Deep Learning Academy · 课时

带 Logits 的二元交叉熵

适用于二分类问题的稳定损失函数

带 Logits 的二元交叉熵 是 CoddyKit 上的免费 Deep Learning Academy 课时。这是第 2 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。

Two Classes, One Decision

When your model answers a yes or no question like spam or not spam, you reach for binary cross-entropy, the loss built for two-class problems. ✅

From Score to Probability

A model first outputs a raw score called a logit. The sigmoid function squashes that logit into a probability between 0 and 1.

import torch
prob = torch.sigmoid(logit)

What Cross-Entropy Rewards

Cross-entropy gives a small loss when the predicted probability is close to the true label, and a large loss when it is confidently wrong.

Confident and Wrong Hurts

Predicting 0.99 when the true label is 0 produces a huge loss. Cross-entropy punishes confident mistakes far more than uncertain ones.

The Naive Loss Choice

You could apply sigmoid yourself, then use nn.BCELoss on the probabilities. It works, but there is a safer way for real training.

import torch.nn as nn
loss_fn = nn.BCELoss()
loss = loss_fn(prob, target)

Feed Logits, Not Probabilities

The preferred loss is BCEWithLogitsLoss. It takes raw logits directly and applies the sigmoid internally for you.

import torch.nn as nn
loss_fn = nn.BCEWithLogitsLoss()
loss = loss_fn(logit, target)

Why Combine the Steps

Fusing sigmoid and the loss uses the log-sum-exp trick, which avoids overflow and gives stable gradients even at extreme logits.

Skip the Extra Sigmoid

Because BCEWithLogitsLoss applies sigmoid itself, your model's last layer should output raw logits with no sigmoid attached.

Targets Are 0 or 1

Pass float targets of 0.0 or 1.0 with the same shape as your logits. A mismatched shape is the classic BCE bug to watch for.

target = torch.tensor([1.0, 0.0, 1.0])

Sigmoid Only at Inference

During training the loss handles the squashing. To read a probability for a prediction, apply sigmoid to the logit at inference time.

Tilt the Balance with pos_weight

If positives are rare, the pos_weight argument scales up their contribution so the model stops ignoring the minority class.

Quick Check

You are doing binary classification and want numerical stability. What should you feed the loss?

Recap: Stable Yes or No

For two-class tasks you feed raw logits to BCEWithLogitsLoss, which squashes and scores in one stable step. Save sigmoid for reading probabilities later. 👍

常见问题解答

「带 Logits 的二元交叉熵」课时是免费的吗？

是的 — 「带 Logits 的二元交叉熵」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。

「带 Logits 的二元交叉熵」这节课中我会学到什么？

适用于二分类问题的稳定损失函数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 2 节课，共 4 节。

「带 Logits 的二元交叉熵」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