about nn.Linear

[ZaidKhan-git]

I want to know why do we use out_features = 4 in multiclass classification and left out_features = 1 in binary classififcation?
Shouldn't it be 2, because we had two classes to predict. If this is right, then we are we considering out features in multi class classification to 4, I mean isn't labels for each set of feature be a single number (i.e its class)

[Muhammad-Ubaid-1166]

Why binary uses out_features=1 but multiclass uses out_features=num_classes

The label is always just 1 number in both cases (e.g. 0/1, or 0/1/2/3). What's different is how many numbers the model needs to output before it decides on that label.

Binary (out_features = 1)
Since there are only 2 classes, they're complements of each other. If P(class 1) = 0.91, then P(class 0) is automatically 1 - 0.91 = 0.09. So you only need 1 raw number → pass through sigmoid → that's your probability for class 1.

Think of it like a boolean flag: you don't store is_active and is_inactive separately, one implies the other.

Multiclass (out_features = num_classes)
With 4+ classes there's no complement trick — knowing the score for class 0 tells you nothing about classes 1, 2, 3. So the model has to output a separate raw score (logit) for every class. 4 classes → 4 numbers, e.g. [1.2, 0.3, -0.8, 3.1]. Pass through softmax → turns them into probabilities that sum to 1 → take argmax to get the predicted class.

So out_features=4 doesn't mean "4 labels" — it means "4 confidence scores, one per class." The model still collapses that down to a single predicted label at the end.

	Binary	Multiclass (4 classes)
out_features	1	4
Output shape	[batch, 1]	[batch, 4]
Activation	Sigmoid	Softmax
Prediction	Threshold 0.5	argmax
Label itself	1 number	1 number