deepellipse-circleid / code /ay_infer.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
11.4 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""AY refine: A from BCE head + fixed Y → Top-K / r rejection CSV."""
from __future__ import annotations
import json
import os
from typing import Dict, List, Optional, Sequence, Tuple
import numpy as np
import pandas as pd
import torch
from sklearn.preprocessing import LabelEncoder
def load_y_logits_json(path: str) -> Tuple[List[str], np.ndarray]:
"""Load deep_ellipse_logits.json → (image_ids, Y[N,C])."""
with open(path, "r", encoding="utf-8") as f:
rows = json.load(f)
image_ids = [str(r["image_id"]) for r in rows]
y = np.asarray([r["logits"] for r in rows], dtype=np.float32)
return image_ids, y
def otsu_threshold_1d(values: np.ndarray) -> float:
"""Otsu on 1-D float array (off-diagonal affinities)."""
v = np.asarray(values, dtype=np.float64).ravel()
v = v[np.isfinite(v)]
if v.size == 0:
return 0.0
vmin, vmax = float(v.min()), float(v.max())
if vmax - vmin < 1e-12:
return vmin
hist, bin_edges = np.histogram(v, bins=256, range=(vmin, vmax))
hist = hist.astype(np.float64)
total = hist.sum()
if total <= 0:
return 0.5 * (vmin + vmax)
bin_centers = 0.5 * (bin_edges[:-1] + bin_edges[1:])
weight1 = np.cumsum(hist)
weight2 = total - weight1
sum_all = np.cumsum(hist * bin_centers)
mean1 = sum_all / np.maximum(weight1, 1e-12)
mean2 = (sum_all[-1] - sum_all) / np.maximum(weight2, 1e-12)
inter = weight1[:-1] * weight2[:-1] * (mean1[:-1] - mean2[:-1]) ** 2
idx = int(np.argmax(inter))
return float(bin_centers[idx])
def build_a_norm(
a: np.ndarray,
*,
power_k: float = 2.0,
alpha: Optional[float] = None,
sample_offdiag: int = 2_000_000,
clamp_neg: bool = True,
) -> Tuple[np.ndarray, float]:
"""
A'[i,j] = max(0, A[i,j] - α) for i != j; diagonal restored as A[i,i];
A_norm = row-normalize (A')^k.
α from Otsu on off-diagonal (optionally subsampled).
clamp_neg=True (default): truncate the shifted affinities at 0, i.e. the
current paper/code construction A' = max(0, A - α) on the off-diagonal.
clamp_neg=False: keep the signed shift A' = A - α (diagonal still restored),
which is the literal main-text formula A' = (A - α) + αI. Because the power
^k is applied AFTER the shift, negative entries fold back to positive for
even k (e.g. k=2: A_ij=0.01 and 0.99 both give 0.2401). A non-integer
power_k with clamp_neg=False would raise (negative base → NaN), so it is
rejected up front.
"""
a = np.asarray(a, dtype=np.float32)
n = a.shape[0]
assert a.shape == (n, n)
if not clamp_neg and float(power_k) != int(power_k):
raise ValueError(
f"clamp_neg=False requires an integer power_k (got {power_k!r}); "
"a fractional power of a negative base is undefined."
)
if alpha is None:
# sample off-diagonal for Otsu
if n * (n - 1) > sample_offdiag:
rng = np.random.default_rng(0)
i = rng.integers(0, n, size=sample_offdiag)
j = rng.integers(0, n, size=sample_offdiag)
mask = i != j
off = a[i[mask], j[mask]]
else:
eye = np.eye(n, dtype=bool)
off = a[~eye]
alpha = otsu_threshold_1d(off)
ap = a - float(alpha)
np.fill_diagonal(ap, a.diagonal()) # restore diag: (A-α)+αI → diag stays A_ii
# Ensure non-negative for power (clamp) unless the un-clamped variant is asked
if clamp_neg:
ap = np.maximum(ap, 0.0)
if power_k != 1.0:
ap = np.power(ap, power_k)
row_sum = ap.sum(axis=1, keepdims=True)
row_sum = np.maximum(row_sum, 1e-12)
a_norm = ap / row_sum
return a_norm.astype(np.float32), float(alpha)
def row_knn_lambda_thresholds(a: np.ndarray, *, knn_n: int = 50) -> np.ndarray:
"""
Per-row lambda = (knn_n+1)-th largest off-diagonal affinity (1-indexed rank).
For knn_n=50, lambda_i is the 51st largest value in row i (excluding diagonal).
"""
a = np.asarray(a, dtype=np.float32)
n = a.shape[0]
assert a.shape == (n, n)
rank_idx = int(knn_n) # 0-based index of (knn_n+1)-th largest
lambdas = np.zeros(n, dtype=np.float32)
for i in range(n):
row = a[i].copy()
row[i] = -np.inf
valid = row[np.isfinite(row) & (row > -np.inf)]
if valid.size <= rank_idx:
lambdas[i] = float(np.min(valid)) if valid.size else 0.0
else:
sorted_desc = np.sort(valid)[::-1]
lambdas[i] = float(sorted_desc[rank_idx])
return lambdas
def build_a_norm_knn_row_lambda(
a: np.ndarray,
*,
knn_n: int = 50,
power_k: float = 2.0,
) -> Tuple[np.ndarray, np.ndarray]:
"""
A'[i,j] = max(0, A[i,j] - lambda_i) for j != i; diag stays A_ii;
A_norm = row-normalize (A')^k with per-row lambda from row_knn_lambda_thresholds.
"""
a = np.asarray(a, dtype=np.float32)
n = a.shape[0]
assert a.shape == (n, n)
lambdas = row_knn_lambda_thresholds(a, knn_n=knn_n)
ap = a - lambdas[:, None]
np.fill_diagonal(ap, a.diagonal())
ap = np.maximum(ap, 0.0)
if power_k != 1.0:
ap = np.power(ap, power_k)
row_sum = ap.sum(axis=1, keepdims=True)
row_sum = np.maximum(row_sum, 1e-12)
a_norm = ap / row_sum
return a_norm.astype(np.float32), lambdas
def ay_refine(
y_raw: np.ndarray, a_norm: np.ndarray, *, graph_hops: int = 1
) -> np.ndarray:
"""Y_refined = A_norm^graph_hops @ Y_raw."""
if graph_hops < 1:
raise ValueError(f"graph_hops must be >= 1, got {graph_hops!r}")
y = y_raw
for _ in range(graph_hops):
y = a_norm @ y
return y
def decode_writer_id(class_idx: int, label_encoder: LabelEncoder) -> str:
return str(label_encoder.classes_[class_idx])
def known_class_indices(label_encoder: LabelEncoder) -> np.ndarray:
"""Indices of known writers: every class slot except the unknown '-1'."""
idxs = [i for i, c in enumerate(label_encoder.classes_) if str(c) != "-1"]
return np.asarray(idxs, dtype=np.int64)
def scores_and_pred_from_y(
y_refined: np.ndarray,
label_encoder: LabelEncoder,
*,
reject_score: str = "logit",
) -> Tuple[np.ndarray, np.ndarray]:
"""
Returns (scores[N], pred_cls[N]).
reject_score='logit': max over known cols of Y_ref
reject_score='softmax': max softmax prob over known cols
"""
if reject_score not in ("logit", "softmax"):
raise ValueError(f"reject_score must be logit|softmax, got {reject_score!r}")
kidx = known_class_indices(label_encoder)
y_k = np.asarray(y_refined, dtype=np.float64)[:, kidx]
if reject_score == "softmax":
y_k = y_k - y_k.max(axis=1, keepdims=True)
exp = np.exp(y_k)
probs = exp / np.maximum(exp.sum(axis=1, keepdims=True), 1e-12)
local = probs.argmax(axis=1)
scores = probs.max(axis=1)
else:
local = y_k.argmax(axis=1)
scores = y_k.max(axis=1)
pred_cls = kidx[local]
return scores.astype(np.float64), pred_cls.astype(np.int64)
def apply_topk_reject(
writers: Sequence[str],
conf: Sequence[float],
*,
reject_ratio: float = 0.65,
reject_score: str = "logit",
) -> Tuple[List[str], List[Optional[str]], Optional[float]]:
"""
Apply Top-r rejection on precomputed argmax writers + confidence scores.
Returns (final_writer_ids, reject_reasons, dyn_thresh).
"""
n = len(writers)
target_unknown = int(n * reject_ratio)
final = [str(w) for w in writers]
conf_arr = np.asarray(conf, dtype=np.float64)
reject_reason: List[Optional[str]] = [None] * n
n_natural = 0
for i in range(n):
if final[i] == "-1":
reject_reason[i] = "invalid_prediction"
n_natural += 1
remaining = [i for i in range(n) if final[i] != "-1"]
remaining_sorted = sorted(remaining, key=lambda i: conf_arr[i])
n_dyn = max(0, target_unknown - n_natural)
n_dyn = min(n_dyn, len(remaining_sorted))
thresh: Optional[float] = None
reason_tag = "low_confidence" if reject_score == "softmax" else "low_logit"
for i in remaining_sorted[:n_dyn]:
final[i] = "-1"
reject_reason[i] = reason_tag
thresh = float(conf_arr[i])
return final, reject_reason, thresh
def _write_reject_csv(
image_ids: Sequence[str],
final: Sequence[str],
conf: Sequence[float],
reject_reason: Sequence[Optional[str]],
*,
out_path: str,
reject_score: str,
thresh: Optional[float],
) -> pd.DataFrame:
n = len(image_ids)
df = pd.DataFrame(
{
"image_id": list(image_ids),
"writer_id": list(final),
"confidence": np.asarray(conf, dtype=np.float64),
"reject_reason": list(reject_reason),
}
)
os.makedirs(os.path.dirname(out_path) or ".", exist_ok=True)
df.to_csv(out_path, index=False)
clean = out_path.replace(".csv", "_submit.csv")
if clean == out_path:
clean = out_path[:-4] + "_submit.csv"
df[["image_id", "writer_id"]].to_csv(clean, index=False)
n_unk = sum(1 for w in final if w == "-1")
print(
f"AY Top-K CSV ({reject_score}): n={n}, unknown={n_unk} "
f"({n_unk / n * 100:.1f}%), "
f"dyn_thresh={thresh if thresh is not None else 'N/A'}"
)
print(f" saved: {out_path}")
print(f" submit: {clean}")
return df
def resweep_reject_from_scored(
image_ids: Sequence[str],
writers: Sequence[str],
conf: Sequence[float],
*,
reject_ratio: float = 0.65,
out_path: str,
reject_score: str = "logit",
) -> pd.DataFrame:
"""Re-apply Top-r on fixed confidence ranking (r-grid without re-inference)."""
final, reject_reason, thresh = apply_topk_reject(
writers, conf, reject_ratio=reject_ratio, reject_score=reject_score
)
return _write_reject_csv(
image_ids,
final,
conf,
reject_reason,
out_path=out_path,
reject_score=reject_score,
thresh=thresh,
)
def topk_reject_csv(
image_ids: Sequence[str],
y_refined: np.ndarray,
label_encoder: LabelEncoder,
*,
reject_ratio: float = 0.65,
out_path: str,
reject_score: str = "logit",
) -> pd.DataFrame:
"""
Score by max logit or max softmax (known classes only); reject lowest
reject_ratio as -1; else argmax → writer_id.
"""
scores, pred_cls = scores_and_pred_from_y(
y_refined, label_encoder, reject_score=reject_score
)
writers = [decode_writer_id(int(c), label_encoder) for c in pred_cls]
final, reject_reason, thresh = apply_topk_reject(
writers, scores, reject_ratio=reject_ratio, reject_score=reject_score
)
return _write_reject_csv(
image_ids,
final,
scores,
reject_reason,
out_path=out_path,
reject_score=reject_score,
thresh=thresh,
)
def align_y_to_ids(
y_ids: Sequence[str],
y: np.ndarray,
target_ids: Sequence[str],
) -> np.ndarray:
"""Reorder Y rows to match target_ids order."""
idx = {str(i): k for k, i in enumerate(y_ids)}
missing = [i for i in target_ids if str(i) not in idx]
if missing:
raise KeyError(f"Y logits missing {len(missing)} ids, e.g. {missing[:3]}")
order = [idx[str(i)] for i in target_ids]
return y[order]