def test_pca__raise_on_missing_data(sample_dataset, sentinel):
ac = sample_dataset["call_alternate_allele_count"]
ac = ac.where(sample_dataset["call_alternate_allele_count"] == 1, sentinel)
ds = sample_dataset.assign(call_alternate_allele_count=ac)
with pytest.raises(ValueError,
match="Input data cannot contain missing values"):
pca.pca(ds, n_components=2)
def test_pca__default_allele_counts_with_index(sample_dataset):
pca.pca(
sample_dataset.drop_vars("call_alternate_allele_count").set_index(
{"variants": ("variant_contig", "variant_position")}),
n_components=2,
merge=False,
).compute()
def test_pca__lazy_evaluation(shape, chunks, algorithm):
# Ensure that all new variables are backed by lazy arrays
if algorithm == "tsqr" and all(c > 0 for c in chunks):
return
ds = simulate_dataset(*shape, chunks=chunks) # type: ignore[misc]
ds = pca.pca(ds, n_components=2, algorithm=algorithm, merge=False)
for v in ds:
assert isinstance(ds[v].data, da.Array)
def test_pca__array_backend(backend, algorithm):
# Ensure that calculation succeeds regardless of array input backend
ds = simulate_dataset(25, 5)
ds["call_alternate_allele_count"] = ds["call_alternate_allele_count"].copy(
data=backend.asarray(ds["call_alternate_allele_count"]))
ds = pca.pca(ds, n_components=2, algorithm=algorithm, merge=False)
for v in ds:
ds[v].compute()
def test_pca__tsqr_allel_comparison(shape, chunks, n_components):
# Validate chunked, non-random implementation vs scikit-allel single chunk results
ds = simulate_dataset(*shape, chunks=chunks) # type: ignore[misc]
ds_sg = pca.pca(ds, n_components=n_components, algorithm="tsqr")
ds_sk = allel_pca(
ds["call_alternate_allele_count"].values.astype("float32"),
n_components=n_components,
scaler="patterson",
randomized=False,
)
assert ds_sg["sample_pca_projection"].values.dtype == np.float32
assert ds_sk["sample_pca_projection"].values.dtype == np.float32
validate_allel_comparison(ds_sg, ds_sk)
def test_pca__stability(stability_test_result, chunks, algorithm):
# Ensure that results are stable across algorithms and that sign flips
# do not occur when chunking changes
if algorithm == "tsqr" and all(c > 0 for c in chunks):
return
shape, expected = stability_test_result
ds = simulate_dataset(*shape, chunks=chunks, n_cohort=3) # type: ignore[misc]
actual = pca.pca(
ds, n_components=2, algorithm=algorithm, n_iter=6, random_state=0, merge=False
)
# Results are expected to change slightly with chunking, but they
# will change drastically (far more than 1e-5) if a sign flip occurs
xr.testing.assert_allclose(expected, actual, atol=1e-5)
def test_pca__randomized_allel_comparison(shape, chunks, n_components):
# Validate chunked, randomized implementation vs scikit-allel single chunk results --
# randomized validation requires more data, more structure, and fewer components in
# order for results to be equal within the same tolerance as deterministic svd.
ds = simulate_dataset(*shape, chunks=chunks, n_cohort=3) # type: ignore[misc]
ds_sg = pca.pca(
ds, n_components=n_components, algorithm="randomized", n_iter=5, random_state=0
)
ds_sk = allel_pca(
ds["call_alternate_allele_count"].values.astype("float32"),
n_components=n_components,
scaler="patterson",
randomized=True,
iterated_power=5,
random_state=0,
)
assert ds_sg["sample_pca_projection"].values.dtype == np.float32
assert ds_sk["sample_pca_projection"].values.dtype == np.float32
validate_allel_comparison(ds_sg, ds_sk)
def stability_test_result(request):
shape = request.param
ds = simulate_dataset(*shape, chunks=(-1, -1), n_cohort=3) # type: ignore[misc]
res = pca.pca(ds, n_components=2, algorithm="tsqr", merge=False)
return shape, res
def test_pca__default_allele_counts(sample_dataset):
pca.pca(
sample_dataset.drop_vars("call_alternate_allele_count"),
n_components=2,
merge=False,
).compute()
