I have a question concerning the computation of the accuracy of a NN. How exactly is it computed and implemented in python? I believe that it has never been mentioned in the course.
You can find examples of accuracy computations in various notebooks. Every time a model is trained, you get accuracy scores on both the training and test data. For example, check the Logistic Regression assignment in Course 1 Week 2. The definition of accuracy is simple: the fraction of predictions that agree with the labels, although in some cases they choose to express it as a percentage. You have the labels Y and you get the output of the model A. Then you have a function, e.g. predict in the Week 2 assignment, that converts the sigmoid values of A into 0 or 1 predictions. So in python, you can do something as simple as:
acc = np.mean(Y == predict(A))
That doesn’t actually use the same function signature as the real predict function, but suffices to make the point here.
1 Like
Thank you for clarification! Your explanation is very straightforward. However, in the programming assignments I cannot always comprehend the computational steps to derive the accuracy, especially in the Week 3 assignment. Here, the steps to compute the accuracy seem similar to parts of the cost function. Why is that?
The fundamental idea is always the same as I described above, but there are lots of ways you can express that in python code. In the Week 3 case, they use dot products. Well, think about it for a second:
If Y is the labels (either 0 or 1) and is a 1 x m vector and predict(A) is the predictions either 0 or one and also a 1 x m vector, then what happens if I compute this dot product:
Y \cdot predict(A)^T
1 x m dotted with m x 1 gives you a 1 x 1 or scalar output, which is this sum in math terms:
\displaystyle \sum_{i = 1}^{m} y_i * predict(a_i)
If either of the terms is 0, the product is zero. So that sum gives the number of cases in which the label is 1 and the prediction is also 1. So it is the number of correct predictions on cases in which the label is 1, right?
Now apply the same reasoning the (1 - Y) case. It just requires reading the code and understanding what it is doing and then thinking about what that means. 
1 Like
Thank you a lot. Your answers are always quite helpful!