Deep Learning Academy · 课时

根据验证损失提前停止

在过拟合前停止训练

根据验证损失提前停止 是 CoddyKit 上的免费 Deep Learning Academy 课时。这是第 4 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 Deep Learning Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Deep Learning Academy 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Training Too Long Hurts

More epochs are not always better. Past a point the model starts memorizing noise, and validation results get worse. Early stopping ends training at the sweet spot.

Watch Validation Loss

The signal you track is validation loss. While it keeps falling the model is still generalizing better, so you let training continue.

The Turning Point

Eventually training loss keeps dropping but validation loss bottoms out and rises. That upward turn is the moment overfitting takes over.

Remember the Best Score

Keep a variable for the best validation loss seen so far. Each epoch you compare the new score against it to decide if progress was made.

best_loss = float('inf')

Count Epochs Without Gain

When an epoch fails to beat the best, increase a counter. A reset to zero happens the moment validation loss improves again.

if val_loss < best_loss:
    best_loss = val_loss
    counter = 0
else:
    counter += 1

Set a Patience

Patience is how many bad epochs you tolerate before quitting. A small value stops fast; a larger one rides out noisy dips.

patience = 5

Break When Patience Runs Out

Once the counter reaches patience, stop the loop. This break ends training before the model wastes time getting worse.

if counter >= patience:
    break

Save the Best Along the Way

Each time validation improves, write a checkpoint. When you finally stop, that saved file holds the strongest model, not the last one.

if val_loss < best_loss:
    torch.save(model.state_dict(), 'best.pt')

Add a Min Delta

Tiny wobbles can look like fake progress. A min_delta requires improvement bigger than a threshold before resetting the patience counter.

if val_loss < best_loss - min_delta:
    best_loss = val_loss

Restore the Best at the End

After the loop, load your saved checkpoint so you actually deploy the best model. Stopping early is pointless if you keep the worse final weights.

model.load_state_dict(torch.load('best.pt'))

A Free Form of Regularization

Early stopping costs nothing extra yet fights overfitting like a regularizer. It pairs nicely with dropout and weight decay for even steadier training.

Quick Check

What does the patience setting control in early stopping?

Recap

Track validation loss, count epochs without gain, stop once patience runs out, and restore the best checkpoint. Early stopping saves time and curbs overfitting. ⏹️

常见问题解答

「根据验证损失提前停止」课时是免费的吗？

是的 — 「根据验证损失提前停止」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Deep Learning Academy 课程的其余内容，请升级到 CoddyKit PRO。 Deep Learning Academy 课程共包含 4 节课。

「根据验证损失提前停止」这节课中我会学到什么？

在过拟合前停止训练你通过在浏览器中直接运行的动手代码来练习 Deep Learning Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Deep Learning Academy 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Deep Learning Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 4 节课，共 4 节。

「根据验证损失提前停止」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Deep Learning Academy 课中编写并运行代码吗？

能。每节 Deep Learning Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 Deep Learning Academy