性与可控性

回调概览

什么是回调？

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")

在训练各阶段添加自定义操作

Lightning 的 ModelCheckpoint 回调

按设定间隔自动保存模型
自选监控指标
仅保留最佳模型

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 的 EarlyStopping 回调

监控某个指标
当指标不再提升时停止训练

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

自定义并使用 Lightning 回调

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

Passons à la pratique !

使用 PyTorch Lightning 构建可扩展 AI 模型