自动扩充的应用

Get the coordinates of conv(e.g., 3x3) and featuremap(e.g., 5x5)

• c_x_y: relative coordinates(1,18) of conv
• p_x_y: coordinates(1, 18, 5, 5) of featuremap

# c_x_y: relative coordinates of conv
# p_x_y:
kernel_size = 3
r = kernel_size // 2
c_x, c_y = torch.meshgrid(torch.linspace(-r, r, 2*r+1)
                       torch.linspace(-r, r, 2*r+1))
c_x_y = torch.cat((x,reshape(-1,1), y.reshape(-1,1)),dim=1)

h=5
w=5
p_x, p_y = torch.meshgrid(torch.linspace(1, h, h),
                          torch.linspace(1, w, w))
p_x = p_x.expand(1, kernel_size * kernel_size, h, w)
p_y = p_y.expand(1, kernel_size * kernel_size, h, w)
p_x_y = torch.cat((p_x, p_y), dim=1)

mershgrid

combine conv and featuremap

• p: it's shape is same as p_x_y
• 

p = c_x_y.reshape(1, 2*kernel_size*kernel_size, 1, 1) 
 +
    p_x_y