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
|
#!/usr/bin/env python3
"""H19 CitationFull-CiteSeer (4.2K, deg 2.5, 6-class) — same regime as Planetoid CiteSeer.
Quick BP + KAFT depth sweep to confirm/extend the 'KAFT wins on real sparse citation' story."""
import torch, sys, numpy as np, time
import torch.nn as nn, torch.nn.functional as F
from torch_geometric.datasets import CitationFull
from torch_geometric.nn import GCNConv
from torch_geometric.utils import add_self_loops, degree
sys.path.insert(0, '/home/yurenh2/graph-grape')
from src.trainers import KAFTTrainer
DATA_ROOT = '/home/yurenh2/graph-grape/data/CFull'
device = torch.device('cuda:0')
def build_A_hat(edge_index, N):
edge_index, _ = add_self_loops(edge_index, num_nodes=N)
row, col = edge_index
deg = degree(row, num_nodes=N, dtype=torch.float)
dis = deg.pow(-0.5); dis[dis == float('inf')] = 0
return torch.sparse_coo_tensor(edge_index, dis[row]*dis[col], (N, N)).coalesce()
def build_row_norm(edge_index, N):
ei, _ = add_self_loops(edge_index, num_nodes=N)
row, col = ei
deg = degree(row, num_nodes=N, dtype=torch.float).clamp(min=1)
A_row = torch.sparse_coo_tensor(ei, 1.0/deg[row], (N,N)).coalesce()
A_row_T = torch.sparse_coo_tensor(ei.flip(0), 1.0/deg[col], (N,N)).coalesce()
return A_row, A_row_T
def make_split(N, seed, y, n_per_class=20, n_val=500):
g = torch.Generator().manual_seed(seed)
train_mask = torch.zeros(N, dtype=torch.bool)
val_mask = torch.zeros(N, dtype=torch.bool)
test_mask = torch.zeros(N, dtype=torch.bool)
C = int(y.max()) + 1
for c in range(C):
idx = (y == c).nonzero().flatten()
idx = idx[torch.randperm(idx.size(0), generator=g)]
train_mask[idx[:n_per_class]] = True
remaining = (~train_mask).nonzero().flatten()
remaining = remaining[torch.randperm(remaining.size(0), generator=g)]
val_mask[remaining[:n_val]] = True
test_mask[remaining[n_val:]] = True
return train_mask, val_mask, test_mask
class GCN(nn.Module):
def __init__(self, in_dim, hidden, out_dim, L, dropout=0.1):
super().__init__()
self.convs = nn.ModuleList([GCNConv(in_dim if i==0 else hidden,
hidden if i<L-1 else out_dim) for i in range(L)])
self.dropout = dropout
def forward(self, x, ei):
for l, c in enumerate(self.convs):
x = c(x, ei)
if l < len(self.convs)-1:
x = F.relu(x)
if self.dropout>0: x = F.dropout(x, self.dropout, self.training)
return x
def bp_one(L, seed, d, train_mask, val_mask, test_mask, epochs=200, lr=5e-3, hidden=128, dropout=0.1):
torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed)
model = GCN(d.x.shape[1], hidden, int(d.y.max())+1, L, dropout=dropout).to(device)
opt = torch.optim.AdamW(model.parameters(), lr=lr)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)
@torch.no_grad()
def ev(m):
model.eval()
out = model(d.x.float(), d.edge_index)
return (out[m].argmax(1) == d.y[m]).float().mean().item()
bv, bt = 0, 0
for ep in range(epochs):
model.train()
out = model(d.x.float(), d.edge_index)
loss = F.cross_entropy(out[train_mask], d.y[train_mask])
opt.zero_grad(); loss.backward(); opt.step()
sched.step()
if ep % 5 == 0:
v = ev(val_mask)
if v > bv: bv = v; bt = ev(test_mask)
return bt
def kaft_one(L, seed, d, A_hat, A_row, A_row_T, train_mask, val_mask, test_mask,
epochs=200, lr=5e-3, hidden=128):
torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed)
data = {
'X': d.x.float(), 'A_hat': A_hat, 'A_row': A_row, 'A_row_T': A_row_T,
'y': d.y, 'train_mask': train_mask, 'val_mask': val_mask, 'test_mask': test_mask,
'num_features': d.x.shape[1], 'num_classes': int(d.y.max())+1,
'num_nodes': d.num_nodes, 'traces': {},
}
trainer = KAFTTrainer(
data=data, hidden_dim=hidden, lr=lr, weight_decay=0.0,
lr_feedback=0.5, num_probes=64, topo_mode='fixed_A', max_topo_power=3,
diffusion_alpha=0.5, diffusion_iters=10,
num_layers=L, residual_alpha=0.0, backbone='gcn',
use_batchnorm=False, dropout=0.0,
)
trainer.align_mode = 'chain_norm'
bv, bt = 0, 0
for ep in range(epochs):
trainer.train_step()
if ep % 5 == 0:
v = trainer.evaluate('val_mask')
if v > bv: bv = v; bt = trainer.evaluate('test_mask')
return bt
def main():
d = CitationFull(root=DATA_ROOT, name='CiteSeer')[0].to(device)
N = d.num_nodes
A_hat = build_A_hat(d.edge_index, N)
A_row, A_row_T = build_row_norm(d.edge_index, N)
print(f'CitationFull-CiteSeer: N={N}, deg={d.edge_index.shape[1]/N:.1f}, C={int(d.y.max())+1}', flush=True)
bp_res, gf_res = {}, {}
for L in [3, 5, 10, 16]:
bp_accs, gf_accs = [], []
for s in [0, 1, 2]:
tm, vm, tem = make_split(N, s, d.y.cpu())
tm, vm, tem = tm.to(device), vm.to(device), tem.to(device)
t0 = time.time()
bp = bp_one(L, s, d, tm, vm, tem)
t1 = time.time()
gf = kaft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem)
t2 = time.time()
bp_accs.append(bp); gf_accs.append(gf)
print(f' L={L} s={s}: BP={bp:.4f}({t1-t0:.0f}s) KAFT={gf:.4f}({t2-t1:.0f}s)', flush=True)
bp_m, bp_sd = np.mean(bp_accs), np.std(bp_accs)
gf_m, gf_sd = np.mean(gf_accs), np.std(gf_accs)
bp_res[L] = (bp_m, bp_sd); gf_res[L] = (gf_m, gf_sd)
print(f'>>> L={L}: BP {bp_m:.4f}±{bp_sd:.4f} KAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True)
if __name__ == '__main__':
main()
|
