Deep Learning Academy · 课时

训练模式与评估模式

了解 model.train() 与 model.eval() 的区别

训练模式与评估模式 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。

A Model Has Two Faces

Your network behaves differently while learning than while being tested. PyTorch calls these two states train mode and eval mode. 🎭

Flip Into Train Mode

Before training you call model.train. This tells every layer to behave the way it should while the model is actively learning from data.

model.train()

Flip Into Eval Mode

Before validating or predicting you call model.eval. Now the network switches to its steady, deterministic behavior for honest measurement.

model.eval()

Dropout Cares

Dropout randomly zeros neurons during training to prevent overfitting. In eval mode it turns fully off so every neuron contributes to the prediction.

Batch Norm Cares Too

Batch norm uses each batch's statistics while training but switches to stored running averages at eval time for stable, consistent outputs.

It Is Only a Switch

These calls do not train or test anything by themselves. They simply flip a flag that tells dropout and norm layers which behavior to use.

Eval Is Not no_grad

A common mix-up: eval changes layer behavior, while no_grad stops gradient tracking. They solve different problems, so you usually use both together.

model.eval()
with torch.no_grad():
    pred = model(x)

Forgetting eval Hurts

If you validate without eval, dropout and batch norm keep their training behavior. Your scores turn noisy and look worse than the model really is.

Forgetting train Hurts

The reverse bites too: leave the model in eval and dropout never activates, so its regularization silently vanishes during training.

Build It Into the Loop

Make the switch a habit: set train before the training pass and eval before validation, every single epoch, so the model never stays in the wrong mode.

for epoch in range(epochs):
    model.train()
    # ... train ...
    model.eval()
    # ... validate ...

Small Call, Big Effect

Two tiny calls, yet they decide whether your metrics are trustworthy. Treat the mode switch as a non-negotiable part of every loop.

Quick Check

What does calling model.eval() actually change?

Recap

You learned the two modes: train activates dropout and batch stats, eval makes them deterministic. Switch every epoch and pair eval with no_grad. 🎉

常见问题解答

「训练模式与评估模式」课时是免费的吗？

是的 — 「训练模式与评估模式」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。

「训练模式与评估模式」这节课中我会学到什么？

了解 model.train() 与 model.eval() 的区别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。

「训练模式与评估模式」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