File size: 5,075 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 | #!/usr/bin/env python3
"""When during an episode does the representation know which cable it is?
Splits each episode by elapsed time and asks, separately per window, how well a
probe recovers cable identity from the encoder features. The question this
answers: during the first seconds the gripper is still reaching and has not
touched the cable, so anything decodable there is visual. If accuracy is already
high while reaching, touch is not what carries the distinction.
Held-out episodes throughout; probes are fit per window on training episodes.
"""
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
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 probe(X, y, tr, te):
if tr.sum() < 30 or te.sum() < 15 or len(set(y[tr])) < 2:
return np.nan
c = make_pipeline(StandardScaler(), LogisticRegression(max_iter=3000))
c.fit(X[tr], y[tr])
return c.score(X[te], y[te])
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--cache", default="outputs/pca/cache_cable_phase.npz")
ap.add_argument("--reach-sec", type=float, default=5.0,
help="frames before this are treated as the reaching phase")
ap.add_argument("--out", default="outputs/pca/cable_phase_analysis.png")
args = ap.parse_args()
d = np.load(args.cache, allow_pickle=True)
X, y, ep, tsec, tpos = d["feats"], d["y"], d["eps"], d["tsec"], d["tpos"]
rng = np.random.default_rng(0)
tr_ep = np.zeros(len(y), bool)
for t in np.unique(y):
e = np.unique(ep[y == t])
tr_ep |= np.isin(ep, rng.choice(e, size=int(len(e) * 0.7), replace=False))
te_ep = ~tr_ep
plt.rcParams.update({"font.family": ["DejaVu Sans"], "font.size": 11})
fig, axes = plt.subplots(1, 3, figsize=(17, 5.2))
# --- 1) accuracy vs elapsed time, in sliding windows
edges = np.arange(0, 31, 2.5)
mids, accs, ns = [], [], []
for a, b in zip(edges[:-1], edges[1:]):
w = (tsec >= a) & (tsec < b)
accs.append(probe(X, y, tr_ep & w, te_ep & w))
mids.append((a + b) / 2); ns.append((te_ep & w).sum())
axes[0].plot(mids, accs, "o-", color="#2B6CB0", lw=2, ms=6)
axes[0].axhline(1 / 3, color="#888", ls="--", lw=1)
axes[0].axvspan(0, args.reach_sec, color="#C05621", alpha=0.10)
axes[0].annotate("reaching\n(no contact)", (args.reach_sec / 2, 0.36), ha="center",
fontsize=9.5, color="#C05621")
axes[0].set_xlabel("time since episode start [s]")
axes[0].set_ylabel("held-out accuracy")
axes[0].set_ylim(0.2, 1.02)
axes[0].set_title("Cable identity over the episode", fontsize=12)
# --- 2) reaching vs sorting, probe trained and tested within each phase
reach = tsec < args.reach_sec
sort = ~reach
bars, labels = [], []
for name, w in (("reaching\n(<%.0fs)" % args.reach_sec, reach),
("sorting\n(>=%.0fs)" % args.reach_sec, sort)):
bars.append(probe(X, y, tr_ep & w, te_ep & w)); labels.append(name)
# cross-phase: train while reaching, test while sorting and vice versa
bars.append(probe(X, y, tr_ep & reach, te_ep & sort)); labels.append("train reach\ntest sort")
bars.append(probe(X, y, tr_ep & sort, te_ep & reach)); labels.append("train sort\ntest reach")
axes[1].bar(labels, bars, color=["#C05621", "#2F855A", "#888", "#888"])
axes[1].axhline(1 / 3, color="#888", ls="--", lw=1)
for i, v in enumerate(bars):
axes[1].annotate(f"{v:.1%}", (i, v), ha="center", va="bottom", fontsize=10)
axes[1].set_ylim(0, 1.08); axes[1].set_ylabel("held-out accuracy")
axes[1].set_title("Reaching vs sorting", fontsize=12)
# --- 3) PCA coloured by phase, to see what actually dominates the variance
Xc = X - X.mean(0)
p = PCA(n_components=4).fit(Xc)
Z = p.transform(Xc)
sc = axes[2].scatter(Z[te_ep, 0], Z[te_ep, 1], c=tsec[te_ep], s=12, alpha=0.7,
cmap="viridis", edgecolors="none")
plt.colorbar(sc, ax=axes[2], label="time since start [s]")
axes[2].set_xlabel(f"PC1 ({p.explained_variance_ratio_[0]:.1%})")
axes[2].set_ylabel(f"PC2 ({p.explained_variance_ratio_[1]:.1%})")
axes[2].set_title("PC1/PC2 coloured by phase,\nnot by cable", fontsize=12)
for ax in axes:
ax.spines[["top", "right"]].set_visible(False)
fig.tight_layout(); fig.savefig(args.out, dpi=150, bbox_inches="tight")
print(f"wrote {args.out}\n")
print(" accuracy by time window (held-out episodes):")
for m, a, n in zip(mids, accs, ns):
print(f" {m:5.1f}s {a:6.1%} (n={n})")
print()
for l, v in zip(labels, bars):
print(f" {l.replace(chr(10), ' '):24s} {v:6.1%}")
if __name__ == "__main__":
main()
|