Self-Attention GAN 中的 self-attention 机制 ： self attention 看成是 feature map 和它自身的转置相乘，让任意两个位置的像素直接发生关系，这样就可以学习到任意两个像素之间的依赖关系，从而得到全局特征了。

self-attention.py

classSelf_Attn(nn.Module):

""" Self attention Layer"""

  def__init__(self,in_dim,activation):

    super(Self_Attn,self).__init__()

    self.chanel_in = in_dim

    self.activation = activation

    self.query_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim// 8, kernel_size= 1)

    self.key_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim// 8, kernel_size= 1)

    self.value_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim , kernel_size= 1)

    self.gamma = nn.Parameter(torch.zeros( 1))

    self.softmax = nn.Softmax(dim= -1) #

  def forward(self,x):

"""

inputs :

x : input feature maps( B X C X W X H)

returns :

out : self attention value + input feature

attention: B X N X N (N is Width*Height)

"""

    m_batchsize,C,width ,height = x.size()

    proj_query = self.query_conv(x).view(m_batchsize, -1,width*height).permute( 0, 2, 1) # B X CX(N)

    proj_key = self.key_conv(x).view(m_batchsize, -1,width*height) # B X C x (*W*H)

    energy = torch.bmm(proj_query,proj_key) # transpose check

    attention = self.softmax(energy) # BX (N) X (N)

    proj_value = self.value_conv(x).view(m_batchsize, -1,width*height) # B X C X N

    out = torch.bmm(proj_value,attention.permute( 0, 2, 1) )

    out = out.view(m_batchsize,C,width,height)

    out = self.gamma*out + x

    return out,attention