Week 1 Programming Assignment 2 Part 1 only tells us what to configure for the ConvNet. But it does not inform us about the reason behind the prescribed configuration settings. For example, why padding is set to 3? why use Conv2D with 32 7x7 filters and stride of 1?
It would be good to know the reason. Thank you in advance
This type of hyperparameter choice is made through experience and experimentation. There is no magic here. We will see many examples as we go through the courses here of particular networks that are good at solving particular types of problems. When you have a new problem you want to solve, you look through examples and start with an architecture that works for a problem that you hope is of roughly the same complexity as your new problem. Then you have to adjust and tune from there.
There are also “libraries” of pretrained networks for various types of image processing and other tasks. Here is the library that TF/Keras provides. Prof Ng will discuss Transfer Learning in Week 2.
We’ll learn more about all this from Prof Ng as we go through Course 4. My suggestion is to “hold that thought” and continue on through the course. After you’ve heard and seen all that is discussed, feel free to open another discussion on any issues like this.