Add Hyperbolic GCN layer of paper [Hyperbolic Graph Convolutional Neural Networks](https://arxiv.org/abs/1910.12933)

CHANGELOG.md

@@ -7,6 +7,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- Added `HGCNConv` layer ([#9423](https://github.com/pyg-team/pytorch_geometric/pull/9423))
 - Added the heterogeneous `HeteroJumpingKnowledge` module ([#9380](https://github.com/pyg-team/pytorch_geometric/pull/9380))
 - Started work on GNN+LLM package ([#9350](https://github.com/pyg-team/pytorch_geometric/pull/9350))
 - Added support for negative sampling in `LinkLoader` acccording to source and destination node weights ([#9316](https://github.com/pyg-team/pytorch_geometric/pull/9316))

examples/hgcn.py

@@ -0,0 +1,108 @@
+import argparse
+import os.path as osp
+import time
+
+import torch
+import torch.nn.functional as F
+
+import torch_geometric.transforms as T
+from torch_geometric.datasets import Planetoid
+from torch_geometric.logging import init_wandb, log
+from torch_geometric.nn import HGCNConv
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--dataset', type=str, default='Cora')
+parser.add_argument('--hidden_channels', type=int, default=128)
+parser.add_argument('--heads', type=int, default=8)
+parser.add_argument('--lr', type=float, default=0.1e-3)
+parser.add_argument('--epochs', type=int, default=100)
+parser.add_argument('--wandb', action='store_true', help='Track experiment')
+args = parser.parse_args()
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+init_wandb(name=f'HGCN-{args.dataset}', heads=args.heads, epochs=args.epochs,
+           hidden_channels=args.hidden_channels, lr=args.lr, device=device)
+
+path = osp.join(osp.dirname(osp.realpath(__file__)), '..', 'data', 'Planetoid')
+dataset = Planetoid(path, args.dataset, transform=T.NormalizeFeatures())
+data = dataset[0].to(device)
+
+
+class LinearDecoder(torch.nn.Module):
+    """MLP Decoder for Hyperbolic/Euclidean node classification models."""
+    def __init__(self, manifold, in_dim, out_dim, c):
+        super(LinearDecoder, self).__init__()
+        self.manifold = manifold
+        self.input_dim = in_dim
+        self.output_dim = out_dim
+        self.cls = torch.nn.Linear(self.input_dim, self.output_dim)
+        self.c = torch.nn.Parameter(torch.FloatTensor([c]))
+
+    def forward(self, x):
+        x = self.manifold.proj_tan0(self.manifold.logmap0(x, c=self.c),
+                                    c=self.c)
+        x = self.cls(x)
+        return x
+
+    def extra_repr(self):
+        return 'in_features={}, out_features={}, c={}'.format(
+            self.input_dim, self.output_dim, self.c)
+
+
+class HGCN(torch.nn.Module):
+    def __init__(self, in_channels, hidden_channels, out_channels):
+        super().__init__()
+        self.conv1 = HGCNConv(in_channels, hidden_channels, dropout=0.5)
+        self.decoder = LinearDecoder(self.conv1.manifold, hidden_channels,
+                                     out_channels, 1.0)
+
+    def forward(self, x, edge_index):
+        # print(x.shape, edge_index.shape)
+        x = self.conv1(x, edge_index)
+        x = F.dropout(x, p=0.6, training=self.training)
+        x = self.decoder(x)
+        x = F.log_softmax(x, dim=-1)
+        # print(x.shape)
+        # exit(0)
+        return x
+
+
+model = HGCN(dataset.num_features, args.hidden_channels,
+             dataset.num_classes).to(device)
+optimizer = torch.optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-3)
+print(model)
+
+
+def train():
+    model.train()
+    optimizer.zero_grad()
+    out = model(data.x, data.edge_index)
+    loss = F.cross_entropy(out[data.train_mask], data.y[data.train_mask])
+    loss.backward()
+    optimizer.step()
+    return float(loss)
+
+
+@torch.no_grad()
+def test():
+    model.eval()
+    pred = model(data.x, data.edge_index).argmax(dim=-1)
+
+    accs = []
+    for mask in [data.train_mask, data.val_mask, data.test_mask]:
+        accs.append(int((pred[mask] == data.y[mask]).sum()) / int(mask.sum()))
+    return accs
+
+
+times = []
+best_val_acc = final_test_acc = 0
+for epoch in range(1, args.epochs + 1):
+    start = time.time()
+    loss = train()
+    train_acc, val_acc, tmp_test_acc = test()
+    if val_acc > best_val_acc:
+        best_val_acc = val_acc
+        test_acc = tmp_test_acc
+    log(Epoch=epoch, Loss=loss, Train=train_acc, Val=val_acc, Test=test_acc)
+    times.append(time.time() - start)
+print(f"Median time per epoch: {torch.tensor(times).median():.4f}s")

test/nn/conv/test_hgcn_conv.py

@@ -0,0 +1,94 @@
+import torch
+
+from torch_geometric.nn.conv.hgcn_conv import (
+    HGCNConv,
+    HypAct,
+    Hyperboloid,
+    HypLinear,
+    PoincareBall,
+)
+from torch_geometric.utils import to_torch_csc_tensor
+
+
+def test_hgcn_conv_hyperboloid_forward():
+    x = torch.randn(4, 16)
+    edge_index = torch.tensor([[0, 0, 0, 1, 2, 3], [1, 2, 3, 0, 0, 0]])
+    value = torch.rand(edge_index.size(1))
+    adj1 = to_torch_csc_tensor(edge_index, size=(4, 4))
+    adj2 = to_torch_csc_tensor(edge_index, value, size=(4, 4))
+
+    conv = HGCNConv(16, 32, manifold='hyperboloid')
+    assert str(conv) == 'HGCNConv(16, 32)'
+
+    out1 = conv(x, edge_index)
+    assert out1.size() == (4, 32)
+    assert torch.allclose(conv(x, adj1.t()), out1, atol=1e-6)
+
+    out2 = conv(x, edge_index, value)
+    assert out2.size() == (4, 32)
+    assert torch.allclose(conv(x, adj2.t()), out2, atol=1e-6)
+    return
+
+
+def test_hgcn_conv_poincareBall_forward():
+    x = torch.randn(4, 16)
+    edge_index = torch.tensor([[0, 0, 0, 1, 2, 3], [1, 2, 3, 0, 0, 0]])
+    value = torch.rand(edge_index.size(1))
+    adj1 = to_torch_csc_tensor(edge_index, size=(4, 4))
+    adj2 = to_torch_csc_tensor(edge_index, value, size=(4, 4))
+
+    conv = HGCNConv(16, 32, manifold='poincare')
+    assert str(conv) == 'HGCNConv(16, 32)'
+
+    out1 = conv(x, edge_index)
+    assert out1.size() == (4, 32)
+    assert torch.allclose(conv(x, adj1.t()), out1, atol=1e-6)
+
+    out2 = conv(x, edge_index, value)
+    assert out2.size() == (4, 32)
+    assert torch.allclose(conv(x, adj2.t()), out2, atol=1e-6)
+    return
+
+
+def test_hgcn_linear_hyperboloid_forward():
+    x = torch.randn(4, 16)
+    manifold = Hyperboloid()
+
+    linaer = HypLinear(manifold, 16, 32, 1.0, 0.0, True)
+
+    out1 = linaer(x)
+    assert out1.size() == (4, 32)
+    return
+
+
+def test_hgcn_linear_poincareBall_forward():
+    x = torch.randn(4, 16)
+    manifold = PoincareBall()
+
+    linaer = HypLinear(manifold, 16, 32, 1.0, 0.0, True)
+
+    out1 = linaer(x)
+    assert out1.size() == (4, 32)
+    return
+
+
+def test_hypact_hyperboloid_forward():
+    x = torch.randn(4, 16)
+    manifold = Hyperboloid()
+
+    hypact = HypAct(manifold, 1.0, 1.0, None)
+
+    out1 = hypact(x)
+    assert out1.size() == (4, 16)
+    return
+
+
+def test_hypact_poincareBall_forward():
+    x = torch.randn(4, 16)
+    manifold = PoincareBall()
+
+    hypact = HypAct(manifold, 1.0, 1.0, None)
+
+    out1 = hypact(x)
+    assert out1.size() == (4, 16)
+    return

torch_geometric/nn/conv/__init__.py

@@ -61,6 +61,7 @@
 from .antisymmetric_conv import AntiSymmetricConv
 from .dir_gnn_conv import DirGNNConv
 from .mixhop_conv import MixHopConv
+from .hgcn_conv import HGCNConv
 
 import torch_geometric.nn.conv.utils  # noqa
 
@@ -131,6 +132,7 @@
     'AntiSymmetricConv',
     'DirGNNConv',
     'MixHopConv',
+    'HGCNConv',
 ]
 
 classes = __all__