Supervise la precisión durante el entrenamiento

A falling loss feels good, but it is abstract. To know if your model is truly useful you also want accuracy, the share of predictions it gets right. 🎯

Your model outputs raw scores per class. To pick a prediction you take the index of the largest score with argmax along the class dimension.

Now compare each predicted class to the real label. The result is a boolean tensor where True marks every example the model got right.

Booleans add up like ones and zeros, so summing the matches gives a count of correct predictions in this batch.

Divide correct predictions by the batch size and you get accuracy, a number from zero to one that is easy to read as a percentage.

Per-batch accuracy is noisy, so keep running totals of correct and seen examples. Their ratio gives a stable epoch accuracy.

At the end of an epoch compute the overall rate and print it beside the loss. Now you see both how confident and how correct the model is.

When you only measure, wrap scoring in no_grad. It stops PyTorch from tracking operations you never plan to backpropagate, saving memory.

On imbalanced data accuracy can fool you: always guessing the common class scores high yet learns nothing. Pair it with loss and other metrics.

One epoch's number means little; the trend is everything. Rising accuracy with falling loss is the clear sign your training is on track.

Tracking accuracy turns training from a black box into something you can watch. Each epoch you get honest feedback on real progress.

How do you turn class scores into a single predicted label?

You learned to track accuracy: argmax the scores, compare to labels, count correct, and divide. Watch the trend alongside loss. 🎉