확률을 위한 Softmax

A classifier's last layer spits out raw scores, one per class. These unbounded numbers are called logits, and they are hard to read directly. 🔢

Softmax turns a vector of logits into a clean probability distribution over all your classes. Now each score becomes a chance.

Softmax promises two things: every output is between 0 and 1, and all of them sum to one. That is exactly what a distribution needs.

Softmax exponentiates each logit, then divides by the total. The exponential makes bigger logits dominate, sharpening the winner.

Softmax cares about the gaps between logits, not their raw size. Adding the same constant to every logit leaves the probabilities unchanged.

Use sigmoid for one yes-or-no output. Use softmax when classes compete and exactly one of several should win.

To get the predicted class, take the argmax of the softmax output. That index is the class with the highest probability.

Always set the right dim so softmax normalizes across classes, not across the batch. A wrong axis quietly ruins your probabilities.

During training you usually do not apply softmax yourself. The cross-entropy loss expects raw logits and applies it internally.

Combining softmax and the log inside one function is more numerically stable. Doing both steps by hand can overflow or lose precision.

Dividing logits by a temperature before softmax controls confidence. Low values sharpen the peak; high values flatten it out.

Softmax converts logits into probabilities that sum to one, perfect for multiclass output. But let cross-entropy handle it during training. 🏁