What if my validation accuracy is 1 but the loss is not 0?

Question

you can see it on this picture

I was wondering if it can be possible

Answer 1

Cross entropy (or logloss) operates on class scores (probabilities). Zero loss can be achieved if your model only yields 1.0 and 0.0 probabilities, and does so correctly. Realistically, you'd likely receive some intermediate values.

Accuracy is treshold sensitive. In simple cases like yours, prediction used for calculating accuracy is likely just the argmax of the class probability list - in a binary case, a class that exceeds 0.5 probability.

Thus, as you can see, a perfect accuracy may still accompany an imperfect loss (which is basically a measure of how close the predicted class probabilities are to the 'true' 0.0 and 1.0).

As it was already mentioned, the fact your validation metrics are better than training ones is a much bigger concern.

Answer 2

This is possible - validation accuracy is the total number of correct predictions divided by the total number of predictions on your validation set. Depending on your validation dataset, it's possible to have perfect prediction (e.g. a validation set of a single entry that is correctly predicted every time). The thing that looks suspicious is that you're getting perfect validation accuracy from the very first epoch.