Activation Functions, or the Lack of Them, under Computer Vision.

[WALTEROBILA]

When we were doing classification, we would use activation functions in the process of converting the predictions from logits, to probabilities, then finally to labels. For the "03_PyTorch_Computer_Vision" module, we only used argmax to convert the logits to labels. Any clarification on why this is the case?

[LuluW8071]

When we were doing classification, we would use activation functions in the process of converting the predictions from logits, to probabilities, then finally to labels. For the "03_PyTorch_Computer_Vision" module, we only used argmax to convert the logits to labels. Any clarification on why this is the case?

Process of converting the predictions from logits, to probabilities is done for Binary Classification mainly for loss_fn = nn.BCEWithLogitsLoss to find probs distribution over classes. We dont need such conversion process in Multiclass classification in "03_PyTorch_Computer_Vision" where there were 10 labels

Similar Discussion in here #314

[WALTEROBILA]

I'll have to differ with this. For the multiclass classification problem under Putting it all together timestamp 13:42:10, we used the torch.softmax activation to convert the logits into probabilities.

[LuluW8071]

Oh thats because we needed to find probability distribution of colored dots.

In the "03_PyTorch_Computer_Vision", the way model is used requires only the prediction of the class with the highest probability, not the full probability distribution over all classes. In such cases, applying the softmax and then taking the argmax might not provide any additional benefits over directly taking the argmax of the logits.

[WALTEROBILA]

This explanation clears it up. Thank you very much.