computation cost is wrong

[Xiaohong-Deng]

Instead of Df * Df * M * N * Dk * Dk, it should be Dg * Dg * M * N * Dk * Dk with Dg replacing Df.

Same applies to depthwise and pointwise convolution.

[mvirgo]

Ended up having a decently long discussion with another Udacian on why this is the case :)

Using Df is actually following the notation used in the paper - check out section 3.1 in the MobileNets paper, for instance. Per the paper, this is actually because they note they assume the output and input dimensions are equal (which is done with stride of 1 and padding). However, if this assumption isn't given, it looks like your computation would be correct.

We've now updated this in the notebook as such!