fairtalking-second-work / scripts /analysis /draw_cta_framework.py
huahua123313's picture
Add files using upload-large-folder tool
eea47ad verified
Raw
History Blame Contribute Delete
11.5 kB
"""Render the CTA (A1_full) framework — cleaned, paper-grade version.
Output: outputs/analysis/figs_framework/cta_framework.{png,pdf}
Design principles:
* Two columns: left = data flow (single chain), right = side notes
(training losses + innovation list).
* Real/Fake sample paths use color (green / red) but only on the few
arrows that actually differ. Most arrows are shared (gray).
* Each innovation is marked with a single ★ inside the box; the
long captions live in the right-side panel, NOT on the diagram.
* Generous padding between modules. Total figure ~ 17 x 11.
Run from project root:
python3 scripts/analysis/draw_cta_framework.py
"""
from __future__ import annotations
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import FancyArrowPatch, FancyBboxPatch
OUT_DIR = Path("outputs/analysis/figs_framework")
OUT_DIR.mkdir(parents=True, exist_ok=True)
# ============================================================
# Palette
# ============================================================
C_REAL = "#1E8449"
C_FAKE = "#C0392B"
C_SHARED = "#7F8C8D"
C_DATA_REAL_FILL = "#E8F8F0"
C_DATA_FAKE_FILL = "#FBEAEA"
C_DATA_AUDIO_FILL = "#FCF3CF"
C_DATA_AUDIO_EDGE = "#B7950B"
C_BB_FILL = "#EAF2F8"
C_BB_EDGE = "#5499C7"
C_CORE_FILL = "#FDEBD0"
C_CORE_EDGE = "#D35400"
C_CORE_TEXT = "#7B241C"
C_NEUTRAL_FILL = "#F4F6F7"
C_NEUTRAL_EDGE = "#7F8C8D"
C_OUT_FILL = "#E8DAEF"
C_OUT_EDGE = "#7D3C98"
C_PANEL_LOSS = "#FEF9E7"
C_PANEL_LOSS_EDGE = "#B7950B"
C_PANEL_INNO = "#FDEDEC"
C_PANEL_INNO_EDGE = "#C0392B"
# ============================================================
# Drawing helpers
# ============================================================
def box(ax, cx, cy, w, h, text, fill, edge, *,
fontsize=10, weight="normal", lw=1.5, padding=0.4, star=False,
text_color="black"):
rect = FancyBboxPatch(
(cx - w/2, cy - h/2), w, h,
boxstyle=f"round,pad={padding}",
facecolor=fill, edgecolor=edge, linewidth=lw,
)
ax.add_patch(rect)
label = (("★ " + text) if star else text)
ax.text(cx, cy, label, ha="center", va="center",
fontsize=fontsize, weight=weight, color=text_color)
def arrow(ax, x1, y1, x2, y2, *, color=C_SHARED, lw=1.6, dashed=False,
mutation=14, alpha=1.0):
ls = (0, (5, 3)) if dashed else "-"
a = FancyArrowPatch(
(x1, y1), (x2, y2),
arrowstyle="->", mutation_scale=mutation,
color=color, linewidth=lw, linestyle=ls, alpha=alpha,
shrinkA=2, shrinkB=2,
)
ax.add_patch(a)
def small(ax, cx, cy, text, *, fontsize=8.5, color="#566573", italic=True,
ha="center", weight="normal"):
style = "italic" if italic else "normal"
ax.text(cx, cy, text, ha=ha, va="center",
fontsize=fontsize, style=style, color=color, weight=weight)
# ============================================================
# Canvas
# ============================================================
fig, ax = plt.subplots(figsize=(17, 11))
ax.set_xlim(0, 100)
ax.set_ylim(0, 100)
ax.set_aspect("equal")
ax.axis("off")
# Title
ax.text(50, 96.5, "CTA: Cross-modal Translation Asymmetry — Framework",
ha="center", va="center", fontsize=18, weight="bold")
small(ax, 50, 92.8,
"★ marks the four innovations (described on the right)",
fontsize=10, italic=True, color="#566573")
# Tiny color legend — top-right corner of the data-flow zone
ax.plot([6, 10], [89, 89], color=C_REAL, lw=2.4)
ax.text(10.6, 89, "real", fontsize=9, va="center", color=C_REAL, weight="bold")
ax.plot([16, 20], [89, 89], color=C_FAKE, lw=2.4, linestyle=(0, (5, 3)))
ax.text(20.6, 89, "fake", fontsize=9, va="center", color=C_FAKE, weight="bold")
ax.plot([26, 30], [89, 89], color=C_SHARED, lw=2.4)
ax.text(30.6, 89, "shared", fontsize=9, va="center", color=C_SHARED)
# ========== LEFT ZONE = single chain (x ∈ [4, 66]) ==========
COL_L, COL_R = 22, 50 # video column / audio column on backbone level
COL_C = (COL_L + COL_R) / 2 # center
# ---- Inputs ----
y_in = 84
box(ax, 12, y_in, 16, 5, "Real video\n(label = 0)",
C_DATA_REAL_FILL, C_REAL, fontsize=10)
box(ax, 31, y_in, 16, 5, "Fake video\n(label = 1)",
C_DATA_FAKE_FILL, C_FAKE, fontsize=10)
box(ax, 56, y_in, 22, 5, "Audio ★ always real",
C_DATA_AUDIO_FILL, C_DATA_AUDIO_EDGE, fontsize=10, weight="bold",
text_color="#7E5109")
# ---- Backbones ----
y_bb = 73
box(ax, COL_L, y_bb, 28, 5,
"VideoMAE-base (freeze 70 %)",
C_BB_FILL, C_BB_EDGE, fontsize=10)
box(ax, 56, y_bb, 22, 5,
"Wav2Vec2-base (freeze 80 %)",
C_BB_FILL, C_BB_EDGE, fontsize=10)
# input → backbone arrows
arrow(ax, 12, y_in - 2.6, COL_L - 5, y_bb + 2.6, color=C_REAL, lw=2.0)
arrow(ax, 31, y_in - 2.6, COL_L + 5, y_bb + 2.6, color=C_FAKE, lw=2.0, dashed=True)
arrow(ax, 56, y_in - 2.6, 56, y_bb + 2.6, color=C_SHARED, lw=2.0)
# Token notation (compact, side-of-arrow)
small(ax, 22, y_bb - 4, "v.tokens / v.pooled",
fontsize=8, italic=True, color="#566573")
small(ax, 56, y_bb - 4, "a.tokens / a.pooled",
fontsize=8, italic=True, color="#566573")
# ---- Predictors (CORE / Innovation 1+2) ----
y_pr = 60
box(ax, 22, y_pr, 28, 7,
"$f_{A \\to V}$ predictor\nTransformer Decoder × 4",
C_CORE_FILL, C_CORE_EDGE, fontsize=10, lw=2.5, star=True,
text_color=C_CORE_TEXT, weight="bold")
box(ax, 56, y_pr, 28, 7,
"$f_{V \\to A}$ predictor\nTransformer Decoder × 4",
C_CORE_FILL, C_CORE_EDGE, fontsize=10, lw=2.5, star=True,
text_color=C_CORE_TEXT, weight="bold")
# small "REAL only" badges below each predictor — Innovation 2 in 5 chars
small(ax, 22, y_pr - 4.5,
"trained on real pairs only",
fontsize=8.2, italic=True, color=C_REAL, weight="bold")
small(ax, 56, y_pr - 4.5,
"trained on real pairs only",
fontsize=8.2, italic=True, color=C_REAL, weight="bold")
# arrows backbone → predictors
# v.tokens → A→V (tgt) and V→A (src). a.tokens → A→V (src) and V→A (tgt).
arrow(ax, 22, y_bb - 2.6, 22, y_pr + 4, color=C_SHARED, lw=1.4) # v→A→V
arrow(ax, 22, y_bb - 2.6, 50, y_pr + 4, color=C_SHARED, lw=1.0,
alpha=0.55)
arrow(ax, 56, y_bb - 2.6, 56, y_pr + 4, color=C_SHARED, lw=1.4) # a→V→A
arrow(ax, 56, y_bb - 2.6, 28, y_pr + 4, color=C_SHARED, lw=1.0,
alpha=0.55)
# ---- Per-sample residuals ----
y_res = 47
box(ax, 22, y_res, 26, 4.5,
"$L_{A \\to V}$ = MSE(v_pred, v.tokens)",
C_CORE_FILL, C_CORE_EDGE, fontsize=10, lw=2.0)
box(ax, 56, y_res, 26, 4.5,
"$L_{V \\to A}$ = MSE(a_pred, a.tokens)",
C_CORE_FILL, C_CORE_EDGE, fontsize=10, lw=2.0)
arrow(ax, 22, y_pr - 4, 22, y_res + 2.5, color=C_SHARED, lw=1.4)
arrow(ax, 56, y_pr - 4, 56, y_res + 2.5, color=C_SHARED, lw=1.4)
# ---- Asymmetry score (Innovation 3) ----
y_asym = 38
box(ax, 39, y_asym, 50, 5,
"$s_{\\rm asym}$ = $L_{V\\to A}$ − $L_{A\\to V}$ "
"$L_{\\rm total}$ = $L_{V\\to A}$ + $L_{A\\to V}$",
C_CORE_FILL, C_CORE_EDGE, fontsize=10.5, lw=2.5, star=True,
text_color=C_CORE_TEXT, weight="bold")
arrow(ax, 22, y_res - 2.3, 30, y_asym + 2.5, color=C_SHARED, lw=1.4)
arrow(ax, 56, y_res - 2.3, 48, y_asym + 2.5, color=C_SHARED, lw=1.4)
# ---- detach (Innovation 4) ----
y_det = 30.5
small(ax, 39, y_det, ".detach() ★ cuts gradient back to predictors",
fontsize=10, italic=True, color=C_CORE_TEXT, weight="bold")
arrow(ax, 39, y_asym - 2.5, 39, y_det + 1, color=C_SHARED, lw=1.4)
# ---- Classifier head ----
y_cls = 22.5
box(ax, 39, y_cls, 56, 6,
"Classifier Head MLP (1538 → 256 → 1)\n"
"input = concat[ v.pooled , a.pooled , $s_{\\rm asym}$ , $L_{\\rm total}$ ]",
C_BB_FILL, C_BB_EDGE, fontsize=10)
arrow(ax, 39, y_det - 0.7, 39, y_cls + 3, color=C_SHARED, lw=1.4)
# pooled features feed in from the side (subtle)
arrow(ax, 22, y_bb - 2.6, 12, y_cls + 1, color=C_SHARED, lw=0.7,
alpha=0.45)
arrow(ax, 56, y_bb - 2.6, 66, y_cls + 1, color=C_SHARED, lw=0.7,
alpha=0.45)
# ---- Output ----
y_out = 13
box(ax, 39, y_out, 28, 4.5,
"score = sigmoid(logit) ∈ [0, 1]\nhigh ⇒ fake",
C_OUT_FILL, C_OUT_EDGE, fontsize=10.5, weight="bold")
arrow(ax, 39, y_cls - 3, 39, y_out + 2.3, color=C_SHARED, lw=1.6)
# ========== RIGHT ZONE = info panels (x ∈ [70, 99]) ==========
# ---- Innovation panel (top-right) ----
inn_w, inn_h = 28, 38
inn_cx, inn_cy = 84, 67
inn = FancyBboxPatch(
(inn_cx - inn_w/2, inn_cy - inn_h/2), inn_w, inn_h,
boxstyle="round,pad=0.5",
facecolor=C_PANEL_INNO, edgecolor=C_PANEL_INNO_EDGE, linewidth=2,
)
ax.add_patch(inn)
ax.text(inn_cx, inn_cy + inn_h/2 - 2,
"Key Innovations ★",
ha="center", fontsize=12.5, weight="bold", color=C_CORE_TEXT)
# Each line: "★ N. Title" then 1-line gloss
ix = inn_cx - inn_w/2 + 1.2
iy = inn_cy + inn_h/2 - 5.0
bullets = [
("★ 1. Bidirectional cross-modal predictors",
" A→V and V→A (Transformer Decoders)"),
("★ 2. Trained on REAL pairs ONLY",
" loss_av, loss_va masked by label = 0"),
("★ 3. Asymmetry score $s_\\mathrm{asym}$ = detector",
" real ≈ 0 ; fake ≪ 0 (OOD signature)"),
("★ 4. detach() before classifier head",
" keeps predictors from leaking to BCE"),
("★ 5. Cross-generator consistency loss",
" same num, alt generator → same $s_\\mathrm{asym}$"),
]
for title, gloss in bullets:
ax.text(ix, iy, title, fontsize=9.4, weight="bold", color=C_CORE_TEXT)
iy -= 1.7
ax.text(ix, iy, gloss, fontsize=8.2, color="#7B241C", style="italic")
iy -= 3.6
# ---- Loss panel (bottom-right) ----
lp_w, lp_h = 28, 36
lp_cx, lp_cy = 84, 25
lp = FancyBboxPatch(
(lp_cx - lp_w/2, lp_cy - lp_h/2), lp_w, lp_h,
boxstyle="round,pad=0.5",
facecolor=C_PANEL_LOSS, edgecolor=C_PANEL_LOSS_EDGE, linewidth=2,
)
ax.add_patch(lp)
ax.text(lp_cx, lp_cy + lp_h/2 - 2.0,
"Training losses & sample masks",
ha="center", fontsize=11.5, weight="bold")
lx = lp_cx - lp_w/2 + 1.2
ly = lp_cy + lp_h/2 - 5.0
losses = [
("$\\mathrm{loss}_{av}$ = mean($L_{A\\to V}$ | label=0)",
"REAL only", C_REAL),
("$\\mathrm{loss}_{va}$ = mean($L_{V\\to A}$ | label=0)",
"REAL only", C_REAL),
("$\\mathrm{loss}_{asym}$ = ReLU($s_\\mathrm{fake}$ − $s_\\mathrm{real}$)",
"BOTH (margin)", "#566573"),
("$\\mathrm{loss}_{cls}$ = BCE(score, label)",
"BOTH", "#566573"),
("$\\mathrm{loss}_{aux}$ = MSE($s_\\mathrm{asym}$, $s_\\mathrm{asym, alt}$)",
"paired FAKEs", C_FAKE),
]
for line, mask, color in losses:
ax.text(lx, ly, line, fontsize=9.0)
ly -= 1.5
ax.text(lx, ly, " sample mask: " + mask,
fontsize=8, color=color, style="italic")
ly -= 2.7
ax.text(lp_cx, lp_cy - lp_h/2 + 3.5,
"Total = 1·$loss_{av}$ + 1·$loss_{va}$ + 0.5·$loss_{asym}$\n"
" + 1·$loss_{cls}$ + 0.1·$loss_{aux}$",
ha="center", va="center", fontsize=8.8, weight="bold")
# ============================================================
# Save
# ============================================================
out_png = OUT_DIR / "cta_framework.png"
out_pdf = OUT_DIR / "cta_framework.pdf"
plt.tight_layout(pad=0.6)
plt.savefig(out_png, dpi=240, bbox_inches="tight", facecolor="white")
plt.savefig(out_pdf, bbox_inches="tight", facecolor="white")
plt.close(fig)
print(f"[draw] wrote {out_png}")
print(f"[draw] wrote {out_pdf}")