forked from Danial-sb/Diffpool
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdiffpool.py
219 lines (168 loc) · 38.7 KB
/
diffpool.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
# -*- coding: utf-8 -*-
"""DiffPool.ipynb
Automatically generated by Colaboratory.
Original file is located at
https://colab.research.google.com/drive/18x0slqSIdZJTMzcyZS69PKc9uqfBnB5a
**First, we look at the concept of DiffPool**
"""
import torch
from math import ceil
"""For doing the computation for the first hierarchical step, we need the node feature matrix and adjacency matrix of the graph.
For instance, imagine out initial graph would be like:
**X_0 = 200 * 32**
**A_0 = 200 * 200**
"""
X_0 = torch.rand(200, 32)
A_0 = torch.rand(200, 200).round().long()
identity = torch.eye(200)
A_0 = A_0 + identity
"""At first step, we have **100** clusters (nodes). For the next step, we want to have these nodes into **10** clusters"""
n_clusters_0 = 200
n_clusters_1 = 10
"""Initialize GNN_weights and GNN_pool"""
w_gnn_weights = torch.rand(32, 16)
w_gnn_pool = torch.rand(32, n_clusters_1)
"""
"""
Z_0 = torch.relu(A_0 @ X_0 @ w_gnn_weights)
S_0 = torch.softmax((A_0 @ X_0 @ w_gnn_pool), dim = 1)
print(f"shape of Z_0: {Z_0.shape}\nshape of S_0: {S_0.shape}")
""""""
X_1 = S_0.t() @ Z_0
A_1 = S_0.t() @ A_0 @ S_0
print(f"shape of Z_0: {X_1.shape}\nshape of S_0: {A_1.shape}")
!pip install -q torch-scatter -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
!pip install -q torch-sparse -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
!pip install -q git+https://github.com/pyg-team/pytorch_geometric.git
import torch.nn.functional as F
from torch_geometric.datasets import TUDataset
import torch_geometric.transforms as T
from torch_geometric.data import DenseDataLoader
from torch_geometric.data import DataLoader
from torch.nn import Linear
from torch_geometric.nn import global_mean_pool
from torch_geometric.nn import DenseGCNConv as GCNConv, dense_diff_pool
#from torch_geometric.nn import GCNConv
max_nodes = 300
class Filter(object):
def __call__(self, data):
return data.num_nodes <= max_nodes
dataset = TUDataset('/content', 'PROTEINS', transform = T.ToDense(max_nodes), pre_filter=Filter())
#dataset = TUDataset('/content', 'PROTEINS', transform = T.ToDense(max_nodes), pre_filter=Filter())
print(f"Number of graphs in the dataset: {len(dataset)}")
print(f"Number of features: {dataset.num_features}")
print(f"Number of classes: {dataset.num_classes}")
dataset.shuffle()
n = (len(dataset)) // 10
test_data = dataset[:2*n]
val_data = dataset[n:2*n]
train_data = dataset[2*n:]
print(f'length of train data: {len(train_data)}')
print(f'length of test data: {len(test_data)}')
print("========================")
test_loader = DenseDataLoader(test_data, batch_size = 32)
val_loader = DenseDataLoader(val_data, batch_size = 32)
train_loader = DenseDataLoader(train_data, batch_size = 32)
for step, data in enumerate(train_loader):
print(f'Step {step + 1}:')
print('=======')
#print(f'Number of graphs in the current batch: {data.num_graphs}')
print(data)
print()
class GNN(torch.nn.Module):
def __init__(self, in_channels, hidden_channels, out_channels,
normalize=True, lin=True):
super(GNN, self).__init__()
self.convs = torch.nn.ModuleList()
self.convs.append(GCNConv(in_channels, hidden_channels, normalize))
self.convs.append(GCNConv(hidden_channels, hidden_channels, normalize))
self.convs.append(GCNConv(hidden_channels, out_channels, normalize))
def forward(self, x, adj, mask=None):
batch_size, num_nodes, in_channels = x.size()
for i in range(len(self.convs)):
x = F.relu(self.convs[i](x, adj, mask))
return x
class DiffPool(torch.nn.Module):
def __init__(self):
super(DiffPool, self).__init__()
num_nodes = ceil(0.25 * max_nodes)
self.gnn1_pool = GNN(dataset.num_features, 64, num_nodes)
self.gnn1_embed = GNN(dataset.num_features, 64, 64)
num_nodes = ceil(0.25 * num_nodes)
self.gnn2_pool = GNN(64, 64, num_nodes)
self.gnn2_embed = GNN(64, 64, 64, lin=False)
self.gnn3_embed = GNN(64, 64, 64, lin=False)
self.lin1 = torch.nn.Linear(64, 64)
self.lin2 = torch.nn.Linear(64, dataset.num_classes)
def forward(self, x, adj, mask=None):
s = self.gnn1_pool(x, adj, mask)
x = self.gnn1_embed(x, adj, mask)
x, adj, l1, e1 = dense_diff_pool(x, adj, s)
s = self.gnn2_pool(x, adj)
x = self.gnn2_embed(x, adj)
x, adj, l2, e2 = dense_diff_pool(x, adj, s)
x = self.gnn3_embed(x, adj)
x = x.mean(dim=1)
x = F.relu(self.lin1(x))
x = self.lin2(x)
return F.log_softmax(x, dim=-1), l1 + l2, e1 + e2
model = DiffPool()
print(model)
for data in train_loader:
print(data.y.size(0))
break
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = DiffPool().to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001, weight_decay=5e-4)
n_total_steps = len(train_loader)
num_epochs = 10
def train():
loss_all = 0
correct = 0
model.train()
for data in train_loader:
data = data.to(device)
optimizer.zero_grad()
output, _, _ = model(data.x, data.adj, data.mask)
loss = F.nll_loss(output, data.y.view(-1))
loss.backward()
loss_all += data.y.size(0) * loss.item()
optimizer.step()
pred = output.argmax(dim = 1)
correct += int((pred == data.y).sum())
return (loss_all / len(train_data)), (correct / len(train_data))
@torch.no_grad()
def test(loader):
model.eval()
correct = 0
for data in loader:
data = data.to(device)
pred = model(data.x, data.adj, data.mask)[0].max(dim=1)[1]
correct += pred.eq(data.y.view(-1)).sum().item()
return correct / len(loader.dataset)
train_loss = []
train_acc = []
for epoch in range(1, 10):
loss, acc = train()
train_loss.append(loss)
train_acc.append(acc)
#if epoch % 10 == 0:
print(f'epoch: {epoch}, Train loss: {loss:.4f}, Train acc: {acc:.4f}')
train_loss = []
train_acc = []
val_loss = []
val_acc = []
best_val_acc = test_acc = 0
for epoch in range(1, 151):
loss, acc = train()
train_loss.append(loss)
train_acc.append(acc)
v_acc = test(val_loader)
if v_acc > best_val_acc:
test_acc = test(test_loader)
best_val_acc = v_acc
#val_loss.append(v_loss)
val_acc.appened(v_acc)
if epoch % 10 == 0:
print(f'epoch: {epoch}, Train loss: {loss:.4f}, Train acc: {acc:.4f}, Val acc: {v_acc:.4f}, Test acc: {test_acc:.4f}')