Datasets:
File size: 4,925 Bytes
2b3cfd1 | 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 | #!/usr/bin/env python3
"""
Scree plot of the NLI accuracy matrix (singular values + cumulative energy).
Uses all_results_summary_fixed.json (45 NLI models x 12 datasets after dropping
models with any cell < min_cell).
Matches ablation_layers_link.png in figsize/fontsize.
Usage:
python scripts/plot_nli_matrix_scree.py
"""
import argparse
import json
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
def build_matrix(results):
models = sorted({r["model_id"] for r in results})
datasets = sorted({r["dataset_id"] for r in results})
m_idx = {m: i for i, m in enumerate(models)}
d_idx = {d: i for i, d in enumerate(datasets)}
M = np.full((len(datasets), len(models)), np.nan, dtype=float)
for r in results:
if r.get("masked"):
continue
M[d_idx[r["dataset_id"]], m_idx[r["model_id"]]] = r["accuracy"]
return M, datasets, models
def drop_low_cell_models(M, models, threshold):
# Ignore NaN (masked) cells in the threshold check
bad = (M < threshold) & ~np.isnan(M)
keep = ~bad.any(axis=0)
return M[:, keep], [m for m, k in zip(models, keep) if k]
def plot_scree(M, out_png, out_pdf):
# Impute masked cells with per-row mean before SVD (keeps them from shifting the mean)
row_mean = np.nanmean(M, axis=1, keepdims=True)
M_imp = np.where(np.isnan(M), row_mean, M)
# Double-centering: remove additive row (dataset-difficulty) and column
# (model-strength) effects; what remains is the interaction matrix.
row_m = M_imp.mean(axis=1, keepdims=True)
col_m = M_imp.mean(axis=0, keepdims=True)
grand = M_imp.mean()
Mc = M_imp - row_m - col_m + grand
U, S, Vt = np.linalg.svd(Mc, full_matrices=False)
# Drop the last (near-zero) singular value — double-centering makes
# the matrix rank-deficient by 1 so the 12th component is meaningless.
S = S[:-1]
energy = (S ** 2) / (S ** 2).sum()
cum = np.cumsum(energy)
# Match ablation_layers_link.png style
plt.rcParams["font.family"] = "sans-serif"
plt.rcParams["font.sans-serif"] = ["Helvetica", "Arial", "DejaVu Sans"]
plt.rcParams["font.size"] = 20
plt.rcParams["axes.spines.top"] = False
plt.rcParams["axes.linewidth"] = 2.0
fig, ax1 = plt.subplots(figsize=(6, 6))
ax1.spines["top"].set_visible(False)
bar_color = "#E74C3C"
line_color = "#2C3E50"
ks = np.arange(1, len(S) + 1)
ax1.bar(ks, S, color=bar_color, alpha=0.85, edgecolor="white", linewidth=0.8)
ax1.set_yscale("log")
ax1.set_xlabel("Rank $k$", fontsize=22, labelpad=8)
ax1.set_ylabel(r"Singular value $\sigma_k$", fontsize=22, color=bar_color, labelpad=8)
ax1.set_yticks([0.1, 0.5, 1.0])
ax1.set_yticklabels(["0.1", "0.5", "1"])
ax1.yaxis.set_minor_locator(plt.NullLocator())
ax1.tick_params(axis="y", labelcolor=bar_color, labelsize=14)
ax1.tick_params(axis="x", labelsize=14)
ax1.set_xticks(ks)
ax2 = ax1.twinx()
ax2.spines["top"].set_visible(False)
ax2.plot(ks, cum, "o-", color=line_color, linewidth=2.2, markersize=7)
ax2.set_ylim(0, 1.05)
ax2.set_ylabel("Cumulative energy", fontsize=22, color=line_color, labelpad=8)
ax2.tick_params(axis="y", labelcolor=line_color, labelsize=14)
# σ_1 annotation
ax1.annotate(rf"$\sigma_1 = {S[0]:.2f}$", xy=(1, S[0]),
xytext=(2.2, S[0] * 1.02), fontsize=18, color=bar_color)
# 90% line
ax2.axhline(0.9, color="#999999", linestyle="--", linewidth=1.2, alpha=0.8)
ax2.text(len(S) - 0.2, 0.91, "90%", fontsize=14, color="#777777",
ha="right", va="bottom")
fig.subplots_adjust(left=0.17, right=0.84, bottom=0.14, top=0.96)
fig.savefig(out_png, dpi=300)
fig.savefig(out_pdf, dpi=300)
plt.close(fig)
print(f"Singular values: {S}")
print(f"Cumulative energy: {cum}")
def main():
p = argparse.ArgumentParser()
p.add_argument("--input", default="all_results_summary_fixed.json")
p.add_argument("--out-dir", default="data/figures")
p.add_argument("--stem", default="nli_matrix_scree")
p.add_argument("--min-cell", type=float, default=0.05,
help="Drop models with any cell below this threshold")
args = p.parse_args()
root = Path(__file__).resolve().parent.parent
in_path = (root / args.input).resolve()
out_dir = (root / args.out_dir).resolve()
out_dir.mkdir(parents=True, exist_ok=True)
data = json.loads(in_path.read_text())
M, datasets, models = build_matrix(data["results"])
if args.min_cell > 0:
M, models = drop_low_cell_models(M, models, args.min_cell)
print(f"Matrix shape: {M.shape} ({len(datasets)} datasets x {len(models)} models)")
png = out_dir / f"{args.stem}.png"
pdf = out_dir / f"{args.stem}.pdf"
plot_scree(M, png, pdf)
print(f"Saved: {png}")
print(f"Saved: {pdf}")
if __name__ == "__main__":
main()
|