In the paper mobilenet, the author uses depthwise convolution and pointwise convolution instead of conventional convolution to reduce the number of parameters.I used the same alternative method on my own model as follows：
def conv2d(in_planes, out_planes, kernel_size, stride, padding=1, dilation=1, bias=True):
return nn.Sequential(
nn.Conv2d(in_planes, in_planes, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation,
groups=in_planes,
bias=True),
nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, dilation=dilation,
bias=True),
)
def deconv2d(in_planes, out_planes, kernel_size, stride, padding=1, dilation=1, bias=True):
return nn.Sequential(
nn.ConvTranspose2d(in_planes, in_planes, kernel_size=kernel_size, stride=stride, padding=padding,
dilation=dilation,
groups=in_planes,
bias=True),
nn.ConvTranspose2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, dilation=dilation,
bias=True),
)
However, as a result, the training parameters were reduced, but the speed became a bit slower.
Can someone tells me what skills are needed to speed up depthwise convolution?