gallery/src/cross-binding-reference_all-classes_python.py at main · libKriging/gallery · GitHub

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
# %% [markdown]
# # Python · all classes — cross-binding consistency: the Python reference
#
# **Idea.** All bindings wrap the same C++ library: on the same data with the same
# options, they must give the same numbers, for every model class and fit method. This
# notebook writes the Python values, the reference of the comparison.
#
# **Reference.** The R, Octave and Julia notebooks `cross-binding_all-classes_*`,
# compared by `tools/report.py`.
#
# **Tags.** class: Kriging, NestedKriging, WarpKriging, MLPKriging · kernel: matern5_2 ·
# objective: LL · method: exact, vecchia, nystrom · noise: none, nugget, variances ·
# warping: kumaraswamy · aggregation: NK · topic: cross-binding
#
# All bindings (Python, R, Octave, Julia) wrap the same C++ library: on the
# same data with the same options they must return the same numbers. Each
# language runs the protocol below on the shared datasets of `data/` and writes
# `results/xbind_<lang>.csv` (`lang,dataset,quantity,index,value`);
# `tools/report.py` then compares every language to Python.
#
# Protocol, for each dataset (Branin n=25, Hartmann 3 n=40, Borehole n=80, on
# the unit cube), kernel `matern5_2`, trend `constant`, θ₀ = 0.3 in every
# dimension:
#
# | quantity | how | tolerance in the report |
# |---|---|---|
# | `fixed_sigma2`, `fixed_beta`, `fixed_ll`, `fixed_loo` | `optim="none"`, θ = θ₀ fixed | rel. 1e-8 (error) |
# | `fixed_mean`, `fixed_stdev` | predict at the first 50 test points | rel. 1e-8 (error) |
# | `llfun`, `llfun_grad` | `logLikelihoodFun(θ₀)` with gradient | rel. 1e-8 (error) |
# | `fit_theta`, `fit_ll`, `fit_mean` | default fit (`BFGS`) | rel. 1e-4 (warn: optimizer paths may differ) |
# | `sim` | `simulate(nsim=3, seed=123)` at the first 5 test points (5 × 3, column-major) | rel. 1e-8 (warn: RNG may be binding-specific) |
#
# Then, on Branin only, every other model class and fit method, so that each
# binding is checked on all of them (the class-specific notebooks are in Python
# only): with fixed hyperparameters the numbers must be identical (error); a
# fit is compared at warn level, as above.
#
# | quantity | how | tolerance in the report |
# |---|---|---|
# | `nugget_fixed_mean`, `nugget_fixed_stdev` | `noise="nugget"`, θ = θ₀, σ² = var(y), nugget = 0.01 var(y), all fixed | rel. 1e-8 (error) |
# | `noise_fixed_mean`, `noise_fixed_stdev` | known noise variances 0.01 var(y), θ = θ₀ and σ² fixed | rel. 1e-8 (error) |
# | `nugget_fit_ll`, `nugget_fit_mean` | `noise="nugget"`, default fit | rel. 1e-4 (warn) |
# | `vecchia_fixed_ll`, `vecchia_fixed_mean`, `nystrom_fixed_ll`, `nystrom_fixed_mean` | `method="vecchia(10)"` / `"nystrom(10)"`, θ = θ₀ fixed; likelihood and prediction with that method | rel. 1e-8 (error) |
# | `warp_fit_mean` | `WarpKriging`, `kumaraswamy` on both inputs, default fit | rel. 1e-4 (warn) |
# | `mlp_fit_mean` | `MLPKriging`, hidden layer of 8 (`tanh`), 2 features, `BFGS10` | rel. 1e-4 (warn) |
# | `nested_fit_mean` | `NestedKriging`, 2 groups (`kmeans`, seed 1), `NK` | rel. 1e-4 (warn) |

# %%
import csv

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import pylibkriging as lk

import gallery
from gallery import compat
from gallery import DATA_DIR as DATA, RESULTS_DIR

suite = gallery.Suite("cross-binding-reference_all-classes_python")
print("pylibkriging", compat.version())
DATASETS = ["branin", "hartmann3", "borehole"]
rows = []


def put(dataset, quantity, values):
    for i, v in enumerate(np.ravel(values), start=1):
        rows.append(("python", dataset, quantity, i, f"{float(v):.17g}"))


def load(name, stem):
    return np.loadtxt(DATA / f"{name}_{stem}.csv", delimiter=",", ndmin=2)

# %%
for name in DATASETS:
    X, y, Xt = load(name, "X"), load(name, "y").ravel(), load(name, "Xt")[:50]
    d = X.shape[1]
    theta0 = np.full((1, d), 0.3)
    kf = lk.Kriging(y, X, "matern5_2", regmodel="constant", optim="none",
                    parameters={"theta": theta0, "is_theta_estim": False})
    put(name, "fixed_sigma2", kf.sigma2())
    put(name, "fixed_beta", kf.beta())
    put(name, "fixed_ll", kf.logLikelihood())
    put(name, "fixed_loo", kf.leaveOneOut())
    m, s, *_ = kf.predict(Xt, return_stdev=True)
    put(name, "fixed_mean", m)
    put(name, "fixed_stdev", s)
    ll, grad, _ = kf.logLikelihoodFun(theta0.ravel(), True)
    put(name, "llfun", ll)
    put(name, "llfun_grad", grad)
    suite.check_close(f"{name}: logLikelihood() == logLikelihoodFun(theta0)", kf.logLikelihood(), ll, rtol=1e-10)
    k = lk.Kriging(y, X, "matern5_2")
    put(name, "fit_theta", k.theta())
    put(name, "fit_ll", k.logLikelihood())
    put(name, "fit_mean", k.predict(Xt, return_stdev=False)[0])
    put(name, "sim", np.ravel(np.asarray(kf.simulate(3, 123, Xt[:5])), order="F"))  # column-major, as R/Octave/Julia

# %% [markdown]
# ## Other classes and fit methods (Branin)

# %%
name = "branin"
X, y, Xt = load(name, "X"), load(name, "y").ravel(), load(name, "Xt")[:50]
theta0 = np.full((1, 2), 0.3)
s2 = float(np.var(y)); nug = 0.01 * s2
fixed = {"theta": theta0, "is_theta_estim": False, "sigma2": s2, "is_sigma2_estim": False}
kn = lk.Kriging(y, X, "matern5_2", noise="nugget", optim="none",
                parameters={**fixed, "nugget": nug, "is_nugget_estim": False})
m, s, *_ = kn.predict(Xt, return_stdev=True)
put(name, "nugget_fixed_mean", m); put(name, "nugget_fixed_stdev", s)
kh = lk.Kriging(y, X, "matern5_2", noise=np.full(len(y), nug), optim="none", parameters=fixed)
m, s, *_ = kh.predict(Xt, return_stdev=True)
put(name, "noise_fixed_mean", m); put(name, "noise_fixed_stdev", s)
kf = lk.Kriging(y, X, "matern5_2", noise="nugget")
put(name, "nugget_fit_ll", kf.logLikelihood()); put(name, "nugget_fit_mean", kf.predict(Xt, return_stdev=False)[0])
for meth in ["vecchia", "nystrom"] if compat.has("method") else []:
    ka = lk.Kriging(y, X, "matern5_2", optim="none", parameters={"theta": theta0, "is_theta_estim": False},
                    method=f"{meth}(10)")
    put(name, f"{meth}_fixed_ll", ka.logLikelihood(method=meth))
    put(name, f"{meth}_fixed_mean", ka.predict(Xt, return_stdev=False, method=meth)[0])
if not compat.has("method"):
    suite.check_skip("vecchia / nystrom quantities", "pylibkriging without the method= argument")
kw = lk.WarpKriging(y, X, ["kumaraswamy", "kumaraswamy"], "matern5_2")
put(name, "warp_fit_mean", kw.predict(Xt, return_stdev=False)[0])
km = lk.MLPKriging(y, X, hidden_dims=[8], d_out=2, activation="tanh", kernel="matern5_2", optim="BFGS10")
put(name, "mlp_fit_mean", km.predict(Xt, return_stdev=False)[0])
kk = lk.NestedKriging(y, X, "matern5_2", nb_groups=2, aggregation="NK", seed=1)
put(name, "nested_fit_mean", kk.predict(Xt, return_stdev=False)[0])

# %%
RESULTS_DIR.mkdir(exist_ok=True)
with open(RESULTS_DIR / "xbind_python.csv", "w", newline="") as fh:
    w = csv.writer(fh)
    w.writerow(["lang", "dataset", "quantity", "index", "value"])
    w.writerows(rows)
print(len(rows), "values written")

# %% [markdown]
# ## The reference values
#
# Number of values per quantity and dataset (the report compares each of them
# with the other bindings), and the default fits they come from: predicted vs
# true values on the 50 test points.

# %%
ref = pd.DataFrame(rows, columns=["lang", "dataset", "quantity", "index", "value"])
ref["value"] = ref["value"].astype(float)
display(ref.groupby(["quantity", "dataset"])["value"].agg(["size", "first"]).unstack("dataset"))
fig, axes = plt.subplots(1, len(DATASETS), figsize=(12, 3.6))
for ax, name in zip(axes, DATASETS):
    yt = load(name, "yt").ravel()[:50]
    m = ref[(ref.dataset == name) & (ref.quantity == "fit_mean")]["value"].to_numpy()
    ax.plot(yt, m, "o", ms=4); ax.plot([yt.min(), yt.max()], [yt.min(), yt.max()], "k--", lw=.8)
    q2 = 1 - np.sum((yt - m) ** 2) / np.sum((yt - yt.mean()) ** 2)
    suite.metric(name, "pylibkriging default fit", q2=q2)
    ax.set_title(f"{name}: Q² = {q2:.3f}"); ax.set_xlabel("true"); ax.set_ylabel("predicted")
plt.tight_layout(); plt.show()

# %% [markdown]
# ## Summary
#
# The metrics of this notebook, the number of checks per status, and every check that
# did not pass (FAIL: error-level, fails the run; WARN: warn-level; KNOWN: listed in
# `tools/known_issues.csv`, with its issue; SKIP: feature absent from the installed version).
# All of them are also written to `results/cross-binding-reference_all-classes_python.checks.csv` and `.metrics.csv`.

# %%
suite.summary()

# %%
suite.finalize()