Harmonic Networks: Deep Translation and Rotation Equivariance

This code requires Tensorflow version 1.0

This code accompanies the paper Harmonic Networks: Deep Translation and Rotation Equivariance

Authors: Daniel E. Worrall, Stephan J. Garbin, Daniyar Turmukhambetov, and Gabriel J. Brostow.

Watch the video

1 Running the code

To run code for a specific experiment, run the file run_<myscript>.py in the relevant folder.

2 Using harmonic convolutions in your code

The core functions for harmonic convolutions can be found in harmonic_network_ops.py. However, the best way to use these operations is via harmonic_network_lite.py. This contains the following functions:

• conv2d
• batch_norm
• non_linearity
• mean_pool
• sum_magnitudes
• stack_magnitudes

Each function takes in a 6D tensor with dimensions: minibatch size, height, width, num rotation orders, num complex channels, num channels. For instance, a real tensor with 16 items of height 128 and width 128, 2 rotation orders and 5 channels would have shape [16,128,128,2,1,5]. Whereas a complex tensor with the same parameters would be of shape [16,128,128,2,2,5].