How many total numbers of filters are there in Conv2D? I have a question regarding filters used in Conv2D. How many total numbers of filters are there? Max number of filters I used is 64. If it is possible to use as many numbers(like 128, 256, 512, 1024) of filters what those filters depict,  what kind of feature it could extract? Where shall I study more about filters?
Does keras randomly generate the values of the filter matrix? or there any set of filters available? If so how many numbers of filters exits?
If my understanding of the filter is wrong please correct me!
 A: There's no upper limit on how many filters that could exist. A filter is characterized by its weights, and there are essentially any number of weights that could be inside a filter.
The purpose of backprop moves the model from a random filter initialization to a filter configuration that makes the loss smaller -- ideally, the loss that's small enough to solve your problem.
A: The filter values are learned.
Keras, along with every other deep learning software, does not randomly produce the values that go in a CNN filter. Through minimization of error of a particular kind defined by the “loss function“ you specify, the best parameters are discovered.
This is quite similar to how OLS linear regression finds the best parameters via $\hat{\beta}_{OLS}=(X^TX)^{-1}X^Ty$, except that the neural network does not have a nice, closed-form solution to the error minimization problem. (For that matter, neither does a logistic regression.)
Also, there is no upper bound on the number of filters you can use.
