import math
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import NNConv, Set2Set
from .pyg_gin import GINLayer, GINMLP
from cogdl.data import DataLoader
from cogdl.utils import (
batch_mean_pooling,
batch_sum_pooling
)
from .. import BaseModel, register_model
[docs]class SUPEncoder(torch.nn.Module):
r"""Encoder used in supervised model with Set2set in paper `"Order Matters: Sequence to sequence for sets"
<https://arxiv.org/abs/1511.06391>` and NNConv in paper `"Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on
Graphs" <https://arxiv.org/abs/1704.02901>`
"""
def __init__(self, num_features, dim, num_layers=1):
super(SUPEncoder, self).__init__()
self.lin0 = torch.nn.Linear(num_features, dim)
nnu = nn.Sequential(nn.Linear(5, 128), nn.ReLU(), nn.Linear(128, dim * dim))
self.conv = NNConv(dim, dim, nnu, aggr='mean', root_weight=False)
self.gru = nn.GRU(dim, dim)
self.set2set = Set2Set(dim, processing_steps=3)
# self.lin1 = torch.nn.Linear(2 * dim, dim)
# self.lin2 = torch.nn.Linear(dim, 1)
[docs] def forward(self, x, edge_index, batch, edge_attr):
out = F.relu(self.lin0(x))
h = out.unsqueeze(0)
feat_map = []
for i in range(3):
m = F.relu(self.conv(out, edge_index, edge_attr))
out, h = self.gru(m.unsqueeze(0), h)
out = out.squeeze(0)
feat_map.append(out)
out = self.set2set(out, batch)
return out, feat_map[-1]
[docs]class Encoder(nn.Module):
r"""Encoder stacked with GIN layers
Parameters
----------
in_feats : int
Size of each input sample.
hidden_feats : int
Size of output embedding.
num_layers : int, optional
Number of GIN layers, default: ``3``.
num_mlp_layers : int, optional
Number of MLP layers for each GIN layer, default: ``2``.
pooling : str, optional
Aggragation type, default : ``sum``.
"""
def __init__(self, in_feats, hidden_dim, num_layers=3, num_mlp_layers=2, pooling='sum'):
super(Encoder, self).__init__()
self.num_layers = num_layers
self.gnn_layers = nn.ModuleList()
self.bn_layers = nn.ModuleList()
for i in range(num_layers):
if i == 0:
mlp = GINMLP(in_feats, hidden_dim, hidden_dim, num_mlp_layers, use_bn=True)
else:
mlp = GINMLP(hidden_dim, hidden_dim, hidden_dim, num_mlp_layers, use_bn=True)
self.gnn_layers.append(GINLayer(
mlp, eps=0, train_eps=True
))
self.bn_layers.append(nn.BatchNorm1d(hidden_dim))
if pooling == 'sum':
self.pooling = batch_sum_pooling
elif pooling == "mean":
self.pooling = batch_mean_pooling
else:
raise NotImplementedError
[docs] def forward(self, x, edge_index, batch, *args):
if x is None:
x = torch.ones((batch.shape[0], 1)).to(batch.device)
layer_rep = []
for i in range(self.num_layers):
x = F.relu(self.bn_layers[i](self.gnn_layers[i](x, edge_index)))
layer_rep.append(x)
pooled_rep = [self.pooling(h, batch) for h in layer_rep]
node_rep = torch.cat(layer_rep, dim=1)
graph_rep = torch.cat(pooled_rep, dim=1)
return graph_rep, node_rep
[docs]class FF(nn.Module):
r"""Residual MLP layers.
..math::
out = \mathbf{MLP}(x) + \mathbf{Linear}(x)
Paramaters
----------
in_feats : int
Size of each input sample
out_feats : int
Size of each output sample
"""
def __init__(self, in_feats, out_feats):
super(FF, self).__init__()
self.block = GINMLP(in_feats, out_feats, out_feats, num_layers=3, use_bn=False)
self.shortcut = nn.Linear(in_feats, out_feats)
[docs] def forward(self, x):
return F.relu(self.block(x)) + self.shortcut(x)
[docs]@register_model("infograph")
class InfoGraph(BaseModel):
r"""Implimentation of Infograph in paper `"InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation
Learning via Mutual Information Maximization" <https://openreview.net/forum?id=r1lfF2NYvH>__. `
Parameters
----------
in_feats : int
Size of each input sample.
out_feats : int
Size of each output sample.
num_layers : int, optional
Number of MLP layers in encoder, default: ``3``.
unsup : bool, optional
Use unsupervised model if True, default: ``True``.
"""
@staticmethod
[docs] def add_args(parser):
parser.add_argument("--hidden-size", type=int, default=512)
parser.add_argument("--batch-size", type=int, default=20)
parser.add_argument("--target", dest='target', type=int, default=0,
help='')
parser.add_argument("--train-num", dest='train_num', type=int, default=5000)
parser.add_argument("--num-layers", type=int, default=1)
parser.add_argument("--sup", dest="sup", action="store_true")
parser.add_argument("--epoch", type=int, default=15)
parser.add_argument("--lr", type=float, default=0.0001)
parser.add_argument("--train-ratio", type=float, default=0.7)
parser.add_argument("--test-ratio", type=float, default=0.1)
@classmethod
[docs] def build_model_from_args(cls, args):
return cls(
args.num_features,
args.hidden_size,
args.num_classes,
args.num_layers,
args.sup
)
@classmethod
[docs] def split_dataset(cls, dataset, args):
if args.dataset == "qm9":
test_dataset = dataset[:10000]
val_dataset = dataset[10000:20000]
train_dataset = dataset[20000:20000 + args.train_num]
return DataLoader(train_dataset, batch_size=args.batch_size), DataLoader(val_dataset, batch_size=args.batch_size),\
DataLoader(test_dataset, batch_size=args.batch_size)
else:
random.shuffle(dataset)
train_size = int(len(dataset) * args.train_ratio)
test_size = int(len(dataset) * args.test_ratio)
bs = args.batch_size
train_loader = DataLoader(dataset[:train_size], batch_size=bs)
test_loader = DataLoader(dataset[-test_size:], batch_size=bs)
if args.train_ratio + args.test_ratio < 1:
valid_loader = DataLoader(dataset[train_size:-test_size], batch_size=bs)
else:
valid_loader = test_loader
return train_loader, valid_loader, test_loader
def __init__(self, in_feats, hidden_dim, out_feats, num_layers=3, sup=False):
super(InfoGraph, self).__init__()
self.sup = sup
self.emb_dim = hidden_dim
self.out_feats = out_feats
self.sem_fc1 = nn.Linear(num_layers*hidden_dim, hidden_dim)
self.sem_fc2 = nn.Linear(hidden_dim, out_feats)
if not sup:
self.unsup_encoder = Encoder(in_feats, hidden_dim, num_layers)
self.register_parameter("sem_encoder", None)
else:
self.unsup_encoder = Encoder(in_feats, hidden_dim, num_layers)
self.sem_encoder = Encoder(in_feats, hidden_dim, num_layers)
self._fc1 = FF(num_layers*hidden_dim, hidden_dim)
self._fc2 = FF(num_layers*hidden_dim, hidden_dim)
self.local_dis = FF(num_layers*hidden_dim, hidden_dim)
self.global_dis = FF(num_layers*hidden_dim, hidden_dim)
self.criterion = nn.MSELoss()
[docs] def reset_parameters(self):
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight.data)
[docs] def forward(self, batch):
if self.sup:
return self.sup_forward(batch.x, batch.edge_index, batch.batch, batch.y, batch.edge_attr)
else:
return self.unsup_forward(batch.x, batch.edge_index, batch.batch)
[docs] def sup_forward(self, x, edge_index=None, batch=None, label=None, edge_attr=None):
node_feat, graph_feat = self.sem_encoder(x, edge_index, batch, edge_attr)
node_feat = F.relu(self.sem_fc1(node_feat))
node_feat = self.sem_fc2(node_feat)
prediction = F.softmax(node_feat, dim=1)
if label is not None:
loss = self.sup_loss(prediction, label)
loss += self.unsup_loss(x, edge_index, batch)[1]
loss += self.unsup_sup_loss(x, edge_index, batch)
return prediction, loss
return prediction, None
[docs] def unsup_forward(self, x, edge_index=None, batch=None):
return self.unsup_loss(x, edge_index, batch)
[docs] def sup_loss(self, prediction, label=None):
sup_loss = self.criterion(prediction, label)
return sup_loss
[docs] def unsup_loss(self, x, edge_index=None, batch=None):
graph_feat, node_feat = self.unsup_encoder(x, edge_index, batch)
local_encode = self.local_dis(node_feat)
global_encode = self.global_dis(graph_feat)
num_graphs = graph_feat.shape[0]
num_nodes = node_feat.shape[0]
pos_mask = torch.zeros((num_nodes, num_graphs)).to(x.device)
neg_mask = torch.ones((num_nodes, num_graphs)).to(x.device)
for nid, gid in enumerate(batch):
pos_mask[nid][gid] = 1
neg_mask[nid][gid] = 0
glob_local_mi = torch.mm(local_encode, global_encode.t())
loss = InfoGraph.mi_loss(pos_mask, neg_mask, glob_local_mi, num_nodes, num_nodes * (num_graphs-1))
return graph_feat, loss
[docs] def unsup_sup_loss(self, x, edge_index, batch):
sem_g_feat, _ = self.sem_encoder(x, edge_index, batch)
un_g_feat, _ = self.unsup_encoder(x, edge_index, batch)
sem_encode = self._fc1(sem_g_feat)
un_encode = self._fc2(un_g_feat)
num_graphs = sem_encode.shape[0]
pos_mask = torch.eye(num_graphs).to(x.device)
neg_mask = 1 - pos_mask
mi = torch.mm(sem_encode, un_encode.t())
loss = InfoGraph.mi_loss(pos_mask, neg_mask, mi, pos_mask.sum(), neg_mask.sum())
return loss
@staticmethod
[docs] def mi_loss(pos_mask, neg_mask, mi, pos_div, neg_div):
pos_mi = pos_mask * mi
neg_mi = neg_mask * mi
pos_loss = (-math.log(2.) + F.softplus(-pos_mi)).sum()
neg_loss = (-math.log(2.) + F.softplus(-neg_mi) + neg_mi).sum()
# pos_loss = F.softplus(-pos_mi).sum()
# neg_loss = (F.softplus(neg_mi)).sum()
# pos_loss = pos_mi.sum()
# neg_loss = neg_mi.sum()
return pos_loss / pos_div + neg_loss / neg_div