File size: 2,139 Bytes
d4bcd5c | 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 | """Results figure: avg relative accuracy vs budget (SPLIT/attn/random + paper),
and TextVQA absolute accuracy vs budget (shows random collapse vs SPLIT robustness)."""
import json
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
agg = json.load(open("outputs/aggregate.json"))
rel = agg["relative_accuracy_avg"]
paper = agg["paper_reference"]["SPLIT"]
budgets = [192, 128, 64]
x = budgets[::-1]
tv = json.load(open("outputs/textvqa_results.json"))["results"]
van_tv = tv["vanilla@576"]["accuracy"]
fig, ax = plt.subplots(1, 2, figsize=(13, 4.6))
# panel 1: avg relative accuracy
def series(m): return [rel[m][str(b)] for b in x]
ax[0].plot(x, [paper[str(b)] for b in x], "s--", color="#888", label="SPLIT (paper, 10 bench)")
ax[0].plot(x, series("split"), "o-", color="#2D5F8B", lw=2.5, label="SPLIT (ours, 3 bench)")
ax[0].plot(x, series("attn"), "^-", color="#C1666B", label="attn-topk (ours)")
ax[0].plot(x, series("random"), "v-", color="#9AA0A6", label="random (ours)")
ax[0].set_xticks(x); ax[0].invert_xaxis()
ax[0].set_xlabel("retained vision tokens"); ax[0].set_ylabel("avg relative accuracy (%)")
ax[0].set_title("Avg relative accuracy vs token budget\n(LLaVA-1.5-7B, POPE+TextVQA+ScienceQA)")
ax[0].axhline(100, color="k", lw=0.5, ls=":"); ax[0].legend(fontsize=9); ax[0].grid(alpha=0.3)
# panel 2: TextVQA absolute
def tvs(m): return [tv[f"{m}@{b}"]["accuracy"] for b in x]
ax[1].axhline(van_tv, color="k", ls=":", lw=1, label=f"vanilla 576 = {van_tv:.1f}")
ax[1].plot(x, tvs("split"), "o-", color="#2D5F8B", lw=2.5, label="SPLIT")
ax[1].plot(x, tvs("attn"), "^-", color="#C1666B", label="attn-topk")
ax[1].plot(x, tvs("random"), "v-", color="#9AA0A6", label="random")
ax[1].set_xticks(x); ax[1].invert_xaxis()
ax[1].set_xlabel("retained vision tokens"); ax[1].set_ylabel("TextVQA accuracy (%)")
ax[1].set_title("TextVQA (OCR, token-sensitive)\nSPLIT robust; random collapses at 64")
ax[1].legend(fontsize=9); ax[1].grid(alpha=0.3)
plt.tight_layout()
plt.savefig("outputs/fig_relative_accuracy.png", dpi=130, bbox_inches="tight")
print("wrote outputs/fig_relative_accuracy.png")
|