/tf_inverse_warp_op

The implementation of bilinear inverse warp operation at TensorFlow

Primary LanguageC++MIT LicenseMIT

tf_inverse_warp_op

The implementation of bilinear inverse warp operation at TensorFlow.

The custom operation supports CPU and GPU devices.

Test Environment

  • Linux (Ubuntu 16.04)
  • TensorFlow 1.9
  • Cuda 9.0

Build and Run

  • Commandmake and generate inverse_warp.so

  • Run example:

    import tensorflow as tf
from inverse_warp_op import inverse_warp

inp = tf.ones([1, 300, 600, 3], tf.float32)
flow = tf.ones([1, 300, 600, 2], tf.float32)
warp_res = inverse_warp(inp, flow)

    ​

API Doc

inverse_warp(
	x,
  	flow
)

Args:

  • x: A Tensor. Must be float32 types. The shape can be 3-D [height, width, channels] or 4-D [batch, height, width, channles].
  • flow: A Tensor. Must be float32 types. The shape can be 3-D [height, width, 2] or 4-D [batch, height, width, 2].

Returns:

A float32 Tensor with the same shape as input x.