File size: 1,422 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 | #!/usr/bin/env python3
"""Full-512-d linear probe on the cached encoder features.
The figure scripts probe only the 2-3 PCA axes they draw, which understates what
the representation holds. This probes the whole feature vector, which is the
number to quote when asking "does the encoder know which cable this is".
Episode-level split: frames inside one episode are near-duplicates, so a
frame-level split would leak the answer.
"""
import argparse
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
ap = argparse.ArgumentParser()
ap.add_argument("--cache", default="cache/cache_cable_feats.npz")
ap.add_argument("--label-key", default="y")
a = ap.parse_args()
d = np.load(a.cache, allow_pickle=True)
X, y, ep = d["feats"], d[a.label_key], d["eps"]
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))
clf = make_pipeline(StandardScaler(), LogisticRegression(max_iter=5000))
clf.fit(X[tr], y[tr])
print(f" classes : {list(np.unique(y))}")
print(f" frames : {tr.sum()} train / {(~tr).sum()} held out (episode-level split)")
print(f" probe accuracy on the full 512-d feature : {clf.score(X[~tr], y[~tr]):.1%}"
f" (chance {1 / len(np.unique(y)):.1%})")
|