File size: 17,746 Bytes
ecc81b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
"""Phase 5 acceptance for td.data. See PLAN.md.

The load-bearing test is that values land in the *right cells*, checked against
an independently built reference. A shape check would pass just as happily on a
mis-shuffled lattice, which is the exact failure this layer exists to prevent.
"""

import pytest
import torch
from torch.utils.data import DataLoader

import torch_dimensions as td
from torch_dimensions.data import (
    LatticeDataset,
    LatticeSource,
    LatticeWindow,
    TensorSource,
    collate_lattice,
    from_coords,
    from_table,
)

STATES = ("CA", "NY", "TX")
SKUS = ("a", "b")


def rows(n_time=8, skip=(("NY", "b"),)):
    """Long-format rows whose value encodes its own coordinates, so a
    misplacement is detectable rather than merely plausible."""
    coords, times, values = [], [], []
    for t in range(n_time):
        for s in STATES:
            for k in SKUS:
                if (s, k) in skip:
                    continue
                coords.append((s, k))
                times.append(2000 + t)
                values.append([t * 100.0 + STATES.index(s) * 10.0 + SKUS.index(k)])
    return coords, times, values


# -- from_coords -------------------------------------------------------------


def test_vocabularies_are_sorted_and_shape_follows():
    cm = from_coords([("TX", "b"), ("CA", "a"), ("CA", "b")], names=("state", "sku"))
    assert cm.vocabs == (("CA", "TX"), ("a", "b"))
    assert cm.lattice.shape == (2, 2)
    assert cm.lattice.axis_names == ("time", "state", "sku")


def test_unobserved_combinations_are_marked_absent():
    cm = from_coords([("CA", "a"), ("CA", "b"), ("TX", "a")])
    assert cm.lattice.n_valid == 3 and cm.lattice.n_cells == 4
    assert not cm.lattice.valid[1, 1]  # TX/b never seen


def test_a_fully_observed_grid_is_dense_not_a_mask_of_all_true():
    cm = from_coords([(s, k) for s in STATES for k in SKUS])
    assert cm.lattice.is_dense


def test_dense_flag_overrides_the_inference():
    cm = from_coords([("CA", "a"), ("TX", "b")], dense=True)
    assert cm.lattice.is_dense and cm.lattice.n_cells == 4


def test_time_axis_is_optional():
    assert not from_coords([("CA", "a")], time=False).lattice.time


def test_index_round_trips_through_decode():
    cm = from_coords([(s, k) for s in STATES for k in SKUS], names=("state", "sku"))
    pairs = [(s, k) for s in STATES for k in SKUS]
    for row, flat in zip(pairs, cm.index.tolist(), strict=True):
        assert cm.decode(flat) == row


def test_encode_matches_the_index_built_at_construction():
    coords = [("TX", "b"), ("CA", "a"), ("NY", "b")]
    cm = from_coords(coords)
    assert torch.equal(cm.encode(coords), cm.index)


def test_encode_refuses_unknown_values_rather_than_folding_them_to_zero():
    cm = from_coords([("CA", "a")], names=("state", "sku"))
    with pytest.raises(KeyError, match="'state'"):
        cm.encode([("ZZ", "a")])
    with pytest.raises(ValueError, match="expected 2 coordinates"):
        cm.encode([("CA",)])


def test_integer_coordinates_work_too():
    cm = from_coords(torch.tensor([[0, 1], [2, 0]]))
    assert cm.lattice.shape == (2, 2)


@pytest.mark.parametrize(
    ("coords", "match"),
    [([], "at least one"), ([("a", "b"), ("c",)], "same length")],
)
def test_from_coords_errors(coords, match):
    with pytest.raises(ValueError, match=match):
        from_coords(coords)


# -- from_table --------------------------------------------------------------


def test_values_land_in_the_cells_their_coordinates_name():
    """The whole point of the layer. Checked against the encoding baked into
    each value, not against another call to our own mapping."""
    coords, times, values = rows()
    table = from_table(coords, times, values, names=("state", "sku"))
    for t in range(len(table.times)):
        for si, s in enumerate(STATES):
            for ki, k in enumerate(SKUS):
                got = table.series[t, si, ki, 0].item()
                if (s, k) == ("NY", "b"):
                    assert got == 0.0, "absent cell must be exactly zero"
                else:
                    assert got == t * 100.0 + si * 10.0 + ki


def test_absent_cells_are_zero_and_marked():
    table = from_table(*rows(), names=("state", "sku"))
    assert table.lattice.n_valid == 5 and table.lattice.n_cells == 6
    assert table.series.masked_select(~table.lattice.valid.reshape(1, 3, 2, 1)).abs().max() == 0


def test_time_axis_is_the_sorted_distinct_timestamps():
    coords, times, values = rows(n_time=4)
    table = from_table(coords, times, values)
    assert table.times == (2000, 2001, 2002, 2003)
    assert len(table) == 4 and table.series.shape[0] == 4


def test_shape_and_feature_count():
    table = from_table(*rows(), names=("state", "sku"))
    assert table.series.shape == (8, 3, 2, 1)
    assert table.n_features == 1
    assert "5/6" in repr(table)


def test_duplicate_time_cell_rows_are_refused():
    """Keeping the last silently would turn a join bug into a plausible
    dataset."""
    with pytest.raises(ValueError, match="duplicate"):
        from_table([("CA", "a"), ("CA", "a")], [2000, 2000], [[1.0], [2.0]])


def test_the_same_cell_at_different_times_is_not_a_duplicate():
    table = from_table([("CA", "a"), ("CA", "a")], [2000, 2001], [[1.0], [2.0]])
    assert table.series.flatten().tolist() == [1.0, 2.0]


@pytest.mark.parametrize(
    ("kwargs", "match"),
    [
        ({"values": [1.0, 2.0]}, r"\(N, F\)"),
        ({"times": [2000]}, "same number of rows"),
    ],
)
def test_from_table_errors(kwargs, match):
    base = {
        "coords": [("CA", "a"), ("CA", "b")],
        "times": [2000, 2000],
        "values": [[1.0], [2.0]],
    }
    with pytest.raises(ValueError, match=match):
        from_table(**{**base, **kwargs})


# -- windowing ---------------------------------------------------------------


def test_windows_tile_the_axis_with_targets_after_inputs():
    w = LatticeWindow(n_time=10, input_len=3, horizon=2)
    assert len(w) == 6
    assert w[0] == (0, 3, 3, 5)
    assert w[-1] == (5, 8, 8, 10)
    for win in w:
        assert win.x1 == win.y0, "a window must never see its own target"


def test_stride_controls_the_step():
    assert [win.x0 for win in LatticeWindow(10, 3, 0, stride=3)] == [0, 3, 6]


def test_horizon_zero_gives_no_target_range():
    w = LatticeWindow(6, 3)
    assert all(win.y0 == win.y1 for win in w)


def test_split_drops_windows_straddling_the_boundary():
    """The gap is deliberate: a straddling window would put post-cut timesteps
    inside a training input."""
    w = LatticeWindow(12, 3, 1)
    before, after = w.split(6)
    assert all(win.y1 <= 6 for win in before)
    assert all(win.x0 >= 6 for win in after)
    assert len(before) + len(after) < len(w)


def test_split_at_time_uses_timestamps():
    w = LatticeWindow(6, 2)
    times = [2000, 2001, 2002, 2003, 2004, 2005]
    by_index = w.split(3)
    by_time = w.split_at_time(times, 2003)
    assert [list(x) for x in by_time] == [list(x) for x in by_index]


def test_split_at_a_time_past_the_end_puts_everything_before():
    w = LatticeWindow(6, 2)
    before, after = w.split_at_time([2000, 2001, 2002, 2003, 2004, 2005], 2099)
    assert len(before) == len(w) and len(after) == 0


def test_slicing_preserves_the_window_type():
    w = LatticeWindow(10, 3)
    assert isinstance(w[:2], LatticeWindow) and len(w[:2]) == 2


@pytest.mark.parametrize(
    ("kwargs", "match"),
    [
        ({"input_len": 0}, "input_len"),
        ({"horizon": -1}, "horizon"),
        ({"stride": 0}, "stride"),
        ({"input_len": 9, "horizon": 5}, "exceeds"),
    ],
)
def test_window_errors(kwargs, match):
    with pytest.raises(ValueError, match=match):
        LatticeWindow(**{"n_time": 10, "input_len": 3, **kwargs})


# -- source / dataset / collate ----------------------------------------------


def test_tensor_source_satisfies_the_protocol():
    table = from_table(*rows(), names=("state", "sku"))
    src = TensorSource(table.series, table.lattice)
    assert isinstance(src, LatticeSource)
    assert len(src) == 8 and src[0:3].shape == (3, 3, 2, 1)


def test_source_rejects_a_series_that_does_not_match_the_lattice():
    table = from_table(*rows(), names=("state", "sku"))
    with pytest.raises(ValueError, match="lattice dims"):
        TensorSource(torch.zeros(8, 9, 9, 1), table.lattice)


def test_a_custom_source_needs_no_inheritance():
    """The protocol is the customization story — this class knows nothing
    about the library."""
    table = from_table(*rows(), names=("state", "sku"))

    class Custom:
        def __init__(self, series, lattice):
            self._s, self._l = series, lattice

        @property
        def lattice(self):
            return self._l

        def __len__(self):
            return self._s.shape[0]

        def __getitem__(self, sl):
            return self._s[sl]

    ds = LatticeDataset(Custom(table.series, table.lattice), LatticeWindow(8, 3, 1))
    assert ds[0]["x"].shape == (3, 3, 2, 1)


def test_dataset_yields_inputs_and_targets_without_the_lattice():
    table = from_table(*rows(), names=("state", "sku"))
    ds = LatticeDataset(TensorSource(table.series, table.lattice), LatticeWindow(8, 3, 1))
    sample = ds[0]
    assert set(sample) == {"x", "y", "window"}, "the lattice is static, not per-sample"
    assert sample.x.shape == (3, 3, 2, 1) and sample.y.shape == (1, 3, 2, 1)
    assert ds.lattice is table.lattice


def test_dataset_refuses_windows_that_run_past_the_source():
    table = from_table(*rows(n_time=4), names=("state", "sku"))
    with pytest.raises(ValueError, match="past the end"):
        LatticeDataset(TensorSource(table.series, table.lattice), LatticeWindow(8, 3, 1))


def test_dataset_refuses_an_empty_window_set():
    table = from_table(*rows(), names=("state", "sku"))
    empty = LatticeWindow(8, 3, 1).split(0)[0]
    with pytest.raises(ValueError, match="empty"):
        LatticeDataset(TensorSource(table.series, table.lattice), empty)


def test_collate_stacks_and_keeps_the_lattice_out():
    table = from_table(*rows(), names=("state", "sku"))
    ds = LatticeDataset(TensorSource(table.series, table.lattice), LatticeWindow(8, 3, 1))
    batch = collate_lattice([ds[0], ds[1], ds[2]])
    assert batch.x.shape == (3, 3, 3, 2, 1)
    assert batch.y.shape == (3, 1, 3, 2, 1)
    assert "lattice" not in batch and len(batch["windows"]) == 3


def test_collate_refuses_ragged_windows_rather_than_padding():
    table = from_table(*rows(), names=("state", "sku"))
    ds = LatticeDataset(TensorSource(table.series, table.lattice), LatticeWindow(8, 3, 1))
    a = ds[0]
    b = dict(a)
    b["x"] = a["x"][:2]
    with pytest.raises(ValueError, match="differing input lengths"):
        collate_lattice([a, b])


def test_collate_refuses_an_empty_batch():
    with pytest.raises(ValueError, match="empty"):
        collate_lattice([])


# -- end to end --------------------------------------------------------------


def test_table_to_dataloader_to_model_to_backward():
    table = from_table(*rows(n_time=16), names=("state", "sku"))
    windows = LatticeWindow(len(table), input_len=4, horizon=1)
    train, _ = windows.split(10)
    ds = LatticeDataset(TensorSource(table.series, table.lattice), train)
    dl = DataLoader(ds, batch_size=2, shuffle=True, collate_fn=collate_lattice)

    model = td.LSTM(d_model=6, n_layers=3, lattice=table.lattice, d_input=table.n_features)
    batch = next(iter(dl))
    out = model(batch.x)
    assert out.shape == (2, 4, 3, 2, 6)
    out.pow(2).mean().backward()
    assert all(p.grad is not None for p in model.parameters())


def test_d_input_is_optional_when_the_data_is_already_d_model_wide():
    lat = td.Lattice(shape=(2, 3), time=True)
    model = td.LSTM(d_model=5, n_layers=3, lattice=lat)
    assert model(torch.randn(1, 4, 2, 3, 5)).shape == (1, 4, 2, 3, 5)


# -- multiprocessing safety ---------------------------------------------------


def _small_dataset():
    lat = td.Lattice(shape=(2, 3), time=True)
    source = TensorSource(torch.randn(8, 2, 3, 4), lat)
    return LatticeDataset(source, LatticeWindow(8, input_len=3, horizon=1))


def test_samples_and_batches_survive_pickling():
    """DataLoader workers send every Sample — and the collated Batch — through
    a pickled queue. `__getattr__ = dict.__getitem__` broke that: pickle probes
    optional dunders with getattr and tolerates AttributeError, not the
    KeyError a dict lookup raises. num_workers>0 crashed outright."""
    import pickle

    sample = _small_dataset()[0]
    back = pickle.loads(pickle.dumps(sample))
    assert torch.equal(back.x, sample.x) and torch.equal(back.y, sample.y)

    batch = collate_lattice([sample, sample])
    back = pickle.loads(pickle.dumps(batch))
    assert torch.equal(back.x, batch.x)


def test_a_missing_field_reads_as_absent_not_as_a_keyerror():
    """`getattr(sample, "y", None)` and `hasattr` must behave; a horizon-0
    sample simply has no target."""
    lat = td.Lattice(shape=(2, 3), time=True)
    source = TensorSource(torch.randn(8, 2, 3, 4), lat)
    sample = LatticeDataset(source, LatticeWindow(8, input_len=3, horizon=0))[0]
    assert getattr(sample, "y", None) is None
    assert not hasattr(sample, "y")
    with pytest.raises(AttributeError):
        _ = sample.y


def test_dataloader_with_worker_processes():
    """The end-to-end form of the pickling guarantee: real worker processes,
    real queues. This is the configuration every user with a large dataset
    reaches for first."""
    dl = DataLoader(_small_dataset(), batch_size=2, num_workers=2, collate_fn=collate_lattice)
    batch = next(iter(dl))
    assert batch.x.shape == (2, 3, 2, 3, 4) and batch.y.shape == (2, 1, 2, 3, 4)


def test_collate_refuses_mixed_target_presence():
    """Keying off samples[0] silently dropped every target whenever the first
    sample happened to lack one."""
    from torch_dimensions.data.source import Sample

    a = Sample(x=torch.zeros(3, 2), window=None)
    b = Sample(x=torch.zeros(3, 2), y=torch.ones(1, 2), window=None)
    with pytest.raises(ValueError, match="mixed-horizon"):
        collate_lattice([a, b])
    with pytest.raises(ValueError, match="mixed-horizon"):
        collate_lattice([b, a])


def test_split_at_time_refuses_unsorted_times():
    """An unsorted timestamp list used to produce a silently nonsensical
    split — the quietest possible leakage bug."""
    w = LatticeWindow(6, input_len=2, horizon=1)
    with pytest.raises(ValueError, match="sorted"):
        w.split_at_time([3, 1, 2, 5, 4, 6], 4)


# -- on-disk source and masked normalization ----------------------------------

# `.npy` is numpy's container; torch does not require numpy, and this project's
# own CI is an install without it.
numpy = pytest.importorskip("numpy")


def test_memmap_source_passes_the_source_conformance_check(tmp_path):
    lat = td.Lattice(shape=(3, 4), names=("h", "w"))
    series = torch.randn(20, 3, 4, 2)
    path = td.data.MemmapSource.write(tmp_path / "series.npy", series)
    source = td.data.MemmapSource(path, lat)
    report = td.testing.check_data_source(source)
    assert report, str(report)
    assert torch.allclose(source[0:20], series, atol=1e-6)


def test_a_memmap_source_reaches_a_worker_process(tmp_path):
    """The failure this class exists to demonstrate: a live mmap handle either
    fails to pickle or pickles into something invalid in the child, and under
    DataLoader(num_workers>0) that presents as a hang (DEBUG.md #9)."""
    import pickle

    lat = td.Lattice(shape=(2, 2), names=("a", "b"))
    series = torch.randn(6, 2, 2, 1)
    path = td.data.MemmapSource.write(tmp_path / "s.npy", series)
    source = td.data.MemmapSource(path, lat)
    _ = source[0:2]  # force the handle open *before* pickling
    revived = pickle.loads(pickle.dumps(source))
    assert revived._array is None, "the mmap handle travelled into the pickle"
    assert torch.allclose(revived[0:6], source[0:6])


def test_a_memmap_source_refuses_a_file_that_is_not_its_lattice(tmp_path):
    path = td.data.MemmapSource.write(tmp_path / "s.npy", torch.randn(5, 9, 9, 1))
    with pytest.raises(ValueError, match="lattice dims"):
        td.data.MemmapSource(path, td.Lattice(shape=(3, 4), names=("h", "w")))


def test_masked_stats_ignore_absent_cells():
    """A mean over a sparse lattice's structural zeros is dragged toward zero
    in proportion to the sparsity, and nothing about the model then looks
    wrong."""
    valid = torch.tensor([[True, False], [True, True]])
    lat = td.Lattice(shape=(2, 2), names=("a", "b"), valid=valid)
    series = torch.full((10, 2, 2, 1), 5.0)
    series[:, 0, 1] = 0.0  # the absent cell, zero as the library guarantees

    naive = series.mean()
    stats = td.data.masked_stats(series, lat)
    assert abs(float(naive) - 3.75) < 1e-5, "the naive mean is dragged toward zero"
    present = stats.mean.reshape(-1)[[0, 2, 3]]
    assert torch.allclose(present, torch.full((3,), 5.0)), stats.mean


def test_normalizer_round_trips():
    lat = td.Lattice(shape=(2, 3), names=("a", "b"))
    series = torch.randn(30, 2, 3, 2) * 4 + 7
    stats = td.data.masked_stats(series, lat)
    assert torch.allclose(stats.invert(stats.apply(series)), series, atol=1e-4)
    normalized = stats.apply(series)
    assert normalized.mean().abs() < 0.1 and abs(float(normalized.std()) - 1) < 0.2


def test_masked_stats_treat_nan_as_absent():
    lat = td.Lattice(shape=(2,), names=("a",))
    series = torch.full((8, 2, 1), 3.0)
    series[0:4, 0] = float("nan")
    stats = td.data.masked_stats(series, lat)
    assert torch.isfinite(stats.mean).all()
    assert abs(float(stats.mean.reshape(-1)[0]) - 3.0) < 1e-5