nd-arrow-array 0.5.0

N-dimensional Arrow arrays as an Arrow extension type
Documentation 
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

"""Generate numpy.json: NumPy reference results for tests/numpy.rs.

Run from the crate root: python3 tests/fixtures/gen_numpy.py
NaN in NumPy maps to null in Arrow. Reductions use the NumPy nan* functions.
"""

import json
import pathlib
import warnings

import numpy as np

rng = np.random.default_rng(42)
DIMS = ["a", "b", "c"]


def encode(array, dims):
    flat = [None if np.isnan(v) else float(v) for v in np.asarray(array, dtype=float).ravel()]
    return {"dims": [[d, int(s)] for d, s in zip(dims, np.shape(array))], "data": flat}


def case(name, op, x, dims, expected, out_dims, **params):
    return {
        "name": name,
        "op": op,
        "input": encode(x, dims),
        "params": params,
        "expected": encode(expected, out_dims),
    }


cases = []

x = rng.normal(size=(3, 4, 5)).round(3)
x[0, 1, 2] = np.nan
x[2, 3, :] = np.nan
x[1, 0, 4] = np.nan

reductions = {
    "sum": np.nansum,
    "mean": np.nanmean,
    "min": np.nanmin,
    "max": np.nanmax,
    "median": np.nanmedian,
    "count": lambda v, axis: np.sum(~np.isnan(v), axis=axis),
    "var0": lambda v, axis: np.nanvar(v, axis=axis, ddof=0),
    "var1": lambda v, axis: np.nanvar(v, axis=axis, ddof=1),
    "std1": lambda v, axis: np.nanstd(v, axis=axis, ddof=1),
    "quantile30": lambda v, axis: np.nanquantile(v, 0.3, axis=axis),
}
axes_options = [None, (0,), (1,), (2,), (0, 2), (0, 1)]

with warnings.catch_warnings():
    warnings.simplefilter("ignore", RuntimeWarning)
    for op, fn in reductions.items():
        for axes in axes_options:
            expected = fn(x, axis=axes)
            kept = [d for i, d in enumerate(DIMS) if axes is not None and i not in axes]
            names = None if axes is None else [DIMS[i] for i in axes]
            cases.append(case(f"{op}_{axes}", op, x, DIMS, expected, kept, axes=names))

# Arg reductions on data without all-NaN groups.
y = rng.normal(size=(3, 4, 5)).round(3)
y[0, 0, 0] = np.nan
for axis in range(3):
    for op, fn in [("argmin", np.nanargmin), ("argmax", np.nanargmax)]:
        kept = [d for i, d in enumerate(DIMS) if i != axis]
        cases.append(case(f"{op}_{axis}", op, y, DIMS, fn(y, axis=axis), kept, axes=[DIMS[axis]]))

for axis in range(3):
    expected = np.nancumsum(x, axis=axis)
    expected[np.isnan(x)] = np.nan
    cases.append(case(f"cumsum_{axis}", "cumsum", x, DIMS, expected, DIMS, dim=DIMS[axis]))
    for n in (1, 2):
        expected = np.diff(x, n=n, axis=axis)
        cases.append(case(f"diff_{axis}_{n}", "diff", x, DIMS, expected, DIMS, dim=DIMS[axis], n=n))

for perm in [(2, 0, 1), (1, 2, 0), (2, 1, 0)]:
    cases.append(
        case(f"transpose_{perm}", "transpose", x, DIMS, np.transpose(x, perm), [DIMS[i] for i in perm],
             order=[DIMS[i] for i in perm])
    )

z = rng.permutation(60).reshape(3, 4, 5).astype(float)
z[1, 2, 3] = np.nan
for axis in range(3):
    cases.append(case(f"sort_{axis}", "sort", z, DIMS, np.sort(z, axis=axis), DIMS, dim=DIMS[axis]))
    cases.append(
        case(f"argsort_{axis}", "argsort", z, DIMS, np.argsort(z, axis=axis, kind="stable"), DIMS,
             dim=DIMS[axis])
    )

sorted_values = np.sort(rng.integers(0, 20, size=15)).astype(float)
needles = rng.uniform(-2, 22, size=10).round(1)
for side in ("left", "right"):
    cases.append({
        "name": f"searchsorted_{side}",
        "op": "searchsorted",
        "input": encode(sorted_values, ["x"]),
        "params": {"side": side, "values": encode(needles, ["q"])},
        "expected": encode(np.searchsorted(sorted_values, needles, side=side), ["q"]),
    })

xp = np.array([0.0, 1.0, 2.5, 4.0, 7.0])
fp = rng.normal(size=(2, 5)).round(3)
targets = np.array([0.0, 0.5, 2.0, 3.9, 6.5, 7.0])
cases.append({
    "name": "interp",
    "op": "interp",
    "input": encode(fp, ["y", "x"]),
    "params": {"x": encode(xp, ["x"]), "new_x": encode(targets, ["x"])},
    "expected": encode(np.stack([np.interp(targets, xp, row) for row in fp]), ["y", "x"]),
})

# Pairwise summation: contiguous float sums must match NumPy bit for bit.
long = rng.normal(size=(3, 1000))
cases.append(case("exact_sum_long_axis", "sum", long, ["a", "b"], np.sum(long, axis=1), ["a"], axes=["b"]))
cases.append(case("exact_mean_long_axis", "mean", long, ["a", "b"], np.mean(long, axis=1), ["a"], axes=["b"]))

a = rng.normal(size=(3, 4)).round(3)
b = rng.normal(size=(4,)).round(3)
for op, fn in [("add", np.add), ("sub", np.subtract), ("mul", np.multiply), ("div", np.divide)]:
    cases.append({
        "name": f"{op}_broadcast",
        "op": op,
        "input": encode(a, ["x", "y"]),
        "params": {"other": encode(b, ["y"])},
        "expected": encode(fn(a, b), ["x", "y"]),
    })

# Functions added after the first kernel set.
m = rng.integers(-9, 10, size=(3, 4)).astype(float)
d = rng.choice([-4.0, -3.0, 2.0, 5.0], size=(4,))
for op, fn in [("floor_divide", np.floor_divide), ("mod", np.mod)]:
    cases.append({
        "name": f"{op}_signs",
        "op": op,
        "input": encode(m, ["x", "y"]),
        "params": {"other": encode(d, ["y"])},
        "expected": encode(fn(m, d), ["x", "y"]),
    })
for axis in range(2):
    dim = ["x", "y"][axis]
    cases.append(case(f"cummax_{axis}", "cummax", m, ["x", "y"], np.maximum.accumulate(m, axis=axis), ["x", "y"], dim=dim))
    cases.append(case(f"cummin_{axis}", "cummin", m, ["x", "y"], np.minimum.accumulate(m, axis=axis), ["x", "y"], dim=dim))
    cases.append(case(f"ptp_{axis}", "ptp", m, ["x", "y"], np.ptp(m, axis=axis), [["y"], ["x"]][axis], axes=[dim]))
    for mode in ("edge", "reflect", "symmetric", "wrap"):
        width = [(0, 0), (0, 0)]
        width[axis] = (2, 3)
        cases.append(case(f"pad_{mode}_{axis}", "pad", m, ["x", "y"], np.pad(m, width, mode=mode), ["x", "y"],
                          dim=dim, before=2, after=3, mode=mode))
    cases.append(case(f"tile_{axis}", "tile", m, ["x", "y"], np.concatenate([m, m, m], axis=axis), ["x", "y"], dim=dim, reps=3))
    cases.append(case(f"repeat_{axis}", "repeat", m, ["x", "y"], np.repeat(m, 2, axis=axis), ["x", "y"], dim=dim, repeats=2))

grid = rng.normal(size=(6, 8))
cases.append(case("coarsen_mean", "coarsen_mean", grid, ["y", "x"],
                  grid.reshape(3, 2, 2, 4).mean(axis=(1, 3)), ["y", "x"], windows=[["y", 2], ["x", 4]]))

samples = rng.normal(size=200)
edges = np.array([-3.0, -1.0, -0.2, 0.0, 0.5, 2.0, 3.0])
counts, _ = np.histogram(samples, bins=edges)
cases.append({
    "name": "histogram",
    "op": "histogram",
    "input": encode(samples, ["x"]),
    "params": {"edges": encode(edges, ["e"])},
    "expected": encode(counts, ["bin"]),
})

out = pathlib.Path(__file__).with_name("numpy.json")
out.write_text(json.dumps({"numpy": np.__version__, "cases": cases}, indent=1))
print(f"wrote {len(cases)} cases to {out}")