pytorch-MobileNet

Simple Code Implementation of "MobileNet" architecture using PyTorch.

For simplicity, i write codes in ipynb. So, you can easliy test my code.

Last update : 2018/12/19

Contributor

• hoya012

Requirements

Python 3.5

numpy
matplotlib
torch=1.0.0
torchvision

Usage

You only run MobileNet-pytorch.ipynb. For test, i used CIFAR-10 Dataset and resize image scale from 32x32 to 224x224. If you want to use own dataset, you can simply resize images.

depthwise convolution and other blocks impelemtation.

In MobileNet, there are many depthwise convolution operation. This is my simple implemenatation.

depthwise convolution operation

class depthwise_conv(nn.Module):
    def __init__(self, nin, kernel_size, padding, bias=False, stride=1):
        super(depthwise_conv, self).__init__()
        self.depthwise = nn.Conv2d(nin, nin, kernel_size=kernel_size, stride=stride, padding=padding, groups=nin, bias=bias)

    def forward(self, x):
        out = self.depthwise(x)
        return out

depthwise block

class dw_block(nn.Module):
    def __init__(self, nin, kernel_size, padding=1, bias=False, stride=1):
        super(dw_block, self).__init__()
        self.dw_block = nn.Sequential(
            depthwise_conv(nin, kernel_size, padding, bias, stride),
            nn.BatchNorm2d(nin),
            nn.ReLU(True)
        )
    def forward(self, x):
        out = self.dw_block(x)
        return out

1x1 block

class one_by_one_block(nn.Module):
    def __init__(self, nin, nout, padding=1, bias=False, stride=1):
        super(one_by_one_block, self).__init__()
        self.one_by_one_block = nn.Sequential(
            nn.Conv2d(nin, nout, kernel_size=1, stride=stride, padding=padding, bias=bias),
            nn.BatchNorm2d(nout),
            nn.ReLU(True)
        )
    def forward(self, x):
        out = self.one_by_one_block(x)
        return out