File size: 11,309 Bytes
6de06a8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | """Single-task GAT for one indirect-control-flow relation."""
import torch
import torch.nn as nn
from common import (
EDGE_TYPE_MAPPING,
PaperHeteroGATLayer,
add_paper_data_structural_features,
build_paper_task_augmented_graph,
extract_node_features,
)
class SingleTaskGAT(nn.Module):
def __init__(
self,
code_feat_dim,
data_feat_dim,
task_type,
hidden_dim=256,
num_heads=8,
num_layers=4,
dropout=0.4,
use_reverse_edges=False,
use_gch=True,
use_gdh=True,
task_aware_routing=True,
gdh_radius=2,
**checkpoint_compatibility,
):
super(SingleTaskGAT, self).__init__()
assert hidden_dim % num_heads == 0, "hidden_dim must be divisible by num_heads"
self.task_type = task_type
self.hidden_dim = hidden_dim
self.num_heads = num_heads
self.num_layers = num_layers
self.use_reverse_edges = use_reverse_edges
self.use_gch = use_gch
self.use_gdh = use_gdh
self.task_aware_routing = task_aware_routing
# Accept metadata written by older checkpoints. The scorer architecture
# itself is fixed to an MLP and is no longer configurable.
checkpoint_compatibility.pop("architecture_version", None)
checkpoint_compatibility.pop("edge_predictor_hidden_dim", None)
if checkpoint_compatibility:
raise TypeError(
"Unexpected model configuration keys: {}".format(
", ".join(sorted(checkpoint_compatibility))
)
)
self.gdh_radius = int(gdh_radius)
# Versioned artifacts may replace this after construction. Keeping it
# on the model makes the default binary decision self-contained.
self.decision_threshold = 0.5
self.edge_type_mapping = EDGE_TYPE_MAPPING
# Input projections
self.code_projection = nn.Linear(code_feat_dim, hidden_dim)
projected_data_dim = data_feat_dim + 2 if data_feat_dim > 0 else data_feat_dim
self.data_projection = nn.Linear(projected_data_dim, hidden_dim) if projected_data_dim > 0 else None
self.hub_embedding = nn.Embedding(2, hidden_dim)
# GAT layers
self.encoder = nn.ModuleList([
PaperHeteroGATLayer(hidden_dim, hidden_dim, num_heads, dropout, use_reverse_edges=use_reverse_edges)
for _ in range(num_layers)
])
# Endpoint scoring is always a two-layer MLP. The MLP width follows the
# encoder width so there is no separate scorer-architecture option.
self.edge_predictor = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, 1),
)
def encode(self, g, feat_dict):
h_dict = {
'code': self.code_projection(feat_dict['code'])
}
if 'data' in feat_dict and self.data_projection is not None:
h_dict['data'] = self.data_projection(feat_dict['data'])
if 'hub' in g.ntypes:
h_dict['hub'] = self.hub_embedding(g.nodes['hub'].data['hub_kind'].long())
for layer in self.encoder:
h_dict = layer(g, h_dict)
return h_dict
def forward(self, g, edge_pairs=None, candidate_meta=None, return_embeddings=False):
"""
Forward pass of the model.
Args:
g: DGL graph
edge_pairs: Edge pairs to predict - can be:
- Tuple of (src_ids, dst_ids)
- Tensor of shape [num_edges, 2] where each row is [src, dst]
- Tensor of shape [2, num_edges] where first row is src, second is dst
return_embeddings: If True, return (embeddings, logits), else return only logits
Returns:
If edge_pairs is None: node embeddings
If return_embeddings is True: (node_embeddings, edge_logits)
Else: edge_logits only
"""
g = build_paper_task_augmented_graph(
g,
candidate_meta,
self.task_type,
use_gch=self.use_gch,
use_gdh=self.use_gdh,
task_aware_routing=self.task_aware_routing,
gdh_radius=self.gdh_radius,
)
feat_dict = extract_node_features(g)
feat_dict = add_paper_data_structural_features(g, feat_dict)
node_embeddings = self.encode(g, feat_dict)
if edge_pairs is None:
return node_embeddings
edge_logits = self.score_edges(node_embeddings, edge_pairs)
if return_embeddings:
return node_embeddings, edge_logits
else:
# Return only edge logits for training/evaluation
return edge_logits
def score_edges(self, node_embeddings, edge_pairs):
"""Score endpoint pairs from one reusable graph encoding."""
# FIXED: Handle different edge_pairs formats
if isinstance(edge_pairs, tuple):
# Format: (src_ids, dst_ids)
src_ids, dst_ids = edge_pairs
elif isinstance(edge_pairs, torch.Tensor):
if edge_pairs.dim() == 2:
if edge_pairs.shape[1] == 2:
# Format: [num_edges, 2] where each row is [src, dst]
src_ids = edge_pairs[:, 0]
dst_ids = edge_pairs[:, 1]
elif edge_pairs.shape[0] == 2:
# Format: [2, num_edges] where first row is src, second is dst
src_ids = edge_pairs[0, :]
dst_ids = edge_pairs[1, :]
else:
raise ValueError(f"Unexpected edge_pairs tensor shape: {edge_pairs.shape}")
else:
raise ValueError(f"Unexpected edge_pairs tensor dimensions: {edge_pairs.dim()}")
else:
raise ValueError(f"Unexpected edge_pairs type: {type(edge_pairs)}")
# Get embeddings for source and destination nodes
src_embs = node_embeddings['code'][src_ids]
dst_embs = node_embeddings['code'][dst_ids]
# Concatenate source and destination embeddings
edge_emb = torch.cat([src_embs, dst_embs], dim=1)
return self.edge_predictor(edge_emb)
def get_loss(self, edge_logits, edge_labels, edge_types, pos_weight=1.0):
task_type_id = self.edge_type_mapping[self.task_type]
type_labels = (edge_types == task_type_id) & (edge_labels == 1)
type_logits = edge_logits.squeeze()
n_pos = type_labels.sum().float()
n_neg = len(type_labels) - n_pos
type_pos_weight = (n_neg / n_pos) * pos_weight if n_pos > 0 else torch.tensor(pos_weight, device=type_logits.device)
loss_fn = nn.BCEWithLogitsLoss(reduction='none')
losses = loss_fn(type_logits, type_labels.float())
if n_pos > 0:
losses[type_labels] *= type_pos_weight
return losses.mean()
def predict_edges(self, g, edge_pairs, threshold=None, candidate_meta=None):
with torch.no_grad():
edge_logits = self.forward(g, edge_pairs, candidate_meta=candidate_meta)
probs = torch.sigmoid(edge_logits.squeeze())
if threshold is None:
threshold = self.decision_threshold
return (probs >= threshold), probs
def evaluate(self, g, pos_edges, pos_types, neg_edges, threshold=None, candidate_meta=None):
"""
Args:
pos_edges: List of tuples [(src, dst), ...] OR tensor of shape [num_edges, 2]
pos_types: List of edge types
neg_edges: List of tuples [(src, dst), ...] OR tensor of shape [num_edges, 2]
"""
task_type_id = self.edge_type_mapping[self.task_type]
# Handle different input formats
if isinstance(pos_edges, torch.Tensor):
# Already tensor format
pos_edges_tensor = pos_edges
else:
# List of tuples format
pos_edges_tensor = torch.tensor(pos_edges, dtype=torch.long, device=g.device)
if isinstance(neg_edges, torch.Tensor):
# Already tensor format
neg_edges_tensor = neg_edges
else:
# List of tuples format
neg_edges_tensor = torch.tensor(neg_edges, dtype=torch.long, device=g.device)
# Concatenate all edges
all_edges = torch.cat([pos_edges_tensor, neg_edges_tensor], dim=0)
all_labels = torch.cat([
torch.ones(len(pos_edges), device=g.device),
torch.zeros(len(neg_edges), device=g.device)
])
all_types = torch.cat([
torch.tensor(pos_types, dtype=torch.long, device=g.device),
torch.full((len(neg_edges),), -1, dtype=torch.long, device=g.device)
])
preds, probs = self.predict_edges(g, all_edges, threshold, candidate_meta=candidate_meta)
type_gt = (all_types == task_type_id)
tp = ((preds == 1) & type_gt).sum().item()
fp = ((preds == 1) & ~type_gt).sum().item()
fn = ((preds == 0) & type_gt).sum().item()
precision = tp / max(tp + fp, 1)
recall = tp / max(tp + fn, 1)
f1 = 2 * precision * recall / max(precision + recall, 1e-6)
return {
"precision": precision,
"recall": recall,
"f1": f1,
"support": type_gt.sum().item()
}
def train_step(self, g, pos_edges, pos_types, neg_edges, optimizer, pos_weight=1.0, candidate_meta=None):
"""
Args:
pos_edges: List of tuples [(src, dst), ...] OR tensor of shape [num_edges, 2]
pos_types: List of edge types
neg_edges: List of tuples [(src, dst), ...] OR tensor of shape [num_edges, 2]
"""
# Handle different input formats
if isinstance(pos_edges, torch.Tensor):
# Already tensor format
pos_edges_tensor = pos_edges
else:
# List of tuples format
pos_edges_tensor = torch.tensor(pos_edges, dtype=torch.long, device=g.device)
if isinstance(neg_edges, torch.Tensor):
# Already tensor format
neg_edges_tensor = neg_edges
else:
# List of tuples format
neg_edges_tensor = torch.tensor(neg_edges, dtype=torch.long, device=g.device)
# Concatenate all edges
edge_pairs = torch.cat([pos_edges_tensor, neg_edges_tensor], dim=0)
edge_labels = torch.cat([
torch.ones(len(pos_edges), device=g.device),
torch.zeros(len(neg_edges), device=g.device)
])
edge_types = torch.cat([
torch.tensor(pos_types, dtype=torch.long, device=g.device),
torch.full((len(neg_edges),), -1, dtype=torch.long, device=g.device)
])
edge_logits = self.forward(g, edge_pairs, candidate_meta=candidate_meta)
loss = self.get_loss(edge_logits, edge_labels, edge_types, pos_weight)
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss.item()
|