多分类交叉熵
了解它为何需要原始 logits，而不是 softmax 结果
多分类交叉熵 是 CoddyKit 上的免费 Deep Learning Academy 课时。 这是第 3 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。
More Than Two Choices
When your model must pick one label from many, like a digit 0 through 9, you use cross-entropy, the standard multiclass classification loss. 🔢
One Logit Per Class
For C classes, your network outputs a vector of C raw scores called logits, one per possible class, for every sample in the batch.
Softmax Turns Scores Into Odds
Softmax exponentiates each logit and normalizes them so they sum to 1, turning raw scores into a clean probability distribution.
import torch
probs = torch.softmax(logits, dim=1)The Loss for Many Classes
PyTorch bundles softmax and the loss into nn.CrossEntropyLoss. It scores how much probability your model put on the correct class.
import torch.nn as nn
loss_fn = nn.CrossEntropyLoss()Pass Raw Logits Directly
This is the key rule: feed CrossEntropyLoss the raw logits, never a softmax output. It applies log-softmax internally for you.
loss = loss_fn(logits, labels)Why Not Softmax First
Applying softmax yourself, then this loss, squashes the values twice. That double pass distorts gradients and quietly wrecks training.
Labels Are Class Indices
Your targets are plain integer indices, like 3 for class three, not one-hot vectors. PyTorch handles the lookup for you.
labels = torch.tensor([3, 0, 7])Mind the Output Shape
Logits have shape batch by classes, while labels have shape batch only. A shape mismatch here is the most common cross-entropy error.
It Is Built From Two Pieces
Under the hood, CrossEntropyLoss equals log-softmax followed by negative log-likelihood. Fusing them keeps the math numerically stable.
Predict With Argmax
To get the final answer, take the argmax over the logits. The largest score is the predicted class, no softmax needed.
preds = logits.argmax(dim=1)Smooth Labels for Generalization
Setting label_smoothing nudges targets slightly away from a hard 1, which discourages overconfidence and often improves test accuracy.
loss_fn = nn.CrossEntropyLoss(label_smoothing=0.1)Quick Check
You are using nn.CrossEntropyLoss for 10 classes. What should the model's final output be?
Recap: One Winner Per Sample
For multiclass tasks you feed raw logits and integer labels to CrossEntropyLoss, which softmaxes internally and scores the right class. Argmax gives the prediction. 🏆
常见问题解答
「多分类交叉熵」课时是免费的吗？
是的 — 「多分类交叉熵」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。
「多分类交叉熵」这节课中我会学到什么？
了解它为何需要原始 logits，而不是 softmax 结果 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 3 节课，共 4 节。
「多分类交叉熵」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。