Spaces:
Sleeping
Sleeping
File size: 12,930 Bytes
8125804 | 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 | from __future__ import annotations
from dataclasses import dataclass
import math
from typing import Any
import torch
import torch.nn.functional as F
from src.model.anchor_types import AnchorRecord
from src.model.future_influence import FutureInfluenceScorer
@dataclass
class AnchorDependencyEdge:
source_id: int
target_id: int
approx_score: float
final_score: float
similarity: float
temporal_prior: float
support_prior: float
viability_prior: float
refined_delta: float = 0.0
is_refined: bool = False
@dataclass
class AnchorDependencyNode:
anchor_id: int
validity: float
soft_confirmation: float
node_pressure: float
predecessor_ids: list[int]
broken_predecessor_ids: list[int]
def _to_float(value: torch.Tensor | float | int | None) -> float:
if value is None:
return 0.0
if isinstance(value, torch.Tensor):
return float(value.detach().item())
return float(value)
def _sigmoid_unit(value: float, threshold: float, slope: float) -> float:
safe_slope = max(float(slope), 1e-6)
tensor = torch.tensor((float(value) - float(threshold)) / safe_slope, dtype=torch.float32)
return float(torch.sigmoid(tensor).item())
def _temporal_prior(source: AnchorRecord, target: AnchorRecord, temporal_window: float) -> float:
distance = max(1, int(target.start_idx) - int(source.end_idx))
return float(math.exp(-(float(distance) - 1.0) / max(float(temporal_window), 1e-6)))
def _approx_dependency_score(
source: AnchorRecord,
target: AnchorRecord,
*,
confirm_threshold: float,
similarity_weight: float,
temporal_weight: float,
support_weight: float,
viability_weight: float,
temporal_window: float,
) -> tuple[float, dict[str, float]]:
source_repr = F.normalize(source.repr.detach().float().unsqueeze(0), dim=-1)
target_repr = F.normalize(target.repr.detach().float().unsqueeze(0), dim=-1)
similarity = max(0.0, float(F.cosine_similarity(source_repr, target_repr, dim=-1).item()))
temporal = _temporal_prior(source, target, temporal_window)
support = min(1.0, max(0.0, 0.5 * (_to_float(source.support) + _to_float(target.support))))
viability = min(
1.0,
max(
0.0,
0.5
* (
_sigmoid_unit(_to_float(source.support), confirm_threshold, 0.10)
+ _to_float(target.viability)
),
),
)
total_weight = max(
float(similarity_weight) + float(temporal_weight) + float(support_weight) + float(viability_weight),
1e-6,
)
score = (
float(similarity_weight) * similarity
+ float(temporal_weight) * temporal
+ float(support_weight) * support
+ float(viability_weight) * viability
) / total_weight
return float(score), {
"similarity": float(similarity),
"temporal_prior": float(temporal),
"support_prior": float(support),
"viability_prior": float(viability),
}
def _compute_counterfactual_scores(
*,
anchors: list[AnchorRecord],
candidate_edges: list[AnchorDependencyEdge],
hidden: torch.Tensor | None,
input_ids: torch.Tensor | None,
attention_mask: torch.Tensor | None,
output_projection: torch.nn.Module | None,
future_scorer: FutureInfluenceScorer | None,
future_window: int,
max_edges: int,
) -> dict[tuple[int, int], float]:
if not candidate_edges or hidden is None or input_ids is None or output_projection is None or future_scorer is None:
return {}
if hidden.ndim != 3 or hidden.size(0) != 1 or input_ids.ndim != 2:
return {}
edge_map = {(edge.source_id, edge.target_id): edge for edge in candidate_edges}
top_edges = sorted(candidate_edges, key=lambda item: item.approx_score, reverse=True)[: max(0, int(max_edges))]
if not top_edges:
return {}
anchor_by_id = {anchor.id: anchor for anchor in anchors}
base_hidden = hidden.detach().clone().requires_grad_(True)
base_logits = output_projection(base_hidden)
base_scores = future_scorer(
hidden=base_hidden,
logits=base_logits,
input_ids=input_ids,
attention_mask=attention_mask,
future_window=future_window,
)["scores"].detach()
deltas: dict[tuple[int, int], float] = {}
unique_source_ids = sorted({edge.source_id for edge in top_edges})
for source_id in unique_source_ids:
source_anchor = anchor_by_id.get(source_id)
if source_anchor is None:
continue
masked_hidden = hidden.detach().clone()
start = max(0, int(source_anchor.start_idx))
end = min(masked_hidden.size(1) - 1, int(source_anchor.end_idx))
masked_hidden[:, start : end + 1, :] = 0.0
masked_hidden = masked_hidden.requires_grad_(True)
masked_logits = output_projection(masked_hidden)
masked_scores = future_scorer(
hidden=masked_hidden,
logits=masked_logits,
input_ids=input_ids,
attention_mask=attention_mask,
future_window=future_window,
)["scores"].detach()
for edge in top_edges:
if edge.source_id != source_id:
continue
target_anchor = anchor_by_id.get(edge.target_id)
if target_anchor is None:
continue
target_start = max(0, int(target_anchor.start_idx))
target_end = min(base_scores.size(1) - 1, int(target_anchor.end_idx))
if target_end < target_start:
continue
delta = (
base_scores[:, target_start : target_end + 1] - masked_scores[:, target_start : target_end + 1]
).abs().mean()
deltas[(edge.source_id, edge.target_id)] = float(torch.nan_to_num(delta, nan=0.0, posinf=1.0, neginf=0.0).item())
return deltas
def build_anchor_dependency_graph(
anchors: list[AnchorRecord],
*,
confirm_threshold: float,
dependency_threshold: float = 0.55,
confirm_slope: float = 0.10,
similarity_weight: float = 0.55,
temporal_weight: float = 0.20,
support_weight: float = 0.15,
viability_weight: float = 0.10,
temporal_window: float = 16.0,
max_predecessors: int = 4,
counterfactual_top_edges: int = 0,
future_scorer: FutureInfluenceScorer | None = None,
hidden: torch.Tensor | None = None,
input_ids: torch.Tensor | None = None,
attention_mask: torch.Tensor | None = None,
output_projection: torch.nn.Module | None = None,
future_window: int = 16,
) -> dict[str, Any]:
if not anchors:
return {
"edges": [],
"nodes": [],
"graph_pressure": 0.0,
"current_graph_pressure": 0.0,
"current_anchor_id": None,
"edge_count": 0,
"broken_anchor_count": 0,
"mean_validity": 1.0,
}
sorted_anchors = sorted(anchors, key=lambda item: (int(item.start_idx), int(item.end_idx), int(item.id)))
anchor_by_id = {anchor.id: anchor for anchor in sorted_anchors}
edges_by_target: dict[int, list[AnchorDependencyEdge]] = {anchor.id: [] for anchor in sorted_anchors}
candidate_edges: list[AnchorDependencyEdge] = []
for source in sorted_anchors:
for target in sorted_anchors:
if int(source.end_idx) >= int(target.start_idx) or source.id == target.id:
continue
approx_score, parts = _approx_dependency_score(
source,
target,
confirm_threshold=confirm_threshold,
similarity_weight=similarity_weight,
temporal_weight=temporal_weight,
support_weight=support_weight,
viability_weight=viability_weight,
temporal_window=temporal_window,
)
if approx_score < float(dependency_threshold):
continue
candidate_edges.append(
AnchorDependencyEdge(
source_id=source.id,
target_id=target.id,
approx_score=float(approx_score),
final_score=float(approx_score),
similarity=float(parts["similarity"]),
temporal_prior=float(parts["temporal_prior"]),
support_prior=float(parts["support_prior"]),
viability_prior=float(parts["viability_prior"]),
)
)
deltas = _compute_counterfactual_scores(
anchors=sorted_anchors,
candidate_edges=candidate_edges,
hidden=hidden,
input_ids=input_ids,
attention_mask=attention_mask,
output_projection=output_projection,
future_scorer=future_scorer,
future_window=future_window,
max_edges=counterfactual_top_edges,
)
for edge in candidate_edges:
edge_key = (edge.source_id, edge.target_id)
delta = float(deltas.get(edge_key, 0.0))
if edge_key in deltas:
edge.refined_delta = delta
edge.is_refined = True
edge.final_score = 0.5 * float(edge.approx_score) + 0.5 * min(1.0, max(0.0, delta))
for anchor in sorted_anchors:
incoming = [edge for edge in candidate_edges if edge.target_id == anchor.id]
incoming.sort(key=lambda item: item.final_score, reverse=True)
edges_by_target[anchor.id] = incoming[: max(1, int(max_predecessors))] if incoming else []
nodes: list[AnchorDependencyNode] = []
node_by_id: dict[int, AnchorDependencyNode] = {}
for anchor in sorted_anchors:
soft_confirmation = _sigmoid_unit(_to_float(anchor.support), confirm_threshold, confirm_slope)
predecessors = edges_by_target[anchor.id]
if predecessors:
total = sum(edge.final_score for edge in predecessors)
weighted_confirmation = sum(
edge.final_score * _sigmoid_unit(
_to_float(anchor_by_id[edge.source_id].support),
confirm_threshold,
confirm_slope,
)
for edge in predecessors
) / max(total, 1e-6)
else:
weighted_confirmation = 1.0
broken_predecessors = [
edge.source_id
for edge in predecessors
if _sigmoid_unit(
_to_float(anchor_by_id[edge.source_id].support),
confirm_threshold,
confirm_slope,
)
< 0.5
]
node_pressure = 1.0 - float(weighted_confirmation) * min(1.0, max(0.0, _to_float(anchor.viability)))
node = AnchorDependencyNode(
anchor_id=anchor.id,
validity=float(weighted_confirmation),
soft_confirmation=float(soft_confirmation),
node_pressure=float(min(1.0, max(0.0, node_pressure))),
predecessor_ids=[edge.source_id for edge in predecessors],
broken_predecessor_ids=broken_predecessors,
)
nodes.append(node)
node_by_id[anchor.id] = node
current_anchor = max(sorted_anchors, key=lambda item: (int(item.end_idx), int(item.start_idx), int(item.id)))
current_graph_pressure = float(node_by_id[current_anchor.id].node_pressure)
graph_pressure = max((node.node_pressure for node in nodes), default=0.0)
return {
"edges": [
{
"source_id": edge.source_id,
"target_id": edge.target_id,
"approx_score": edge.approx_score,
"final_score": edge.final_score,
"similarity": edge.similarity,
"temporal_prior": edge.temporal_prior,
"support_prior": edge.support_prior,
"viability_prior": edge.viability_prior,
"refined_delta": edge.refined_delta,
"is_refined": edge.is_refined,
}
for target_edges in edges_by_target.values()
for edge in target_edges
],
"nodes": [
{
"anchor_id": node.anchor_id,
"validity": node.validity,
"soft_confirmation": node.soft_confirmation,
"node_pressure": node.node_pressure,
"predecessor_ids": node.predecessor_ids,
"broken_predecessor_ids": node.broken_predecessor_ids,
}
for node in nodes
],
"graph_pressure": float(graph_pressure),
"current_graph_pressure": float(current_graph_pressure),
"current_anchor_id": int(current_anchor.id),
"edge_count": int(sum(len(edges) for edges in edges_by_target.values())),
"broken_anchor_count": int(sum(1 for node in nodes if node.broken_predecessor_ids)),
"mean_validity": float(sum(node.validity for node in nodes) / max(len(nodes), 1)),
}
|