File size: 11,654 Bytes
c28f0c7 | 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 | import os
import sys
import argparse
from typing import List, Dict, Optional
from matplotlib import pyplot as plt
from tqdm import tqdm
import numpy as np
import pandas as pd
import torch
import faiss
from pathlib import Path
_PROJECT_ROOT = str(Path.cwd())
_RESULTS_DIR = os.path.join(_PROJECT_ROOT, "results")
if _PROJECT_ROOT not in sys.path:
sys.path.insert(0, _PROJECT_ROOT)
from src.metrics import roc
from src.utils import last_token_stack,MODEL2HF,DATA2HF,MODEL2LAYER
# Defaults aligned with xp_prj1/get_activation.py
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(description="Compute LID metrics over saved activations")
parser.add_argument(
"-m",
"--model",
default="llama_instruct",
choices=MODEL2HF.keys(),
)
parser.add_argument(
"-d",
"--dataset",
default="hotpotqa",
choices=DATA2HF.keys(),
)
parser.add_argument(
"--basepath2prepared_data",
type=str,
default=os.path.join(_PROJECT_ROOT, "prepared_data"),
help="Base path where get_activation.py wrote activations and labels",
)
parser.add_argument(
"-k",
"--num-neighbors",
type=int,
default=50,
dest="num_neighbors",
help="Number of nearest neighbors used for LID computation.",
)
# Output directory for metrics CSVs
parser.add_argument(
"--results-base",
dest="results_base",
type=str,
default=_RESULTS_DIR,
help="Base directory to write results/metrics.csv",
)
parser.add_argument("--zero-shot", type=bool, default=True)
parser.add_argument("--all_data",action="store_true",help='if set, run LID over all datasets')
return parser
class LID:
"""
Local Intrinsic Dimensionality (LID) scorer.
Mirrors the computation in LID-HallucinationDetection/calculate_lid.py, but
packaged as a reusable class and aligned with paths produced by
xp_prj1/get_activation.py.
Usage example:
lid = LID(basepath2prepared_data=..., model="llama2-7b", dataset="triviaqa", num_neighbors=500)
df = lid.compute_per_layer(write_csv=True)
"""
def __init__(
self,
args: Optional[argparse.Namespace] = None,
) -> None:
"""Initialize LID configuration.
You can either pass parameters directly, or provide an argparse
Namespace via `args`, or omit both and this will call build_parser().parse_args().
Explicit kwargs, when provided, override values from `args`.
"""
# Prefer explicit kwargs, otherwise fall back to args
self.args=args
self.basepath2prepared_data = getattr(args, "basepath2prepared_data")
self.model = getattr(args, "model")
self.dataset = getattr(args, "dataset")
self.num_neighbors = int(getattr(args, "num_neighbors"))
self.zero_shot = bool(getattr(args, "zero_shot"))
self.results_base = getattr(args, "results_base")
# Resolve paths
self.base = os.path.join(self.basepath2prepared_data, self.model, self.dataset)
self.results_dir = os.path.join(self.results_base, self.model, self.dataset,f'{self.__class__.__name__}')
self.train_set_dir=os.path.join(self.base, "train",'activations')
self.train_acts_dir=os.path.join(self.train_set_dir,'predicted')
self.train_labels_dir=os.path.join(self.train_set_dir,'labels')
self.test_set_dir=os.path.join(self.base, "test",'activations')
self.test_acts_dir=os.path.join(self.test_set_dir,'predicted')
self.test_labels_dir=os.path.join(self.test_set_dir,'labels')
self.log_name=f"lid_metrics_{self.args.num_neighbors}"
torch.manual_seed(42)
np.random.seed(42)
# ------------------------
# Public API
# ------------------------
def _load_layer_data(self, layer: int):
train_acts_path=os.path.join(
self.train_acts_dir, f"{self.model}_{self.dataset}_layer_{layer}_pred.pt"
)
train_label_path=os.path.join(
self.train_labels_dir, f"{self.model}_{self.dataset}_all_layer_label.pt"
)
test_acts_path=os.path.join(
self.test_acts_dir, f"{self.model}_{self.dataset}_layer_{layer}_pred.pt"
)
test_label_path=os.path.join(
self.test_labels_dir, f"{self.model}_{self.dataset}_all_layer_label.pt"
)
train_acts= torch.load(train_acts_path,weights_only=True)
test_acts= torch.load(test_acts_path,weights_only=True)
train_labels= torch.load(train_label_path,weights_only=True)
test_labels= torch.load(test_label_path,weights_only=True)
return train_acts, train_labels, test_acts, test_labels
def _compute_for_layer(self, layer: int) -> Optional[Dict[str, float]]:
"""
Computes LID-related metrics for a single layer and returns a row dict containing:
- 'layer'
- 'auroc'
- 'lid_mean'
- 'lid_std'
Ready to be appended directly in compute_per_layer.
"""
train_acts, train_labels, test_acts, test_labels = self._load_layer_data(layer)
# use only label=1 train_acts as the reference set
pos_mask = (train_labels == 1)
pos_train_acts = train_acts[pos_mask]
# convert to numpy
pos_train_acts = pos_train_acts.to(dtype=torch.float32, device="cpu").numpy()
test_acts = test_acts.to(dtype=torch.float32, device="cpu").numpy()
test_labels = test_labels.numpy()
if pos_train_acts.shape[0] == 0:
# no positive samples in edge case — skip this layer
return None
k = int(min(pos_train_acts.shape[0], self.num_neighbors))
lids = self._lid_new(
feats=test_acts,
num_neighbors=[k],
ref_feats=[pos_train_acts],
)
lids = np.asarray(lids)
auroc = roc(test_labels, -lids)
lid_mean = float(np.mean(lids))
lid_std = float(np.std(lids))
# return the row dict directly (including the layer field)
return {
"layer": int(layer),
"auroc": float(auroc),
"lid_mean": lid_mean,
"lid_std": lid_std,
}
def _plot_per_layer_metrics(self, df) -> None:
"""
Given a DataFrame with 'layer', 'lid_mean', 'auroc' columns, plots a dual-axis figure and saves it.
Subclasses can override this method to plot different metrics:
def _plot_per_layer_metrics(self, df, basename):
...
"""
if df.empty:
return
layers = df["layer"].values
lid_mean = df["lid_mean"].values
auroc = df["auroc"].values
fig, ax1 = plt.subplots(figsize=(8, 5))
# left axis: LID mean
ax1.plot(layers, lid_mean, marker="o", linestyle="-", label="LID mean")
ax1.set_xlabel("Layer")
ax1.set_ylabel("LID mean")
ax1.grid(True, linestyle="--", alpha=0.5)
# right axis: AUROC
ax2 = ax1.twinx()
ax2.plot(layers, auroc, marker="s", linestyle="-", label="AUROC", color="tab:orange")
ax2.set_ylabel("AUROC")
# merge legend
lines1, labels1 = ax1.get_legend_handles_labels()
lines2, labels2 = ax2.get_legend_handles_labels()
ax1.legend(lines1 + lines2, labels1 + labels2, loc="best")
fig.tight_layout()
fig_path = os.path.join(self.results_dir, f"{self.log_name}.png")
plt.savefig(fig_path, bbox_inches="tight")
plt.close()
print(f"Saved LID metrics figure to {fig_path}")
# ------------------------
# Public API
# ------------------------
def compute_per_layer(self) -> pd.DataFrame:
"""
Compute LID scores for each saved layer and return a DataFrame with
per-layer AUROC and LID statistics.
"""
layer_ids = list(range(MODEL2LAYER[self.model]))
metrics_rows = []
for layer in tqdm(layer_ids):
row = self._compute_for_layer(layer)
if row is not None:
metrics_rows.append(row)
df = pd.DataFrame(metrics_rows)
self.save_data(df)
print(f"{self.__class__.__name__} computation complete for model={self.model}, dataset={self.dataset}")
return df
def save_data(self,df):
os.makedirs(self.results_dir, exist_ok=True)
csv_path = os.path.join(self.results_dir, f"{self.log_name}.csv")
df.to_csv(csv_path, index=False)
print(f"Saved {self.__class__.__name__} metrics CSV to {csv_path}")
self._plot_per_layer_metrics(df)
# ------------------------
# Core LID kernels (ported)
# ------------------------
@staticmethod
def _lid_new(
feats: np.ndarray,
num_neighbors: List[int],
faiss_instances: Optional[List[faiss.IndexFlatL2]] = None,
ref_feats: Optional[List[np.ndarray]] = None,
) -> np.ndarray:
"""
Compute LID of `feats` using `ref_feats` as the reference points.
Returns a numpy array of shape (batch,).
Ported from LID-HallucinationDetection/calculate_lid.py:_lid_new
"""
if faiss_instances is None:
if ref_feats is None:
raise ValueError("Provide either faiss_instances or ref_feats.")
faiss_instances = []
for rf in ref_feats:
inst = faiss.IndexFlatL2(rf.shape[1])
inst.add(np.ascontiguousarray(rf))
faiss_instances.append(inst)
else:
if ref_feats is not None:
# Prefer the provided FAISS indices; ignore ref_feats
pass
k_avg_lids = [] # (num_ks, batch)
for k in num_neighbors:
bootstrap_avg_lids = [] # (num_bootstrap, batch)
for faiss_instance in faiss_instances:
b, _ = faiss_instance.search(feats, k)
b = np.sqrt(b)
# Remove self-distance if present
filtered = []
for row in b:
if row[0] < 1e-6:
filtered.append(np.expand_dims(row[1:], axis=0))
else:
filtered.append(np.expand_dims(row[:-1], axis=0))
D = np.concatenate(filtered, axis=0)
rk = np.max(D, axis=1)
rk[rk == 0] = 1e-8
lids = D / rk[:, None]
lids = np.clip(lids, 1e-6, 1.0)
logs = np.log(lids)
denom = np.mean(logs, axis=1)
denom = np.where(np.abs(denom) < 1e-5, 1e-5 * np.sign(denom), denom)
lids = -1.0 / denom
lids[np.isinf(lids)] = feats.shape[1]
lids = np.nan_to_num(lids, nan=feats.shape[1])
bootstrap_avg_lids.append(lids.tolist())
bootstrap_avg_lids = np.array(bootstrap_avg_lids).mean(axis=0)
k_avg_lids.append(bootstrap_avg_lids.tolist())
k_avg_lids = np.array(k_avg_lids).mean(axis=0)
return k_avg_lids
if __name__ == "__main__":
parser = build_parser()
args = parser.parse_args()
if args.all_data:
for dataset_name in ['coqa','hotpotqa','squad','triviaqa','psiloqa','math']:
args.dataset=dataset_name
lid=LID(args=args)
df=lid.compute_per_layer()
else:
lid = LID(args=args)
df = lid.compute_per_layer()
# print(df)
|