Sir said we start from smaller sized filter and then larger .Why so?
I didn't get it why 32?
It is done to increase the information retained. For example, after every Conv or Max Pool operation, you’re decreasing the width and height of the image, effectively reducing the info you have. To retain this, we increase the number of filters and retain much of the lost info (transforming the image along the way).
Also, 32 is just a hyper-parameter, you can use any other number. It is convention to use a power of 2 though.
Hope this helps!