ml-mobileone
ml-mobileone copied to clipboard
Conv2D in Stage0 (Beginning) not mentioned in paper
Hey, I just have a look at your implementation and the resulting reparametrized graph and see that you have a conv2D at the very beginning (stage0) of your network, which is not stated in the paper (neither the conv2d nor stage0, the paper begins with stage1)... Is there an issue ? Am I wrong ?
Besides that, counting the depthwise+pointwiseconv2d layer in your model yield to 42 layer in total (excluding the first one mentioned above).. According to your paper that should be 44 depthwise+pointwise conv2d layers (22 Blocks)... Where and why did you leave away some layers ? For the sake of runtime ?