mnist_grid2.py

import torch
import torch.nn.functional as F
from torchvision.datasets import MNIST
from torchvision.transforms import ToTensor
from torch.utils.data import DataLoader

from torch_geometric.utils import grid
from torch_geometric.nn import SplineConv

train_dataset = MNIST('/tmp/MNIST', train=True, transform=ToTensor())
test_dataset = MNIST('/tmp/MNIST', train=False, transform=ToTensor())
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True,
                          drop_last=True)
test_loader = DataLoader(test_dataset, batch_size=64, drop_last=True)


def to_batch(edge_index, pos, batch_size):
    edge_indices = [edge_index + pos.size(0) * i for i in range(batch_size)]
    edge_index = torch.cat(edge_indices, dim=1)
    pos = torch.cat([pos] * batch_size, dim=0)
    edge_attr = pos[edge_index[0]] - pos[edge_index[1]]
    return edge_index, (edge_attr + 1.) / 2.


edge_index1, edge_attr1 = to_batch(*grid(28, 28), batch_size=64)
edge_index2, edge_attr2 = to_batch(*grid(14, 14), batch_size=64)


class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = SplineConv(1, 32, dim=2, kernel_size=3)
        self.conv2 = SplineConv(32, 64, dim=2, kernel_size=3)
        self.fc1 = torch.nn.Linear(3136, 512)
        self.fc2 = torch.nn.Linear(512, 10)

    def forward(self, x):
        x = x.view(-1, 1)
        x = F.elu(self.conv1(x.view(-1, 1), edge_index1, edge_attr1))
        x = x.view(64, 28, 28, 32).permute(0, 3, 1, 2)
        x = F.max_pool2d(x, kernel_size=2)
        x = x.permute(0, 2, 3, 1).contiguous().view(-1, 32)
        x = F.elu(self.conv2(x, edge_index2, edge_attr2))
        x = x.view(64, 14, 14, 64).permute(0, 3, 1, 2)
        x = F.max_pool2d(x, kernel_size=2)
        x = x.contiguous().view(64, -1)

        x = F.elu(self.fc1(x))
        return F.log_softmax(self.fc2(x), dim=1)


device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net().to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
edge_index1, edge_attr1 = edge_index1.to(device), edge_attr1.to(device)
edge_index2, edge_attr2 = edge_index2.to(device), edge_attr2.to(device)


def train(epoch):
    model.train()

    for i, (x, y) in enumerate(train_loader):
        x, y = x.to(device), y.to(device)
        optimizer.zero_grad()
        loss = F.nll_loss(model(x), y)
        loss.backward()
        optimizer.step()
        print(i, len(train_loader), loss.item())


def test():
    model.eval()

    correct = 0
    for i, (x, y) in enumerate(test_loader):
        x, y = x.to(device), y.to(device)
        pred = model(x).max(1)[1]
        correct += pred.eq(y).sum().item()
    return correct / len(test_dataset)


for epoch in range(1, 50):
    train(epoch)
    test_acc = test()
    print('Epoch: {:02d}, Test: {:.4f}'.format(epoch, test_acc))

In SplineCNN paper, you used Graclus based pooling rather than max_pool_2D, right?

After 50 epochs, I got around 98.8% accuracy, which is still not 99.33%. I used a filter size of 5. What's the idea behind using contiguous (is it for the reason described in the solution here: https://discuss.pytorch.org/t/when-and-why-do-we-use-contiguous/47588)? The reason you are using max_pool2D is that the graph is a grid graph here, right?

For the grid experiment, we use normal 2D max pooling since we cannot flatten the output features otherwise. To increase accuracy, adding two-hop neighbors to the grid helps. contiguous needs to be called since permute permutes the dimensions of the node features, and GNN operators generally expect inputs with contiguous memory layout.

What's two-hop neighbors? Is it similar to strides in normal CNN? How do I add that to the grid?