| |
| """Cable clusters before vs after contact, in the ACT encoder space. |
| |
| Contact onset is detected per episode from finger torque, so "pre" really means |
| the GelSight has not touched the cable yet. That split is what makes the tactile |
| question answerable: anything decodable pre-contact must be visual. |
| |
| Panels: t-SNE of each phase separately (clusters by diameter), and the diameter |
| regression axis fitted within each phase. If pre-contact already separates as |
| well as post-contact, touch adds nothing to cable identity. |
| """ |
| 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 LogisticRegression, RidgeCV |
| from sklearn.manifold import TSNE |
| from sklearn.pipeline import make_pipeline |
| from sklearn.preprocessing import StandardScaler |
|
|
| DIA = {"cable 1.70 sq": 1.7, "blue cable": 2.6, "cable 3.30 sq": 4.5} |
| NAME = {1.7: "1.7 mm", 2.6: "2.6 mm", 4.5: "4.5 mm"} |
| COLOR = {1.7: "#C05621", 2.6: "#2B6CB0", 4.5: "#2F855A"} |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--cache", default="outputs/pca/cache_cable_contact.npz") |
| ap.add_argument("--out", default="outputs/pca/cable_contact_clusters.png") |
| ap.add_argument("--perplexity", type=float, default=30.0) |
| args = ap.parse_args() |
|
|
| d = np.load(args.cache, allow_pickle=True) |
| X, y, ep, ph = d["feats"], d["y"], d["eps"], d["phase"] |
| dia = np.array([DIA[t] for t in y]) |
|
|
| rng = np.random.default_rng(0) |
| tr = np.zeros(len(y), bool) |
| for t in np.unique(y): |
| e = np.unique(ep[y == t]) |
| tr |= np.isin(ep, rng.choice(e, size=int(len(e) * 0.7), replace=False)) |
| te = ~tr |
| order = sorted(set(y), key=lambda s: DIA[s]) |
|
|
| plt.rcParams.update({"font.family": ["DejaVu Sans"], "font.size": 11}) |
| fig, axes = plt.subplots(2, 2, figsize=(13, 11)) |
|
|
| for col, (pname, label) in enumerate([("pre", "pre-contact (GelSight untouched)"), |
| ("post", "post-contact (cable in hand)")]): |
| w = ph == pname |
| Xp = X[w] |
| Xc = Xp - Xp.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) |
| trp, tep, yp = tr[w], te[w], y[w] |
|
|
| clf = make_pipeline(StandardScaler(), LogisticRegression(max_iter=3000)).fit(Xp[trp], yp[trp]) |
| acc = clf.score(Xp[tep], yp[tep]) |
|
|
| ax = axes[0, col] |
| for t in order: |
| m = tep & (yp == t) |
| ax.scatter(emb[m, 0], emb[m, 1], s=16, alpha=0.65, edgecolors="none", |
| c=COLOR[DIA[t]], label=NAME[DIA[t]]) |
| ax.set_xlabel("t-SNE 1"); ax.set_ylabel("t-SNE 2") |
| ax.set_title(f"{label}\nt-SNE — {acc:.1%} accuracy (chance 33.3%)", fontsize=12) |
| if col == 0: |
| ax.legend(frameon=False, title="cable diameter") |
|
|
| reg = make_pipeline(StandardScaler(), RidgeCV(alphas=np.logspace(-2, 4, 25))).fit( |
| Xp[trp], dia[w][trp]) |
| pred = reg.predict(Xp) |
| r2 = reg.score(Xp[tep], dia[w][tep]) |
| ax = axes[1, col] |
| for t in order: |
| m = tep & (yp == t) |
| ax.scatter(pred[m], Z[m, 1], s=16, alpha=0.65, edgecolors="none", |
| c=COLOR[DIA[t]], label=NAME[DIA[t]]) |
| for v in (1.7, 2.6, 4.5): |
| ax.axvline(v, color=COLOR[v], lw=1.1, ls="--", alpha=0.7) |
| ax.set_xticks([1.7, 2.6, 4.5]) |
| ax.set_xlabel("predicted cable diameter [mm]"); ax.set_ylabel("PC2") |
| ax.set_title(f"diameter axis — held-out $R^2$ = {r2:.2f}", fontsize=12) |
|
|
| print(f" {pname:4s} accuracy {acc:.1%} diameter R^2 {r2:.2f}", flush=True) |
| for t in order: |
| m = tep & (yp == t) |
| print(f" {NAME[DIA[t]]:>7} -> {pred[m].mean():.2f} +- {pred[m].std():.2f} mm", flush=True) |
|
|
| for ax in axes.ravel(): |
| ax.spines[["top", "right"]].set_visible(False) |
| fig.tight_layout(); fig.savefig(args.out, dpi=150, bbox_inches="tight") |
| print(f"\nwrote {args.out}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
