The Global Convolution Network (GCN) Block is essentially a kx1 followed by 1xk convolution summed with a parallely computed 1xk followed by kx1 convolution. This results in a large kxk kernel with dense connections.

global_convolution_tf.py

class GCN(tf.Module):
    def __init__(self, filters, k=7):
        super(GCN, self).__init__()
        self.padding_l1 = tf.keras.layers.ZeroPadding2D(padding=((k - 1) // 2, 0))
        self.conv_l1 = tf.keras.layers.Conv2D(filters, kernel_size=(k, 1))
        self.padding_l2 = tf.keras.layers.ZeroPadding2D(padding=(0, (k - 1) // 2))
        self.conv_l2 = tf.keras.layers.Conv2D(filters, kernel_size=(1, k))

        self.padding_r1 = tf.keras.layers.ZeroPadding2D(padding=((k - 1) // 2, 0))
        self.conv_r1 = tf.keras.layers.Conv2D(filters, kernel_size=(1, k))
        self.padding_r2 = tf.keras.layers.ZeroPadding2D(padding=(0, (k - 1) // 2))
        self.conv_r2 = tf.keras.layers.Conv2D(filters, kernel_size=(k, 1))

    def __call__(self, x):
        x_l = self.padding_l1(x)
        x_l = self.conv_l1(x_l)
        x_l = self.padding_l2(x_l)
        x_l = self.conv_l2(x_l)

        x_r = self.padding_r1(x)
        x_r = self.conv_r1(x_r)
        x_r = self.padding_r2(x_r)
        x_r = self.conv_r2(x_r)

        x = x_l + x_r

        return x