The 3x3 stacking argument is about parameters as well as receptive field. Three 3x3 layers use 27 weights per channel pair, against 49 for a single 7x7 layer, for the same reach.
Implement stacked_params(n, c, k), the weight count of n convolutions with c channels each.
Multiply the number of layers, the kernel area and the square of the channel count.
No results yet
Run your code to check it against the examples, or submit to run every test.