Skip to content

SciPy

The SciPy pack has 30 rules, and xpd can fix 18 of them.

The SciPy rules report code that raises or warns on SciPy 1.18.1, code that loses digits, and SciPy calls made once per value, in each file that imports scipy. The examples assume import numpy as np.

When a message uses a word from the table, the word means the result of a check on SciPy 1.18.1 with NumPy 2.5.5.

Word in the message What was measured
removed, deprecated The old form raises under python -W error.
inexact The old form lost at least 3 decimal digits against mpmath at 50 digits, and the form to write lost fewer than 1.
faster The range over repeated runs on one core, at 1,000,000 values unless the rule says otherwise. The faster form was not slower at 1,000 values.
legacy, preferred SciPy's own source calls the old form legacy or prefers the new one. Such rules have severity info and claim nothing about speed or accuracy.

A rule reports a name only when the dotted name starts with scipy, or when the file's own imports bind it from scipy. A form made only of NumPy names, such as np.linalg.inv(a) @ b, belongs to the NumPy rules. A rule that asks for an API newer than the project's SciPy is skipped.

Fix What it promises
safe The same object, or a value of the same type that is equal bit for bit. xpd offers no safe fix where the file binds the name again or the call spreads *args or **kwargs.
unsafe The message names what can differ.

Replace each name that SciPy removed or deprecated with the name the rule gives, because a removed name raises on import or on the call.

Rule Severity Fix What it finds
removed-in-scipy error unsafe These names were removed from SciPy and raise AttributeError on SciPy 1.18.1.
deprecated-in-scipy warning unsafe These names warn on SciPy 1.18.1, often because they come from a private module.

Change each argument that SciPy removed, deprecated or ignores, and move each legacy function to the interface that SciPy's documentation names as its replacement.

Rule Severity Fix What it finds
removed-keyword error unsafe These keywords and argument values were removed from SciPy and raise TypeError on SciPy 1.18.1.
deprecated-argument warning unsafe These arguments warn on SciPy 1.18.1, and SciPy will remove them.
default-that-warns warning unsafe These calls rely on a default that warns on SciPy 1.18.1 because it will change.
minimize-unknown-option error none minimize warns and ignores an option that the chosen method does not know, so the setting has no effect.
minimize-unused-argument warning none minimize ignores an argument such as jac= when the chosen method does not use it, and warns.
rng-keyword info unsafe Since SciPy 1.15, these functions take rng=, and SciPy's docs say new code should use it in place of random_state= or seed=.
legacy-optimize info none SciPy lists these optimize functions as legacy and not recommended for new scripts.
legacy-odeint info none SciPy's docs say new code should use solve_ivp instead of odeint.
legacy-interpolate info none SciPy marks interp1d and these other interpolation classes as legacy.
legacy-signal info safe SciPy marks signal.stft, spectrogram and similar functions as legacy.
legacy-api info safe SciPy keeps these names only for old code.

Use the SciPy function that computes the quantity directly, because the two-step form loses digits or overflows where the direct function doesn't.

Rule Severity Fix What it finds
log-of-special warning unsafe A scipy.special function such as ndtr underflows to 0 far in a tail, so its log is -inf there.
log-of-dist-method warning unsafe A distribution's pdf, cdf or sf underflows to 0 far in a tail, so np.log of it gives -inf there.
cdf-complement warning unsafe 1 - dist.cdf(x) drops the digits of a small right tail and gives 0 once the cdf rounds to 1.
special-complement warning unsafe 1 - special.f(x) drops the digits of a small result and gives 0 once the function rounds to 1.
ppf-of-complement warning unsafe ppf(1 - p) rounds away the digits of a small p, so it loses precision far in the tail.
comb-by-factorials warning unsafe factorial overflows to inf above 170, so a binomial coefficient built from factorials gives inf or nan for larger numbers.
log-of-scipy-det warning unsafe linalg.det overflows or underflows for a large matrix, so its log is inf or -inf.

SciPy calls made once per value or once per pair

Section titled “SciPy calls made once per value or once per pair”

Replace each SciPy call that Python makes once per value, once per pair of points or once per pass of a loop with one call on whole arrays, which does the same work in compiled code. Each speedup is a median of 7 runs on one core with SciPy 1.18.1, on float64 and float32, at 1,000,000 values unless the rule names another size.

Rule Severity Fix What it finds
distance-per-pair warning none Calling a distance function once for every pair of points runs Python for each pair.
callable-metric warning unsafe Passing a metric function such as euclidean to cdist or pdist calls it once for every pair in Python.
per-value-scipy-call warning none Calling a SciPy function that takes whole arrays once for every value was 7 to 23 times slower.
query-per-point warning none Querying a KD tree once for every point makes one call for each point.
solve-in-loop warning none spsolve or solve inside a loop factors the same matrix again on every pass.
sparse-inv-then-multiply warning unsafe The inverse of most sparse matrices is dense, so inv(A) @ b builds a huge array.

Build a sparse matrix in a format made for writes, and use the sparse array classes in place of the matrix classes.

Rule Severity Fix What it finds
empty-csr-write warning none Writing into a CSR or CSC matrix that started empty warns, raises under python -W error, and copies the whole matrix each time.
sparse-matrix-api info none SciPy expects to deprecate the sparse matrix classes.

The API that SciPy's docs prefer on SciPy 1.18

Section titled “The API that SciPy's docs prefer on SciPy 1.18”

Use the class or module that SciPy's documentation prefers, because the older name stays only for old code.

Rule Severity Fix What it finds
ckdtree info unsafe cKDTree is the same as KDTree, and SciPy keeps the name only for old code.
legacy-fftpack info none scipy.fftpack is legacy, and SciPy says new code should use scipy.fft.