Rotational Equivariant Vector Field Networks (RotEqNet) for PyTorch

This is a PyTorch implementation of the method proposed in: Rotation equivariant vector field networks, ICCV 2017, Diego Marcos, Michele Volpi, Nikos Komodakis, Devis Tuia.

https://arxiv.org/abs/1612.09346

The original MATLAB implementation can be found at:

https://github.com/dmarcosg/RotEqNet

The goal of this code is to provide an implementation of the new network layers proposed in the paper. In addition we try to reproduce the results the MNIST-rot dataset to verify the implementation.

Example usage

from __future__ import division
from layers_2D import RotConv, VectorMaxPool, VectorBatchNorm, Vector2Magnitude, VectorUpsampling
from torch import nn


class MnistNet(nn.Module):
    def __init__(self):
        super(MnistNet, self).__init__()

        self.main = nn.Sequential(           
            RotConv(1, 6, [9, 9], 1, 9 // 2, n_angles=17, mode=1), #The first RotConv must have mode=1 
            VectorMaxPool(2),
            VectorBatchNorm(6),
            
            RotConv(6, 16, [9, 9], 1, 9 // 2, n_angles=17, mode=2), #The next RotConv has mode=2 (since the input is vector field)
            VectorMaxPool(2),
            VectorBatchNorm(16),
            
            RotConv(16, 32, [9, 9], 1, 1, n_angles=17, mode=2),
            Vector2Magnitude(), #This call converts the vector field to a conventional multichannel image/feature image
            
            nn.Conv2d(32, 128, 1),
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.Dropout2d(0.7),
            nn.Conv2d(128, 10, 1),
            
        )

    def forward(self,x):
        x = self.main(x)
        return  x

Dependencies

The following python packages are required:

torch
numpy
scipy

To download and setup the MNIST-rot dataset, cd into the MNIST-folder and run:

python download_mnist.py
python make_mnist_rot.py

To run the MNIST-test:

python mnist_test.py

Results from the MNIST-rot test

The MNIST-experiment in the orignial paper was obtained by:

• training on 10 000 images from the MNIST-rot dataset + applying random rotation as augmentation
• validating on 2000 images from the MNIST-rot dataset
• testing on 10 0000 images from the MNIST-rot dataset + with test-time augmentation as described in the paper

Using this implementation, we obtain a test accuracy of 1.2%, while the original paper reports 1.1%.

Known issues:

• The interpolation of filters (apply_transformation in utils.py) sometimes causes "CUDA runtime error 59". This error disappears when we use "torch.gather" to collect the samples, but this does reduce the best test error rate to ~3%.

Contact

Anders U. Waldeland
Norwegian Computing Center
anders@nr.no