C3W2B ExpandingBlock in UNet

Course Q&A Generative Adversarial Networks (GANS) Apply Generative Adversarial Networks
1

Hi,

This post is just to clarify the structure of the ExpandingBlock of the UNet model in Pix2Pix, the question in fact is:

In the video you talk about transpose convolutions, in the expanding part of the model, which is very logical, because i want to expand the image for create a nxn resolution image, but in the code you don’t use a transpose convolution instead you use a normal 2d convolution (nn.Conv2D):

class ExpandingBlock(nn.Module):
    '''
    ExpandingBlock Class:
    Performs an upsampling, a convolution, a concatenation of its two inputs,
    followed by two more convolutions with optional dropout
    Values:
        input_channels: the number of channels to expect from a given input
    '''
    def __init__(self, input_channels, use_dropout=False, use_bn=True):
        super(ExpandingBlock, self).__init__()
        self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
        self.conv1 = nn.Conv2d(input_channels, input_channels // 2, kernel_size=2)
        self.conv2 = nn.Conv2d(input_channels, input_channels // 2, kernel_size=3, padding=1)
        self.conv3 = nn.Conv2d(input_channels // 2, input_channels // 2, kernel_size=2, padding=1)
        if use_bn:
            self.batchnorm = nn.BatchNorm2d(input_channels // 2)
        self.use_bn = use_bn
        self.activation = nn.ReLU()
        if use_dropout:
            self.dropout = nn.Dropout()
        self.use_dropout = use_dropout

So, i want to know exactly why i shouldn’t use a transpose convolution instead (nn.ConvTranspose2d)??

Thanks so much, i have learned a lot with these courses.

2

Hi @Daniel_Mendoza_Castr

You can use both. While nn.ConvTranspose2d is effective, you can sometimes see artifacts like checkerboard patterns because of uneven overlap of the kernels. However, in this case, upsampling increases the spatial dimensions by interpolation, and then a regular convolution is applied to learn features on the upsampled image.

Hope it helps! Feel free to ask if you need further assistance.

3 Likes