File size: 17,370 Bytes
5f0e48f | 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 | """Paper figures from saved run outputs; each figure is skipped while its inputs are missing.
python make_figs.py stats --device cpu # style statistics of the test images through S2 (t-SNE input), once
python make_figs.py figs # writes ../figs/*.pdf from whatever exists in ../runs
python make_figs.py export --device cuda:0 # runs/fig_examples.pt for paper/figs.py (journal-style figures)
"""
import argparse
import csv
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt # noqa: E402
import torch # noqa: E402
from gridkd import * # noqa: F401,F403,E402
R, F = Path("../runs"), Path("../figs")
BENCH = {"A": ("../datasets/A_weather", ("clear", "fog")), "B": ("../datasets/B_region", ("cn", "br"))}
plt.rcParams.update({"font.size": 8, "axes.titlesize": 8, "figure.dpi": 200, "savefig.bbox": "tight"})
def auc(neg, pos): # P(pos > neg), ties count half
d = pos[None, :] - neg[:, None]
return float((d > 0).float().mean() + 0.5 * (d == 0).float().mean())
def card_names(b, k=8):
return [n for n, x in load_plain(R / f"{b}_Lt/weights/best.pt").model.named_modules() if isinstance(x, MultiLoRAConv)][:k]
def cmd_stats(a):
from deploy_route import stats
for b, (d, doms) in BENCH.items():
out = R / f"{b}_style_test.pt"
if out.exists():
continue
names = card_names(b)
torch.save({dom: torch.cat(stats(load_plain(R / f"{b}_S2/weights/best.pt"), names, f"{d}/{dom}.yaml", "test", a.device), 1)
for dom in doms}, out)
print("wrote", out)
def fig_layer_auc():
"""Per-layer and first-k domain separability: BN statistics (all 81 layers) and density cards (first 16 LoRA inputs)."""
have = [b for b in BENCH if (R / f"{b}_bnllr_on.pt").exists()]
if not have:
return
fig, axs = plt.subplots(1, len(have), figsize=(3.4 * len(have), 2.4), squeeze=False)
for ax, b in zip(axs[0], have):
on, off = torch.load(R / f"{b}_bnllr_on.pt"), torch.load(R / f"{b}_bnllr_off.pt")
(s1, t1, _), (s0, t0, _) = on.values(), off.values()
S, T = s1[..., 0] - s0[..., 1], t1[..., 0] - t0[..., 1] # each model on its own path: target-BN vs source-BN fit
L = range(1, S.shape[1] + 1)
ax.plot(L, [auc(S[:, i], T[:, i]) for i in range(S.shape[1])], lw=0.8, c="C0", alpha=0.6, label="BN, single layer")
ax.plot(L, [auc(S[:, :k].sum(1), T[:, :k].sum(1)) for k in L], lw=1.2, c="C0", label="BN, first k layers")
lab = R / f"{b}_lab_s0.pt"
if lab.exists():
p = torch.load(lab)["per_layer"]
P0, P1 = p["test0"], p["test1"]
K = range(1, P0.shape[1] + 1)
ax.plot(K, [auc(P0[:, i], P1[:, i]) for i in range(P0.shape[1])], lw=0.8, c="C3", alpha=0.6, label="card, single layer")
ax.plot(K, [auc(P0[:, :k].sum(1), P1[:, :k].sum(1)) for k in K], lw=1.2, c="C3", label="card, first k layers")
for rule, se, c, lab in (("deep", "base/4", "C2", "deep part (layer-10 features)"), ("fused", "base", "k", "fused card")):
v = [float(x["auc"]) for x in lab_rows(b) if x["rule"] == rule and x["setting"] == se and x["q"] == "raw"]
if v:
ax.axhline(v[0], c=c, lw=0.8, ls="--", label=lab)
doms = BENCH[b][1]
ax.set(title=f"{b}: {doms[0]} vs {doms[1]} (test)", xlabel="layer (network order)", ylabel="AUC", xscale="log", ylim=(0, 1.02))
ax.axhline(0.5, c="k", lw=0.5, ls=":")
ax.grid(alpha=0.3)
axs[0][0].legend(loc="lower right", fontsize=6)
fig.savefig(F / "layer_auc.pdf")
print("wrote layer_auc.pdf")
def fig_tsne():
from sklearn.manifold import TSNE
have = [b for b in BENCH if (R / f"{b}_style_test.pt").exists()]
if not have:
return
fig, axs = plt.subplots(1, len(have), figsize=(3.0 * len(have), 2.8), squeeze=False)
g = torch.Generator().manual_seed(0)
for ax, b in zip(axs[0], have):
st = torch.load(R / f"{b}_style_test.pt")
xs = [v[torch.randperm(len(v), generator=g)[:800]] for v in st.values()] # ponytail: 800 per domain keeps t-SNE fast
x = torch.cat(xs)
x = ((x - x.mean(0)) / (x.std(0) + 1e-6)).numpy()
e = TSNE(2, init="pca", perplexity=30, random_state=0).fit_transform(x)
i = 0
for c, (dom, v) in zip(("C0", "C3"), zip(st, xs)):
ax.scatter(*e[i:i + len(v)].T, s=2, c=c, alpha=0.5, lw=0, label=dom)
i += len(v)
ax.set(title=f"{b}: style statistics, first 8 LoRA inputs (S2)", xticks=[], yticks=[])
ax.legend(markerscale=4, fontsize=6)
fig.savefig(F / "tsne_style.pdf")
print("wrote tsne_style.pdf")
def lab_rows(b, seed="s0"):
p = R / f"{b}_lab_{seed}.csv"
return list(csv.DictReader(open(p))) if p.exists() else []
def fig_llr_hist():
"""Image-level log-likelihood ratio on both test sets: early-layer card (own path, 5% threshold) and fused card
(base path, threshold 0)."""
have = [b for b in BENCH if (R / f"{b}_lab_s0.pt").exists() and "fused/base" in torch.load(R / f"{b}_lab_s0.pt")["scores"]]
if not have:
return
fig, axs = plt.subplots(2, len(have), figsize=(3.2 * len(have), 4.2), squeeze=False)
for j, b in enumerate(have):
sc = torch.load(R / f"{b}_lab_s0.pt")["scores"]
for i, (key, title, q) in enumerate((("card/own/8/4/ms", "early-layer card", 0.05), ("fused/base", "fused card", None))):
ax, s = axs[i][j], sc[key]
r = {k: v[:, 1] - v[:, 0] for k, v in s.items()}
tau = float(r["val1"].quantile(q)) if q else 0.0
sl = lambda t: torch.sign(t) * torch.log10(1 + t.abs()) # noqa: E731 signed log: the ratios span 1e0-1e4
r, tau = {k: sl(v) for k, v in r.items()}, float(sl(torch.tensor(tau)))
for c, k, dom in zip(("C0", "C3"), ("test0", "test1"), BENCH[b][1]):
ax.hist(r[k].numpy(), bins=80, color=c, alpha=0.55, density=True, label=dom)
ax.axvline(tau, c="k", lw=0.8, label=r"$\tau$" + (" (5% of new-site val)" if q else " = 0"))
ax.set(title=f"{b}: {title} (test)", yticks=[], xlabel="signed log10(1 + |LLR|), LLR = log p(new) - log p(old)" if i else None)
ax.legend(fontsize=6)
fig.tight_layout()
fig.savefig(F / "llr_hist.pdf")
print("wrote llr_hist.pdf")
def fig_depth():
"""Fused card with the deep part read at layer 4 / 6 / 8 / 10 of the base model: AUC against the share of the base
model's MACs spent before the decision."""
have = {b: r for b in BENCH for st_ in ("s2", "s1", "s0") if (r := [x for x in lab_rows(b, st_) if x["rule"] in ("deep", "fused") and x["q"] == "raw"])
and any("/L" in x["setting"] for x in r)}
if not have:
return
from deploy_route import macs_by_layer
macs = macs_by_layer(load_plain(R / "A_S2/weights/best.pt").model.float(), 640)
share = {L: sum(v for i, v in macs.items() if i <= L) / sum(macs.values()) for L in (4, 6, 8, 10)}
fig, ax = plt.subplots(figsize=(3.4, 2.4))
for b, rows in have.items():
get = lambda rule, se: next(float(x["auc"]) for x in rows if x["rule"] == rule and x["setting"] == se) # noqa: E731
for rule, ls, lab in (("fused", "-", "fused card"), ("deep", ":", "deep part only")):
ys = [get(rule, ("base" if rule == "fused" else "base/4") + ("" if L == 10 else f"/L{L}")) for L in (4, 6, 8, 10)]
ax.plot([share[L] for L in (4, 6, 8, 10)], ys, ls=ls, marker="o", ms=3, c="C0" if b == "A" else "C3", label=f"{b}: {lab}")
for L in (4, 6, 8, 10):
ax.annotate(f"L{L}", (share[L], 0.5), fontsize=6, ha="center")
ax.set(xlabel="share of base-model MACs before the decision", ylabel="AUC (test)", ylim=(0.45, 1.01))
ax.grid(alpha=0.3)
ax.legend(fontsize=6)
fig.savefig(F / "depth.pdf")
print("wrote depth.pdf")
def fig_card_grid():
"""Early part of the fused card (base path, threshold 0): mean test mAP50-95 over both sites against layers read (k)
and GMM components (K)."""
have = [b for b in BENCH if lab_rows(b)]
if not have:
return
fig, axs = plt.subplots(1, 2 * len(have), figsize=(2.3 * 2 * len(have), 2.0), squeeze=False)
for j, b in enumerate(have):
rows = [r for r in lab_rows(b) if r["rule"] == "card" and r["q"] == "raw"]
for ax, feat in zip(axs[0][2 * j:2 * j + 2], ("ms", "m")):
ks = sorted({int(r["setting"].split("/")[1]) for r in rows})
Ks = sorted({int(r["setting"].split("/")[2]) for r in rows})
M = torch.full((len(Ks), len(ks)), float("nan"))
for r in rows:
path, k, K, f = r["setting"].split("/")
if path == "base" and f == feat:
M[Ks.index(int(K)), ks.index(int(k))] = (float(r["map_0"]) + float(r["map_1"])) / 2
im = ax.imshow(M.numpy(), cmap="viridis", origin="lower")
for (yy, xx), v in zip(((yy, xx) for yy in range(len(Ks)) for xx in range(len(ks))), M.flatten()):
ax.text(xx, yy, f"{v:.3f}", ha="center", va="center", fontsize=5, c="w")
ax.set(xticks=range(len(ks)), xticklabels=ks, yticks=range(len(Ks)), yticklabels=Ks, xlabel="k (layers)",
ylabel="K (components)" if feat == "ms" else None, title=f"{b}: {'mean + log-std' if feat == 'ms' else 'mean only'}")
fig.tight_layout()
fig.savefig(F / "card_grid.pdf")
print("wrote card_grid.pdf")
def gt_boxes(img, w, h):
p = Path(str(img).replace("/images/", "/labels/")).with_suffix(".txt")
t = torch.tensor([[float(v) for v in ln.split()[1:5]] for ln in p.read_text().splitlines() if ln.strip()]).view(-1, 4)
return torch.stack([(t[:, 0] - t[:, 2] / 2) * w, (t[:, 1] - t[:, 3] / 2) * h, (t[:, 0] + t[:, 2] / 2) * w, (t[:, 1] + t[:, 3] / 2) * h], 1)
def fig_examples():
"""One random test image per site (seed 0, >= 2 insulators): ground truth, old-site model, new-site model; the
panel of the model the fused card picks (threshold 0) is framed."""
from PIL import Image
have = [b for b in BENCH if (R / f"{b}_lab_s0.pt").exists()]
if not have:
return
rows = [(b, i) for b in have for i in range(2)]
fig, axs = plt.subplots(len(rows), 3, figsize=(6.6, 1.9 * len(rows)), squeeze=False)
g = torch.Generator().manual_seed(0)
for b in have:
lab = torch.load(R / f"{b}_lab_s0.pt")
s, tau = lab["scores"]["fused/base"], 0.0
ms = [load_plain(R / f"{b}_{w}/weights/best.pt") for w in ("S2", "Lt")]
for i in range(2):
files = lab["files"][i]
for j in torch.randperm(len(files), generator=g).tolist():
im = Image.open(files[j]).convert("RGB")
gt = gt_boxes(files[j], *im.size)
if len(gt) >= 2:
break
llr = float(s[f"test{i}"][j, 1] - s[f"test{i}"][j, 0])
pick = int(llr > tau)
ax = axs[rows.index((b, i))]
pr = [m.predict(files[j], conf=0.25, device="cpu", verbose=False)[0].boxes for m in ms]
for c, (bx, title) in enumerate(zip([gt] + [p.xyxy for p in pr],
["ground truth", "old-site model (S2)", "new-site model (LoRA)"])):
ax[c].imshow(im)
for x0, y0, x1, y1 in bx.tolist():
ax[c].add_patch(plt.Rectangle((x0, y0), x1 - x0, y1 - y0, fill=False, lw=0.8, ec="lime" if c == 0 else "yellow"))
ax[c].set(xticks=[], yticks=[], title=title if c == 0 else f"{title}: {len(bx)} box" + ("" if len(bx) == 1 else "es"))
if c == pick + 1: # the router's choice
for sp in ax[c].spines.values():
sp.set(ec="red", lw=2.5)
ax[0].set_ylabel(f"{b} {BENCH[b][1][i]}\nfused-card LLR {llr:.0f}", fontsize=7)
fig.savefig(F / "examples.pdf")
print("wrote examples.pdf")
SHOW = {"T": "teacher (YOLO11l)", "S0": "student (YOLO11n)", "S2": "KD + projection (old site)", "FT": "full fine-tune", "J": "joint training (both sites)",
"S3nt": "multi-LoRA + router", "Lt": "LoRA, source-free", "Lt_deploy_q5": "LoRA + density cards (deployed)",
"S2p2": "pruned 30%", "S3p2": "pruned + multi-LoRA", "S3ntp": "pruned + multi-LoRA + router", "Ltp_deploy_q5": "pruned + LoRA + cards (deployed)",
"Lt_deploy_fused": "LoRA + fused cards (deployed)", "Ltp_deploy_fused": "pruned + LoRA + fused cards (deployed)"}
def fig_tradeoff():
"""Mean test mAP50-95 over both sites against GFLOPs per image (mean over both test sets)."""
pts = {b: {} for b in BENCH}
for p in sorted(R.glob("*_eval*.csv")):
if any(t in p.stem for t in ("_v1", "_nanbug", "_seed", "_cards_s")):
continue
for r in csv.DictReader(open(p)):
b, m = r["model"].split("_", 1)
if b in pts and m in SHOW:
pts[b].setdefault(m, [{}, {}]) # per domain: GFLOPs (routed models vary with the routed share), mAP; later files win
pts[b][m][0][r["domain"]], pts[b][m][1][r["domain"]] = float(r["gflops"]), float(r["map50_95"])
have = [b for b in BENCH if pts[b]]
fig, axs = plt.subplots(1, len(have), figsize=(3.0 * len(have), 2.6), squeeze=False)
style = {m: (f"C{i % 10}", "osD^v<>p"[i % 8]) for i, m in enumerate(SHOW)}
for ax, b in zip(axs[0], have):
for m in SHOW:
if m in pts[b] and len(pts[b][m][1]) == 2:
gf, v = sum(pts[b][m][0].values()) / 2, list(pts[b][m][1].values())
ax.scatter(gf, sum(v) / 2, s=16, c=style[m][0], marker=style[m][1], label=SHOW[m], ec="k", lw=0.3)
ax.set(xscale="log", xlabel="GFLOPs per image", title=b)
ax.grid(alpha=0.3)
axs[0][0].set_ylabel("mean test mAP50-95 (both sites)")
fig.subplots_adjust(wspace=0.25)
h = {l: x for ax in axs[0] for x, l in zip(*ax.get_legend_handles_labels())}
fig.legend(h.values(), h.keys(), loc="upper center", bbox_to_anchor=(0.5, -0.02), ncol=3, fontsize=6, frameon=False)
fig.savefig(F / "tradeoff.pdf")
print("wrote tradeoff.pdf")
def fig_confusion():
"""Three-site sequence (cn -> cnfog -> br): where each site's test images are routed, per rule (threshold 0)."""
p = R / "B_lab3.pt"
if not p.exists():
return
sc, names = torch.load(p)["scores"], ("cn", "cnfog", "br")
rules = [(k, t) for k, t in (("fused/base", "fused card"), ("card/base/8/4/ms", "early-layer style only"), ("deep/base/4", "deep part only"),
("bn-miyazaki", "BN statistics (Miyazaki)"), ("nm-mean", "nearest mean")) if k in sc]
fig, axs = plt.subplots(1, len(rules), figsize=(2.0 * len(rules), 2.2), squeeze=False)
for ax, (k, title) in zip(axs[0], rules):
M = torch.stack([torch.bincount(sc[k][f"test{i}"].argmax(1), minlength=3).float() / len(sc[k][f"test{i}"]) for i in range(3)])
ax.imshow(M.numpy(), cmap="Blues", vmin=0, vmax=1)
for i in range(3):
for j in range(3):
ax.text(j, i, f"{M[i, j]:.2f}", ha="center", va="center", fontsize=6, c="w" if M[i, j] > 0.6 else "k")
ax.set(title=title, xticks=range(3), xticklabels=names, yticks=range(3), yticklabels=names if ax is axs[0][0] else [], xlabel="routed to")
if ax is axs[0][0]:
ax.set_ylabel("test images of")
fig.tight_layout()
fig.savefig(F / "confusion3.pdf")
print("wrote confusion3.pdf")
def cmd_export(a):
"""The example images of fig_examples (same seed and order), downscaled, with ground-truth and predicted boxes, so
paper/figs.py can draw them off the server: runs/fig_examples.pt."""
import numpy as np
from PIL import Image
out, g = [], torch.Generator().manual_seed(0)
for b in BENCH:
lab = torch.load(R / f"{b}_lab_s0.pt")
s = lab["scores"]["fused/base"]
ms = [load_plain(R / f"{b}_{w}/weights/best.pt") for w in ("S2", "Lt")]
for i in range(2):
files = lab["files"][i]
for j in torch.randperm(len(files), generator=g).tolist():
im = Image.open(files[j]).convert("RGB")
gt = gt_boxes(files[j], *im.size)
if len(gt) >= 2:
break
pr = [m.predict(files[j], conf=0.25, device=a.device, verbose=False)[0].boxes.xyxy.cpu() for m in ms]
sc = 640 / max(im.size)
small = im.resize((round(im.width * sc), round(im.height * sc)))
out.append({"b": b, "i": i, "file": str(files[j]), "img": torch.from_numpy(np.asarray(small).copy()),
"boxes": [x * sc for x in [gt] + pr], "llr": float(s[f"test{i}"][j, 1] - s[f"test{i}"][j, 0])})
torch.save(out, R / "fig_examples.pt")
print("wrote", R / "fig_examples.pt")
def cmd_figs(a):
F.mkdir(exist_ok=True)
for f in (fig_layer_auc, fig_tsne, fig_llr_hist, fig_depth, fig_card_grid, fig_examples, fig_tradeoff, fig_confusion):
f()
if __name__ == "__main__":
p = argparse.ArgumentParser()
p.add_argument("cmd", choices=("stats", "figs", "export"))
p.add_argument("--device", default="cpu")
a = p.parse_args()
{"stats": cmd_stats, "figs": cmd_figs, "export": cmd_export}[a.cmd](a)
|