File size: 5,209 Bytes
3c41fb5 | 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 | #!/usr/bin/env python3
"""t-SNE of four cable sizes, testing whether 3.00 sq lands between blue and 3.30.
The claim under test is an ordering one, so it is checked numerically rather than
by eye: a diameter/size axis is regressed on held-out episodes and the class means
along it are compared. t-SNE distances are not metric, so the map alone cannot
support a "lies between" statement -- the axis can.
Confound worth stating: 3.00 sq comes from a different dataset than the other
three. Anything session-specific (lighting, gel state, day) travels with that
label. The ordering test is the guard: a session artefact has no reason to place
3.00 sq *between* two other sizes rather than off to one side.
"""
import argparse
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from sklearn.decomposition import PCA
from sklearn.linear_model import RidgeCV
from sklearn.manifold import TSNE
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
# Conductor cross-section [sq mm]; used only to order the classes.
SQ = {"1.70 sq": 1.70, "blue (2.6 mm)": 2.30, "3.00 sq": 3.00, "3.30 sq": 3.30}
COLOR = {"1.70 sq": "#C05621", "blue (2.6 mm)": "#2B6CB0",
"3.00 sq": "#805AD5", "3.30 sq": "#2F855A"}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--cache", default="outputs/pca/cache/cache_4sizes.npz")
ap.add_argument("--outdir", default="outputs/pca/4sizes")
ap.add_argument("--perplexity", type=float, default=30.0)
ap.add_argument("--blue-sq", type=float, default=2.30,
help="cross-section assigned to the blue cable for the ordering test")
args = ap.parse_args()
SQ["blue (2.6 mm)"] = args.blue_sq
import os
os.makedirs(args.outdir, exist_ok=True)
d = np.load(args.cache, allow_pickle=True)
X, lab, ep = d["feats"], d["label"], d["eps"]
order = sorted(set(lab), key=lambda s: SQ[s])
rng = np.random.default_rng(0)
tr = np.zeros(len(lab), bool)
for t in np.unique(lab):
e = np.unique(ep[lab == t])
tr |= np.isin(ep, rng.choice(e, size=int(len(e) * 0.7), replace=False))
te = ~tr
Xc = X - X.mean(0)
Z = PCA(n_components=30).fit_transform(Xc)
emb = TSNE(n_components=2, perplexity=args.perplexity, init="pca",
random_state=0).fit_transform(Z)
plt.rcParams.update({"font.family": ["DejaVu Sans"], "font.size": 12})
# --- t-SNE on its own
fig, ax = plt.subplots(figsize=(7.6, 6.8))
for t in order:
m = te & (lab == t)
ax.scatter(emb[m, 0], emb[m, 1], s=16, alpha=0.65, edgecolors="none",
c=COLOR[t], label=t)
ax.set_xlabel("t-SNE 1"); ax.set_ylabel("t-SNE 2")
ax.set_title(f"t-SNE of four cable sizes (perplexity {args.perplexity:g})", fontsize=13)
ax.legend(frameon=False, title="cable")
ax.spines[["top", "right"]].set_visible(False)
fig.tight_layout(); fig.savefig(f"{args.outdir}/tsne_4sizes.png", dpi=160, bbox_inches="tight")
# --- ordering test on a regressed size axis
sq = np.array([SQ[t] for t in lab])
reg = make_pipeline(StandardScaler(), RidgeCV(alphas=np.logspace(-2, 4, 25))).fit(X[tr], sq[tr])
pred = reg.predict(X)
r2 = reg.score(X[te], sq[te])
fig, ax = plt.subplots(figsize=(8.4, 6.0))
for t in order:
m = te & (lab == t)
ax.scatter(pred[m], Z[m, 1], s=16, alpha=0.6, edgecolors="none", c=COLOR[t], label=t)
ax.axvline(pred[m].mean(), color=COLOR[t], lw=1.3, ls="--", alpha=0.85)
ax.set_xlabel("predicted conductor cross-section [sq mm]")
ax.set_ylabel("PC2")
ax.set_title(f"Size axis — held-out $R^2$ = {r2:.2f}", fontsize=13)
ax.legend(frameon=False, title="cable")
ax.spines[["top", "right"]].set_visible(False)
fig.tight_layout(); fig.savefig(f"{args.outdir}/size_axis_4sizes.png", dpi=160, bbox_inches="tight")
# --- centroid distances in the raw feature space: is 3.00 between blue and 3.30?
cent = {t: X[te & (lab == t)].mean(0) for t in order}
def dist(a, b): return float(np.linalg.norm(cent[a] - cent[b]))
print(f"wrote {args.outdir}/tsne_4sizes.png")
print(f"wrote {args.outdir}/size_axis_4sizes.png\n")
print(f" size axis held-out R^2 = {r2:.2f}")
print(" mean position on the size axis (held-out frames):")
means = {}
for t in order:
m = te & (lab == t)
means[t] = pred[m].mean()
print(f" {t:14s} true {SQ[t]:.2f} -> {pred[m].mean():.2f} +- {pred[m].std():.2f}")
b, t300, t330 = means["blue (2.6 mm)"], means["3.00 sq"], means["3.30 sq"]
print(f"\n 3.00 sq between blue and 3.30 sq on this axis: "
f"{'YES' if b < t300 < t330 or t330 < t300 < b else 'NO'}")
print("\n centroid distances in the 512-d feature space:")
print(f" blue <-> 3.00 {dist('blue (2.6 mm)', '3.00 sq'):.2f}")
print(f" 3.00 <-> 3.30 {dist('3.00 sq', '3.30 sq'):.2f}")
print(f" blue <-> 3.30 {dist('blue (2.6 mm)', '3.30 sq'):.2f} "
f"(should be the largest if 3.00 sits between)")
print(f" 1.70 <-> blue {dist('1.70 sq', 'blue (2.6 mm)'):.2f}")
if __name__ == "__main__":
main()
|