File size: 5,546 Bytes
25055a4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Claim 6: evaluate Algorithms 3 (DoWS) and 4 (T-DoWS) against a primal-dual
baseline on SVM classification with real datasets.

The logbook stated it does not reproduce the real-data SVM comparison and ran no
SVM experiment. Two of the three datasets are run here; MNIST 3-vs-5 could not be
retrieved (OpenML returned HTTP 504).

SVM as constrained optimisation, which is the form the algorithms address:
    minimise  f(w,b) = 0.5||w||^2
    subject to  y_i (w.x_i + b) >= 1   (one constraint per sample)

Randomised feasibility: sample a violated constraint and project onto its
halfspace, rather than forming the full constraint set.
"""
import json, pickle, numpy as np

RES = {}


def load():
    D = pickle.load(open("svm_data.pkl", "rb"))
    out = {}
    for k, (X, y) in D.items():
        X = (X-X.mean(0))/(X.std(0)+1e-9)
        out[k] = (X, y.astype(float))
    return out


C_SLACK = 1.0


def obj(w, b, xi=None):
    """SOFT-margin objective. Hard margin is INFEASIBLE on banknote (min margin
    -0.767 under an exact LinearSVC fit), so violations there could never reach 0
    and the feasibility comparison was measuring an infeasible program."""
    v = 0.5*float(w @ w)
    return v+C_SLACK*float(np.sum(xi)) if xi is not None else v


def viol(X, y, w, b, xi):
    return float(np.maximum(0.0, 1.0-xi-y*(X @ w+b)).max())


def project_one(X, y, w, b, xi, i):
    """Project (w,b,xi_i) onto y_i(w.x_i+b) + xi_i >= 1, xi_i >= 0."""
    a = y[i]*X[i]; c = y[i]
    g = a @ w+c*b+xi[i]-1.0
    if g < 0:
        n2 = a @ a+c*c+1.0
        lam = -g/max(n2, 1e-12)
        w = w+lam*a; b = b+lam*c; xi[i] = xi[i]+lam
    xi[i] = max(0.0, xi[i])
    return w, b


def run(X, y, alg, T=4000, seed=0, nproj=None):
    """nproj scales with the CONSTRAINT COUNT. Claim 4 of this same paper (already
    verified) says infeasibility decays geometrically in the NUMBER of feasibility
    updates; with a fixed 5 projections/step over 569-1372 constraints most
    constraints are almost never visited and violation plateaus."""
    rng = np.random.default_rng(seed)
    n, d = X.shape
    if nproj is None: nproj = max(5, n//4)
    w = np.zeros(d); b = 0.0; w0 = w.copy(); xi = np.ones(n)
    G2 = 1e-12; rmax = 1e-8
    lam_d = 0.0
    hist = []
    for t in range(1, T+1):
        g = w.copy()                                   # grad of 0.5||w||^2
        if alg == "primal_dual":
            i = int(rng.integers(n))
            s = 1.0-y[i]*(X[i] @ w+b)
            gv = -y[i]*X[i] if s > 0 else np.zeros(d)
            gb = -y[i] if s > 0 else 0.0
            eta = 0.5/np.sqrt(t)
            w = w-eta*(g+lam_d*gv); b = b-eta*(lam_d*gb)
            xi[i] = max(0.0, xi[i]-eta*(C_SLACK-lam_d))
            lam_d = max(0.0, lam_d+eta*(s-xi[i]))
        else:
            G2 += g @ g
            rmax = max(rmax, float(np.linalg.norm(w-w0)))
            if alg == "dows":
                eta = rmax/np.sqrt(G2)
            else:                                       # t_dows: tamed, no bounded-Y assumption
                eta = rmax/np.sqrt(G2*np.log(np.e+t))
            w = w-eta*g
            xi = np.maximum(0.0, xi-eta*C_SLACK)
            for _ in range(nproj):                      # randomised feasibility updates
                i = int(rng.integers(n))
                w, b = project_one(X, y, w, b, xi, i)
        if t % 200 == 0:
            hist.append({"t": t, "obj": obj(w, b, xi), "max_violation": viol(X, y, w, b, xi)})
    acc = float(np.mean(np.sign(X @ w+b) == y))
    return w, b, hist, acc


def main():
    data = load()
    rows = []
    for name, (X, y) in data.items():
        for alg in ("dows", "t_dows", "primal_dual"):
            o, v, a = [], [], []
            for s in range(3):
                w, b, h, acc = run(X, y, alg, seed=s)
                o.append(h[-1]["obj"]); v.append(h[-1]["max_violation"]); a.append(acc)
            rows.append({"dataset": name, "n": int(X.shape[0]), "d": int(X.shape[1]),
                         "algorithm": alg, "final_objective": round(float(np.mean(o)), 5),
                         "final_max_violation": round(float(np.mean(v)), 6),
                         "train_accuracy": round(float(np.mean(a)), 4)})
            print("  %-14s %-12s  objective=%9.4f   max violation=%9.5f   accuracy=%.4f"
                  % (name, alg, np.mean(o), np.mean(v), np.mean(a)), flush=True)
    RES["claim6_svm"] = {"rows": rows, "seeds": 3, "iterations": 4000,
        "datasets_run": sorted(data.keys()),
        "mnist_note": "MNIST 3-vs-5 unavailable: OpenML returned HTTP 504"}
    # comparison summary
    summ = []
    for name in data:
        sub = {r["algorithm"]: r for r in rows if r["dataset"] == name}
        summ.append({"dataset": name,
                     "dows_beats_pd_on_violation": bool(sub["dows"]["final_max_violation"] <= sub["primal_dual"]["final_max_violation"]),
                     "t_dows_beats_pd_on_violation": bool(sub["t_dows"]["final_max_violation"] <= sub["primal_dual"]["final_max_violation"]),
                     "dows_acc_minus_pd": round(sub["dows"]["train_accuracy"]-sub["primal_dual"]["train_accuracy"], 4)})
        print("  [%s] DoWS beats PD on feasibility: %s | T-DoWS: %s | accuracy delta %+.4f"
              % (name, summ[-1]["dows_beats_pd_on_violation"], summ[-1]["t_dows_beats_pd_on_violation"],
                 summ[-1]["dows_acc_minus_pd"]), flush=True)
    RES["claim6_svm"]["summary"] = summ
    json.dump(RES, open("svm_results.json", "w"), indent=1)


if __name__ == "__main__":
    main(); print("DONE")