File size: 7,728 Bytes
ce85a7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
"""Figures for the BALLAST reproduction logbook (Plotly HTML + raw CSV)."""

from __future__ import annotations

import json
import os
import sys

import numpy as np
import plotly.graph_objects as go

OUT = "outputs"
COL = {
    "unif": "#7f8c8d", "sobol": "#2980b9", "dist_sep": "#8e44ad",
    "eig": "#e67e22", "ballast_opt": "#16a085", "ballast_true": "#c0392b",
}
NAME = {
    "unif": "UNIF", "sobol": "SOBOL", "dist_sep": "DIST-SEP", "eig": "EIG",
    "ballast_opt": "BALLAST-opt", "ballast_true": "BALLAST-true",
}
LAYOUT = dict(
    template="plotly_white", font=dict(size=13), height=420,
    margin=dict(l=60, r=20, t=50, b=50),
)


def _save(fig, name, csv_rows=None, header=None):
    os.makedirs(OUT, exist_ok=True)
    fig.write_html(f"{OUT}/{name}.html", include_plotlyjs="cdn")
    if csv_rows is not None:
        with open(f"{OUT}/{name}.csv", "w") as f:
            f.write(",".join(header) + "\n")
            for r in csv_rows:
                f.write(",".join(str(x) for x in r) + "\n")
    print("wrote", name)


def fig_spde_cost(path="outputs/raw/spde_cost/spde_cost.json"):
    d = json.load(open(path))
    rows = d["rows"]
    n = [r["n_sampT"] for r in rows]
    sp = [r["spde_s"] for r in rows]
    nv = [r["naive_s"] for r in rows]
    ok = [i for i, v in enumerate(nv) if v is not None]

    fig = go.Figure()
    fig.add_trace(go.Scatter(x=n, y=sp, name="SPDE (Sec. 4.1)", mode="lines+markers",
                             line=dict(color="#16a085", width=3)))
    fig.add_trace(go.Scatter(x=[n[i] for i in ok], y=[nv[i] for i in ok],
                             name="naive dense", mode="lines+markers",
                             line=dict(color="#c0392b", width=3)))
    # extrapolate the naive cubic to show where it goes
    i0 = ok[-1]
    ext = [x for x in n if x > n[i0]]
    if ext:
        fig.add_trace(go.Scatter(
            x=[n[i0]] + ext,
            y=[nv[i0]] + [nv[i0] * (x / n[i0]) ** 3 for x in ext],
            name="naive (cubic extrapolation, out of memory)", mode="lines",
            line=dict(color="#c0392b", width=2, dash="dot")))
    fig.update_layout(
        title="Posterior sampling cost: SPDE stays linear, naive dense goes cubic then out of memory",
        xaxis=dict(title="N_sampT (sampled time slices)", type="log"),
        yaxis=dict(title="seconds for J=20 samples (A100, fp64)", type="log"),
        legend=dict(x=0.02, y=0.98), **LAYOUT)
    _save(fig, "claim2_spde_cost",
          [(r["n_sampT"], r["spde_s"], r["naive_s"], r["naive_cov_gb"]) for r in rows],
          ["n_sampT", "spde_seconds", "naive_seconds", "naive_cov_gb"])


def fig_ablation(path="outputs/claim5.json"):
    d = json.load(open(path))
    fig = go.Figure()
    colors = {"3.0": "#2980b9", "5.0": "#16a085", "7.0": "#c0392b"}
    rows = []
    for t, r in d.items():
        J = np.array(r["J"])
        m = np.array(r["gap_mc_mean"])
        s = np.array(r["gap_mc_se2"])
        c = colors.get(str(t), "#333")
        fig.add_trace(go.Scatter(x=J, y=m, name=f"BALLAST t={t}", mode="lines",
                                 line=dict(color=c, width=2.5)))
        fig.add_trace(go.Scatter(
            x=np.concatenate([J, J[::-1]]), y=np.concatenate([m + s, (m - s)[::-1]]),
            fill="toself", fillcolor=c.replace("#", "rgba(").replace("", "") if False else c,
            opacity=0.15, line=dict(width=0), showlegend=False, hoverinfo="skip"))
        fig.add_trace(go.Scatter(x=[J[0], J[-1]], y=[r["eig_gap_mc"]] * 2,
                                 name=f"EIG t={t}", mode="lines",
                                 line=dict(color=c, width=1.5, dash="dash")))
        fig.add_trace(go.Scatter(x=[J[0], J[-1]], y=[r["unif_gap_mc"]] * 2,
                                 name=f"UNIF t={t}", mode="lines",
                                 line=dict(color=c, width=1.5, dash="dot")))
        for j, v in zip(J, m):
            rows.append((t, j, v))
    fig.add_hline(y=1.0, line=dict(color="black", width=1))
    fig.add_annotation(x=np.log10(120), y=np.log10(1.0), text="1% cut-off",
                       showarrow=False, yshift=10)
    fig.add_vline(x=20, line=dict(color="#888", width=1, dash="dash"))
    fig.update_layout(
        title="Percentage utility gap vs Monte Carlo sample number J (2 s.e. bands)",
        xaxis=dict(title="J (posterior field samples)", type="log"),
        yaxis=dict(title="Monte Carlo % utility gap", type="log"),
        **LAYOUT)
    _save(fig, "claim5_ablation", rows, ["decision_time", "J", "gap_mc_pct"])


def fig_policy(path, tag, title):
    d = json.load(open(path))
    pols = d["policies"]
    n_dep = d["n_deploy"]
    x = np.arange(1, n_dep + 1)

    # --- average rank
    fig = go.Figure()
    for i, p in enumerate(pols):
        m = np.array(d["rank_mean"][i])
        s = np.array(d["rank_se2"][i])
        fig.add_trace(go.Scatter(x=x, y=m, name=NAME[p], mode="lines+markers",
                                 line=dict(color=COL[p], width=2.5),
                                 error_y=dict(type="data", array=s, visible=True, thickness=1)))
    fig.update_layout(
        title=f"{title}: average policy rank (1 = best), {d['n_runs']} runs",
        xaxis=dict(title="drifters deployed"),
        yaxis=dict(title="average rank"), **LAYOUT)
    _save(fig, f"{tag}_rank",
          [(p, i + 1, d["rank_mean"][j][i]) for j, p in enumerate(pols) for i in range(n_dep)],
          ["policy", "n_drifters", "mean_rank"])

    # --- iso-performance
    fig = go.Figure()
    for p in pols:
        if p == "unif":
            continue
        m = np.array(d["iso"][p]["mean"])
        s = np.array(d["iso"][p]["se2"])
        fig.add_trace(go.Scatter(x=x, y=m, name=NAME[p], mode="lines+markers",
                                 line=dict(color=COL[p], width=2.5),
                                 error_y=dict(type="data", array=s, visible=True, thickness=1)))
    fig.add_hline(y=0, line=dict(color="#7f8c8d", width=1, dash="dash"))
    fig.update_layout(
        title=f"{title}: drifters saved vs UNIF (iso-performance), {d['n_runs']} runs",
        xaxis=dict(title="drifters deployed"),
        yaxis=dict(title="drifters saved (positive = better)"), **LAYOUT)
    _save(fig, f"{tag}_iso",
          [(p, i + 1, d["iso"][p]["mean"][i], d["iso"][p]["se2"][i])
           for p in pols for i in range(n_dep)],
          ["policy", "n_drifters", "drifters_saved", "se2"])

    # --- error curves
    fig = go.Figure()
    for p in pols:
        m = np.array(d["err_mean"][p])
        s = np.array(d["err_se2"][p])
        fig.add_trace(go.Scatter(x=x, y=m, name=NAME[p], mode="lines",
                                 line=dict(color=COL[p], width=2.5),
                                 error_y=dict(type="data", array=s, visible=True, thickness=1)))
    fig.update_layout(
        title=f"{title}: field error vs drifters deployed, {d['n_runs']} runs",
        xaxis=dict(title="drifters deployed"),
        yaxis=dict(title="mean L2 error of posterior mean field"), **LAYOUT)
    _save(fig, f"{tag}_error",
          [(p, i + 1, d["err_mean"][p][i]) for p in pols for i in range(n_dep)],
          ["policy", "n_drifters", "mean_l2_error"])


if __name__ == "__main__":
    which = sys.argv[1] if len(sys.argv) > 1 else "all"
    if which in ("all", "spde"):
        fig_spde_cost()
    if which in ("all", "ablation") and os.path.exists("outputs/claim5.json"):
        fig_ablation()
    if which in ("all", "synth") and os.path.exists("outputs/claim3.json"):
        fig_policy("outputs/claim3.json", "claim3", "Temporal Helmholtz ground truth")
    if which in ("all", "suntans") and os.path.exists("outputs/claim4.json"):
        fig_policy("outputs/claim4.json", "claim4", "SUNTANS ground truth")