How exactly does Keras calculate the validation accuracy?

Question

After each epoch, Keras provides the following evaluations (depending on how the model is compiled):

1. train_accuracy
2. train_loss
3. validation_loss
4. validation_accuracy

Keras evaluates the performance of the model using the validation set at the end of each epoch. But how does Keras do this? Assume that we are performing binary classification and using a binary loss function. Assume that there are 100 images in the validation set.

QUESTION: Can the accuracy can be found by simply calculating accuracy for each item and then average them (/100)?

CONTEXT: I want to implement a custom loss function that calculates the loss towards reducing validation loss (increasing validation accuracy). Therefore, I need to understand how the accuracy is calculated for the validation set.

Answer 1

After each epoch, the model is applied to the features of the validation set (x_val) to predict the probability of the labels. Then, the probabilities are thresholded to generate the predicted label (y_pred). Finally, the accuracy is calculated as follows:

$$ Accuracy\ =\ \frac{Concordant\ samples}{Total\ samples} $$

, where concordant samples are those for which the predicted (y_pred) and true labels (y_val) agree.

Assume we have a binary classification task and 5 validation samples. If y_pred = [0, 1, 1, 0, 1] and y_val = [0, 1, 1, 1, 0], then there are 3 concordant samples, giving an accuracy of 3/5 = 60%.