File size: 15,416 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 | """Multi-task GAT for joint indirect-control-flow prediction."""
import torch
import torch.nn as nn
import dgl
from common import (
EDGE_TYPE_MAPPING,
PaperHeteroGATLayer,
add_paper_data_structural_features,
build_paper_mtl_augmented_graph,
extract_node_features,
)
class MultiTaskGAT(nn.Module):
def __init__(
self,
code_feat_dim,
data_feat_dim,
hidden_dim=256,
num_heads=8,
shared_layers=4,
task_layers=1,
dropout=0.4,
use_reverse_edges=False,
task_weights=None,
use_gch=True,
use_gdh=True,
task_aware_routing=True,
gdh_radius=2,
**checkpoint_compatibility,
):
super(MultiTaskGAT, self).__init__()
assert hidden_dim % num_heads == 0, "hidden_dim must be divisible by num_heads"
self.edge_type_mapping = EDGE_TYPE_MAPPING
self.code_feat_dim = code_feat_dim
self.data_feat_dim = data_feat_dim
self.shared_layers = shared_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 the redundant field written by older checkpoints.
checkpoint_compatibility.pop("architecture_version", None)
if checkpoint_compatibility:
raise TypeError(
"Unexpected model configuration keys: {}".format(
", ".join(sorted(checkpoint_compatibility))
)
)
self.gdh_radius = int(gdh_radius)
self.task_layers = task_layers
self.hidden_dim = hidden_dim
self.num_heads = num_heads
self.dropout = dropout
# Task weights for loss balancing
self.task_weights = task_weights or {et: 1.0 for et in self.edge_type_mapping}
self.task_weights["edge_type"] = self.task_weights.get("edge_type", 1.0)
# 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)
# Shared encoder layers
self.shared_encoder = nn.ModuleList([
PaperHeteroGATLayer(
hidden_dim,
hidden_dim,
num_heads,
dropout,
use_reverse_edges=use_reverse_edges,
)
for _ in range(shared_layers)
])
# Task-specific encoder layers
self.task_encoders = nn.ModuleDict({
task: nn.ModuleList([
PaperHeteroGATLayer(
hidden_dim,
hidden_dim,
num_heads,
dropout,
use_reverse_edges=use_reverse_edges,
)
for _ in range(task_layers)
]) for task in self.edge_type_mapping
})
# Both endpoint matching and edge-type prediction always use two-layer
# MLPs. Their hidden width follows the encoder width, leaving no
# alternative scorer architecture to configure.
self.edge_heads = nn.ModuleDict({
task: nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, 1),
)
for task in self.edge_type_mapping
})
self.edge_type_head = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, len(self.edge_type_mapping)),
)
def encode(self, g, task=None):
"""Encode graph nodes with optional task-specific layers"""
feat_dict = extract_node_features(g)
feat_dict = add_paper_data_structural_features(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())
# Apply shared layers
for layer in self.shared_encoder:
h_dict = layer(g, h_dict)
# Apply task-specific layers if specified
if task is not None and task in self.task_encoders:
for layer in self.task_encoders[task]:
h_dict = layer(g, h_dict)
return h_dict
def forward(self, g, edge_pairs=None, candidate_meta=None, return_embeddings=False):
"""
Forward pass supporting multiple input formats like SingleTaskGAT
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 along with predictions
Returns:
If edge_pairs is None: node embeddings
If return_embeddings is True: (node_embeddings, edge_logits, edge_type_logits)
Else: (edge_logits, edge_type_logits)
"""
model_graph = build_paper_mtl_augmented_graph(
g,
candidate_meta,
use_gch=self.use_gch,
use_gdh=self.use_gdh,
task_aware_routing=self.task_aware_routing,
gdh_radius=self.gdh_radius,
)
# Run the shared encoder once on the unified graph.
node_embs_shared = self.encode(model_graph)
if edge_pairs is None:
return node_embs_shared
# FIXED: Handle different edge_pairs formats like SingleTaskGAT
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)}")
# Shared embeddings for edge type classification
shared_edge_embs = torch.cat([
node_embs_shared['code'][src_ids],
node_embs_shared['code'][dst_ids]
], dim=1)
# Task-specific edge predictions
edge_logits = {}
for task in self.edge_type_mapping:
node_embs_task = dict(node_embs_shared)
for layer in self.task_encoders[task]:
node_embs_task = layer(model_graph, node_embs_task)
task_edge_embs = torch.cat([
node_embs_task['code'][src_ids],
node_embs_task['code'][dst_ids]
], dim=1)
edge_logits[task] = self.edge_heads[task](task_edge_embs)
# Edge type classification
edge_type_logits = self.edge_type_head(shared_edge_embs)
if return_embeddings:
return node_embs_shared, edge_logits, edge_type_logits
else:
return edge_logits, edge_type_logits
def get_loss(
self,
edge_logits,
edge_labels,
edge_types,
pos_weight=1.0,
edge_type_logits=None,
negative_class_weights=None,
hard_negative_focus=None,
):
"""Compute multi-task loss with proper weighting"""
bce = nn.BCEWithLogitsLoss(reduction='none')
total_loss = 0
losses = {}
for task_name, type_id in self.edge_type_mapping.items():
# Each classifier consumes only its own positives and independently
# sampled negatives, as required by Section 4.4.
type_mask = (edge_types == type_id)
if not type_mask.any():
continue
task_labels = edge_labels[type_mask].float()
logits = edge_logits[task_name].reshape(-1)[type_mask]
# Calculate class weights
n_pos = task_labels.sum().float()
n_neg = len(task_labels) - n_pos
if n_pos > 0:
weight = (n_neg / n_pos) * pos_weight
loss = bce(logits, task_labels)
loss[task_labels == 1] *= weight
else:
loss = bce(logits, task_labels)
negative_mask = task_labels == 0
if negative_mask.any():
class_weight = float(
(negative_class_weights or {}).get(task_name, 1.0)
)
if class_weight != 1.0:
loss[negative_mask] *= class_weight
focus = (hard_negative_focus or {}).get(task_name)
if focus is not None:
alpha, gamma = map(float, focus)
negative_scores = torch.sigmoid(
logits[negative_mask].detach()
)
loss[negative_mask] *= (
1.0 + alpha * negative_scores.pow(gamma)
)
losses[task_name] = loss.mean()
total_loss += self.task_weights[task_name] * losses[task_name]
# Edge type classification loss (only for positive edges)
if edge_type_logits is not None:
pos_mask = edge_labels == 1
if pos_mask.sum() > 0:
type_loss = nn.CrossEntropyLoss()(
edge_type_logits[pos_mask],
edge_types[pos_mask]
)
losses["edge_type"] = type_loss
total_loss += self.task_weights["edge_type"] * type_loss
return total_loss, losses
def predict_edges(self, g, edge_pairs, threshold=0.5, candidate_meta=None):
"""Predict edges for all tasks"""
with torch.no_grad():
edge_logits, edge_type_logits = self.forward(
g,
edge_pairs,
candidate_meta=candidate_meta,
)
predictions = {}
probabilities = {}
for task in self.edge_type_mapping:
probs = torch.sigmoid(edge_logits[task].squeeze())
preds = (probs >= threshold)
predictions[task] = preds
probabilities[task] = probs
# Edge type predictions
type_probs = torch.softmax(edge_type_logits, dim=1)
type_preds = torch.argmax(type_probs, dim=1)
return predictions, probabilities, type_preds, type_probs
def evaluate(self, g, pos_edges, pos_types, neg_edges, threshold=0.5, candidate_meta=None):
"""Evaluate model performance on all tasks"""
# Handle different input formats like SingleTaskGAT
if isinstance(pos_edges, torch.Tensor):
pos_edges_tensor = pos_edges
else:
pos_edges_tensor = torch.tensor(pos_edges, dtype=torch.long, device=g.device)
if isinstance(neg_edges, torch.Tensor):
neg_edges_tensor = neg_edges
else:
neg_edges_tensor = torch.tensor(neg_edges, dtype=torch.long, device=g.device)
# Combine 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)
])
# Get predictions
predictions, probabilities, type_preds, type_probs = self.predict_edges(
g,
all_edges,
threshold,
candidate_meta=candidate_meta,
)
# Calculate metrics for each task
results = {}
for task_name, type_id in self.edge_type_mapping.items():
type_gt = (all_types == type_id) & (all_labels == 1)
task_preds = predictions[task_name]
tp = (task_preds & type_gt).sum().item()
fp = (task_preds & ~type_gt).sum().item()
fn = (~task_preds & 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)
results[task_name] = {
"precision": precision,
"recall": recall,
"f1": f1,
"support": type_gt.sum().item()
}
return results
def train_step(self, g, pos_edges, pos_types, neg_edges, optimizer, pos_weight=1.0, candidate_meta=None):
"""Training step with improved input handling"""
# Handle different input formats like SingleTaskGAT
if isinstance(pos_edges, torch.Tensor):
pos_edges_tensor = pos_edges
else:
pos_edges_tensor = torch.tensor(pos_edges, dtype=torch.long, device=g.device)
if isinstance(neg_edges, torch.Tensor):
neg_edges_tensor = neg_edges
else:
neg_edges_tensor = torch.tensor(neg_edges, dtype=torch.long, device=g.device)
# Combine edges - use [num_edges, 2] format
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, device=g.device),
torch.full((len(neg_edges),), -1, dtype=torch.long, device=g.device)
])
# Forward pass
edge_logits, edge_type_logits = self.forward(
g,
edge_pairs,
candidate_meta=candidate_meta,
)
# Compute loss
loss, type_losses = self.get_loss(edge_logits, edge_labels, edge_types, pos_weight, edge_type_logits)
# Backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss.item(), {k: v.item() for k, v in type_losses.items()}
|