Define conv for generator. Args: x: Input. cnum: Channel number. ksize: Kernel size. Stride: Convolution stride.

Define conv for discriminator. Activation is set to leaky_relu. Args: x: Input. cnum: Channel number. ksize: Kernel s

Define deconv for generator. The deconv is defined to be a x2 resize_nearest_neighbor operation with additional gen_conv operation. Args:

Inpaint network. Args: x: incomplete image, [-1, 1] mask: mask region {0, 1} Returns: [-1, 1] as

Generate mask tensor from bbox. Args: bbox: tuple, (top, left, height, width) Returns: tf.Tensor: output with shape [1, H, W

Generate mask tensor from bbox. Returns: tf.Tensor: output with shape [1, H, W, 1]

Contextual attention layer implementation. Contextual attention is first introduced in publication: Generative Image Inpainting with Con

Tensorflow ops for computing flow to image.

Generate a random tlhw. Returns: tuple: (top, left, height, width)

Resize mask like shape of x. Args: mask: Original mask. x: To shape of x. Returns: tf.Tensor: resized mask

Test contextual attention layer with 3-channel image input (instead of n-channel feature).

Transfer flow map to image. Part of code forked from flownet.

Convert flow into middlebury color code image.

Tensorflow ops for highlight flow.

Convert image to edges.

Crop local patch according to bbox. Args: x: input bbox: (top, left, height, width) Returns: tf.Tensor: local patch