EMD: Earth Mover Distance

It contains the Earth Mover Distance function with CUDA implementation for PyTorch as well as Tensorflow.

Test Entire Package:

chmod +x test.sh\ ./test.sh

Some Important Points about usage:

1. Both Inputs should have dimension (Batch x Number of Points x Dimension)
2. Both Inputs should be defined on GPU + CUDA devices.

EMD with PyTorch:

Copy the [emd_torch] folder to desired location.

Compile the CUDA code

cd emd_torch\ sudo python3 setup.py install

Sample to Test the EMD with PyTorch

import torch\ from emd import EMDLoss\ cost = EMDLoss()\ p1, p2 = torch.rand(32,1024,3).cuda().double(), torch.rand(32,1024,3).cuda().double()\ p1.requires_grad = True\ p2.requires_grad = True\ loss_val = cost(p1, p2)\ loss_val.backward()\ print(loss_val)\ print(p1.grad, p2.grad)

EMD with Tensorflow:

Copy the [emd_tf] folder to desired location.\ Change the import path in tf_util_loss.py.

Compile the CUDA code

cd emd_tf/pc_distance\ make -f makefile

Sample to Test the EMD with Tensorflow

import tensorflow as tf\ from emd_tf import tf_util_loss\ loss_val = tf_util_loss.earth_mover(xyz1, xyz2)\ loss_val = sess.run(loss, feed_dict={xyz1: xyz1_numpy, xyz2: xyz2_numpy})\

[Note: xyz1, xyz2 should be placeholders defined on GPU device]