depthwise "separable" convolution not implemented correctly?

[semin-park]

First of all, great work :)

I'm just a little bit concerned about your implementation of the depthwise separable convolution. AFAIK, it's supposed to be a depthwise convolution followed by a pointwise convolution.

What I am thinking of is something like this:

cin, cout = channels_in, channels_out

depthwise = nn.Conv2d(cin, cin, kernel_size=3, padding=(1,1), groups=cin)
pointwise = nn.Conv2d(cin, cout, kernel_size=1)

depthwise_separable_convolution = nn.Sequential(
    depthwise,
    pointwise
)

But in your implementation, you implemented it as a single convolution, which I believe is a depthwise convolution, but not a depthwise "separable" convolution.

[tonylins]

Please see the InvertedResidual class.