An empirical study of rotation equivariant neural networks

Charles Gallay, Michaël Defferrard, Nathanaël Perraudin

This work is a semester project done at EPFL as part of the master program of data science.

Goal

The goal of the project is to test different configuration in order to see the importance of being equivariant to rotation for a CNN.

As a first measure, classical CNN and equivariant to rotation CNN are compare on the task of image classification and more specifically on the CIFAR10 dataset.

Setup

1. Clone this repository.

   git clone [email protected]:cgallay/GraphSymmetries.git
   cd GraphSymmetries

2. Install the dependencies.

   pip install -r requirements.txt
   pip install -e .

Example

import torch.nn as nn

from graphSym.graph_conv import GridGraphConv
from graphSym.graph_pool import GraphMaxPool2d

class Net(nn.Module):
    """
    Network with Horizontal symetry
    """
    def __init__(self, input_shape=(32,32), nb_class=5):
        super().__init__()
        underlying_graphs = [['left', 'right'], ['top'], ['bottom']]
        conv1 = GridGraphConv(3, 30, merge_way='cat', underlying_graphs=underlying_graphs)
        conv2 = GridGraphConv(30, 60, merge_way='cat', underlying_graphs=underlying_graphs)
        
        pool1 = GraphMaxPool2d(input_shape=input_shape)
        out_shape = (16, 16)
        
        conv3 = GridGraphConv(60, 60, input_shape=out_shape, merge_way='cat', underlying_graphs=underlying_graphs)
        conv4 = GridGraphConv(60, nb_class, input_shape=out_shape, merge_way='mean', underlying_graphs=underlying_graphs)
        
        self.seq = nn.Sequential(conv1, conv2, pool1, conv3, conv4)
        
    def forward(self, x):
        out = self.seq(x)
        out = out.mean(2)
        return out

net = Net()

License

The content of this repository is released under the terms of the MIT license