[Numpy-discussion] ENH: Add "where" argument to reduction functions that are missing it
The following reduction functions receive a "where" argument: - https://numpy.org/doc/stable/reference/generated/numpy.max.html - https://numpy.org/doc/stable/reference/generated/numpy.min.html - https://numpy.org/doc/stable/reference/generated/numpy.sum.html - https://numpy.org/doc/stable/reference/generated/numpy.prod.html - https://numpy.org/doc/stable/reference/generated/numpy.mean.html - https://numpy.org/doc/stable/reference/generated/numpy.var.html - https://numpy.org/doc/stable/reference/generated/numpy.std.html and their nan* counterparts, where applicable. Feature request: Add "where" argument to - https://numpy.org/doc/stable/reference/generated/numpy.ptp.html - https://numpy.org/doc/stable/reference/generated/numpy.average.html - https://numpy.org/doc/stable/reference/generated/numpy.median.html - https://numpy.org/doc/stable/reference/generated/numpy.quantile.html - https://numpy.org/doc/stable/reference/generated/numpy.percentile.html - https://numpy.org/doc/stable/reference/generated/numpy.nanmedian.html and their nan* counterparts, where applicable. ptp should be especially straightforward, since it's just max - min. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Re: ENH: Add "where" argument to reduction functions that are missing it
This also applies to - https://numpy.org/doc/stable/reference/generated/numpy.argmax.html - https://numpy.org/doc/stable/reference/generated/numpy.argmin.html and their nan* counterparts. An `initial` argument could be added to handle the empty case, as with np.max and np.min. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: Add more default arguments to various functions
Suggestion: Add the following default arguments to the following functions: - https://numpy.org/doc/stable/reference/generated/numpy.zeros.html: shape=() - https://numpy.org/doc/stable/reference/generated/numpy.ones.html: shape=() - https://numpy.org/doc/stable/reference/generated/numpy.empty.html: shape=() - https://numpy.org/doc/stable/reference/generated/numpy.heaviside.html: x2=0.5 Feel free to add more examples. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: Uniform interface for accessing minimum or maximum value of a dtype
As discussed
[here](https://github.com/numpy/numpy/issues/5032#issuecomment-1830838701),
[here](https://github.com/numpy/numpy/issues/5032#issuecomment-2307927804), and
[here](https://github.com/google/jax/issues/18661#issuecomment-1829031914), I'm
interested in a uniform interface for accessing the minimum or maximum value of
a given dtype.
Currently, this requires branching on the type of dtype (boolean, integer, or
floating point) and then (for the latter two) calling either
[iinfo](https://numpy.org/doc/stable/reference/generated/numpy.iinfo.html) or
[finfo](https://numpy.org/doc/stable/reference/generated/numpy.finfo.html),
respectively. It would be more ergonomic to have a single, uniform interface
for accessing this information that is dtype-independent.
Possible interfaces include:
```python3
import numpy as np
dt = np.dtype('int32')
dt.min
np.dtypes.info(dt).min
np.dtypes.min(dt)
np.dtypes.min_value(dt)
```
___
NumPy-Discussion mailing list -- numpy-discussion@python.org
To unsubscribe send an email to numpy-discussion-le...@python.org
https://mail.python.org/mailman3/lists/numpy-discussion.python.org/
Member address: arch...@mail-archive.com
[Numpy-discussion] Extend ldexp to handle complex inputs
Currently, [ldexp](https://numpy.org/doc/stable/reference/generated/numpy.ldexp.html) throws a TypeError on a complex input: ```python3 import numpy as np def naive_ldexp(x, n): return x * 2**n def new_ldexp(x, n): if np.iscomplex(x): y = np.empty_like(x) y.real = np.ldexp(x.real, n) y.imag = np.ldexp(x.imag, n) return y else: return np.ldexp(x, n) def main(): x = 2.2 + 3.3j n = 3 print(naive_ldexp(x, n)) print(new_ldexp(x, n)) print(np.ldexp(x, n)) if __name__ == "__main__": main() ``` ``` (17.6+26.4j) (17.6+26.4j) Traceback (most recent call last): File "test.py", line 34, in main() File "test.py", line 30, in main print(np.ldexp(x, n)) ^^ TypeError: ufunc 'ldexp' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe'' ``` This can easily be solved by applying ldexp to the real and imaginary parts of the input, as shown in the function new_ldexp above. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Feature request: Add out-of-place (i.e. pure) versions of all existing in-place operations
Feature request: Add out-of-place (i.e. pure) versions of all existing in-place operations, such as the following: - https://numpy.org/doc/stable/reference/generated/numpy.fill_diagonal.html - https://numpy.org/doc/stable/reference/generated/numpy.put.html - https://numpy.org/doc/stable/reference/generated/numpy.place.html - https://numpy.org/doc/stable/reference/generated/numpy.put_along_axis.html - https://numpy.org/doc/stable/reference/generated/numpy.putmask.html This can be done by adding a boolean argument to each function called "inplace", which is True by default, but when set to False, makes the function return a modified copy of the array. This is the approach taken by JAX for its counterparts of these functions: - https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.fill_diagonal.html - https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.put.html - https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.place.html This would have the following advantages: - Making it easier to perform these operations in a functional way (potentially less error-prone). - Harmonizing the interfaces of NumPy and JAX. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Add diagonal offset argument to all functions that are missing it
The following functions accept a diagonal offset argument: - https://numpy.org/doc/stable/reference/generated/numpy.diag.html - https://numpy.org/doc/stable/reference/generated/numpy.diagflat.html - https://numpy.org/doc/stable/reference/generated/numpy.diagonal.html - https://numpy.org/doc/stable/reference/generated/numpy.eye.html - https://numpy.org/doc/stable/reference/generated/numpy.trace.html - https://numpy.org/doc/stable/reference/generated/numpy.tri.html - https://numpy.org/doc/stable/reference/generated/numpy.tril.html - https://numpy.org/doc/stable/reference/generated/numpy.tril_indices.html - https://numpy.org/doc/stable/reference/generated/numpy.tril_indices_from.html - https://numpy.org/doc/stable/reference/generated/numpy.triu.html - https://numpy.org/doc/stable/reference/generated/numpy.triu_indices.html - https://numpy.org/doc/stable/reference/generated/numpy.triu_indices_from.html The following functions lack such an argument: - https://numpy.org/doc/stable/reference/generated/numpy.diag_indices.html - https://numpy.org/doc/stable/reference/generated/numpy.diag_indices_from.html - https://numpy.org/doc/stable/reference/generated/numpy.fill_diagonal.html - https://github.com/numpy/numpy/issues/14402 - https://github.com/numpy/numpy/issues/18000 - https://github.com/numpy/numpy/pull/15079 - https://numpy.org/doc/stable/reference/generated/numpy.identity.html Feature request: Add a diagonal offset argument to all of the functions that are missing it. (fill_diagonal is already tracked by the aforementioned issues, but the rest aren't, to my knowledge.) ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Add a "broadcast" option to numpy.concatenate
Often, I've wanted to concatenate arrays with different ndims along a
particular axis, broadcasting the other axes as needed. Others have sought this
functionality as well:
- https://stackoverflow.com/questions/56357047
- https://github.com/numpy/numpy/issues/2115
-
https://stackoverflow.com/questions/52733240/concatenate-1d-array-to-a-3d-array
-
https://stackoverflow.com/questions/46700081/concat-two-arrays-of-different-dimensions-numpy
-
https://stackoverflow.com/questions/55879664/how-to-concatenate-a-2d-array-into-every-3d-array
- https://stackoverflow.com/questions/63453495/combining-two-numpy-arrays
However,
[`numpy.concatenate`](https://numpy.org/doc/stable/reference/generated/numpy.concatenate.html)
raises an error if the ndims don't match. For example:
```python3
import numpy as np
a = np.full([2], 0.1)
b = np.full([3, 2], 0.2)
c = np.full([5, 3, 2], 0.3)
arrays = [a, b, c]
axis = -1
try:
np.concatenate(arrays, axis)
except ValueError as e:
print(repr(e))
```
```
ValueError('all the input arrays must have same number of dimensions, but the
array at index 0 has 1 dimension(s) and the array at index 1 has 2
dimension(s)')
```
It would be convenient to add to `numpy.concatenate` an optional boolean
argument called `broadcast` that broadcasts the input arrays along the axes
that are *not* the concatenation axis, before concatenating them. Its default
value can be `False`, which is the current behavior.
Below is an example implementation:
```python3
def tuple_replace(tupl, index, item):
return tupl[:index] + (item,) + tupl[index:][1:]
def broadcast_concat(arrays, axis):
shape = np.broadcast_shapes(*(tuple_replace(a.shape, axis, 0) for a in
arrays))
bcast_arrays = [
np.broadcast_to(a, tuple_replace(shape, axis, a.shape[axis])) for a in
arrays
]
return np.concatenate(bcast_arrays, axis)
output = broadcast_concat(arrays, axis)
assert output.shape[axis] == sum(a.shape[axis] for a in arrays)
```
If desired, I can submit a PR for this.
___
NumPy-Discussion mailing list -- numpy-discussion@python.org
To unsubscribe send an email to numpy-discussion-le...@python.org
https://mail.python.org/mailman3/lists/numpy-discussion.python.org/
Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: Add saturating arithmetic functions
Saturating arithmetic (https://en.wikipedia.org/wiki/Saturation_arithmetic) is important in digital signal processing and other areas. Feature request: Add saturating arithmetic functions for the following basic operations: - addition (C++ counterpart: https://en.cppreference.com/w/cpp/numeric/add_sat) - subtraction (C++ counterpart: https://en.cppreference.com/w/cpp/numeric/sub_sat) - multiplication (C++ counterpart: https://en.cppreference.com/w/cpp/numeric/mul_sat) - division (C++ counterpart: https://en.cppreference.com/w/cpp/numeric/div_sat) - casting (C++ counterpart: https://en.cppreference.com/w/cpp/numeric/saturate_cast) - negation I've implemented these for JAX at https://gist.github.com/carlosgmartin/b32fa6fed3aa82f83dfbaac4b6345672. Corresponding issue: https://github.com/jax-ml/jax/issues/26566. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Add a spectral_radius function
Add a function called `spectral_radius` that computes the [spectral_radius](https://en.wikipedia.org/wiki/Spectral_radius) of a given matrix. A naive way to do this is `np.max(np.abs(np.linalg.eigvals(a)))`, but there are more efficient methods. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Re: Add a bit_width function
Hi Stefan, It should work on unsigned integer types and use exact computation (not floating-point computations like np.log2). For an example, see https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Add a bit_width function
Feature request: Add a `bit_width` function to NumPy's [bit-wise operations](https://numpy.org/doc/stable/reference/routines.bitwise.html) that computes the [bit-width](https://en.wikipedia.org/wiki/Bit-width) (also called bit-length) of an input. For an example, see C++'s [`bit_width`](https://en.cppreference.com/w/cpp/numeric/bit_width) function. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Re: ENH: Let numpy.size accept multiple axes
Correction: [`math.prod`](https://docs.python.org/3/library/math.html#math.prod). ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3//lists/numpy-discussion.python.org Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: Let numpy.size accept multiple axes
Let the `axis` argument of [`numpy.size`](https://numpy.org/doc/stable/reference/generated/numpy.size.html) accept a tuple of ints (like other functions that take an `axis` argument) to measure the size for multiple axes. This should be straightforward to implement with [`itertools.product`](https://docs.python.org/3/library/itertools.html#itertools.product). ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3//lists/numpy-discussion.python.org Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: More user-friendly way to pad a specific axis
It is common to want to pad an array along a *specific* axis. Examples:
-
https://stackoverflow.com/questions/72106542/how-to-pad-the-i-j-axes-of-a-3d-np-array-without-padding-its-k-axis
- https://stackoverflow.com/questions/56076094/zero-pad-ndarray-along-axis
-
https://stackoverflow.com/questions/66225047/padding-a-3-dimensional-numpy-array-with-the-medians-alongside-specific-axis
-
https://stackoverflow.com/questions/19349410/how-to-pad-with-zeros-a-tensor-along-some-axis-python
-
https://stackoverflow.com/questions/74966041/how-to-pad-a-specific-dimension-of-a-numpy-array
Doing so with
[`numpy.pad`](https://numpy.org/doc/stable/reference/generated/numpy.pad.html)
requires constructing a list of pairs of length equal to the ndim of the array,
with exactly one of those pairs at the right position containing the desired
pad widths. This can be verbose and clumsy when there are several axes.
I propose a new, more user-friendly way to pad along a specific axis (or axes):
Let the `pad_width` argument accept a *dictionary* whose keys are axes and
whose values are the `(before, after)` pair (or perhaps single number) for the
corresponding axis. Example:
```python3
# before
np.pad(array, [(0, 0), (0, 0), (1, 2), (0, 0), (0, 0)])
# after
np.pad(array, {-3: (1, 2)})
```
This should require only minor modification to the implementation of
`numpy.pad`. If others like this idea, I can create a PR for it.
___
NumPy-Discussion mailing list -- numpy-discussion@python.org
To unsubscribe send an email to numpy-discussion-le...@python.org
https://mail.python.org/mailman3//lists/numpy-discussion.python.org
Member address: arch...@mail-archive.com
[Numpy-discussion] Re: ENH: Add reverse argument to ufunc.accumulate, cumsum, nancumsum, cumprod, nancumprod
My particular use case was computing cumulative returns in the context of reinforcement learning. A GitHub code search suggests that reverse cumsum is a fairly common operation, with 3.8k hits: https://github.com/search?q=%2Fcumsum%5C%28.*%5C%5B%3A%3A-1%5C%5D%5C%29%5C%5B%3A%3A-1%5C%5D%2F&type=code. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3//lists/numpy-discussion.python.org Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: Add reverse argument to ufunc.accumulate, cumsum, nancumsum, cumprod, nancumprod
Add a reverse argument to accumulating functions that, when true, causes the accumulation to be performed in reverse. Examples: - ufunc.accumulate: https://numpy.org/doc/stable/reference/generated/numpy.ufunc.accumulate.html - cumsum: https://numpy.org/doc/stable/reference/generated/numpy.cumsum.html - nancumsum: https://numpy.org/doc/stable/reference/generated/numpy.nancumsum.html - cumprod: https://numpy.org/doc/stable/reference/generated/numpy.cumprod.html - nancumprod: https://numpy.org/doc/stable/reference/generated/numpy.nancumprod.html Currently, this can be done via two flips (https://numpy.org/doc/stable/reference/generated/numpy.flip.html) before and after the operation. But it might be more efficient to do it directly at the low level. ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3//lists/numpy-discussion.python.org Member address: arch...@mail-archive.com
[Numpy-discussion] ENH: Add a non-circular shift function
Feature request: Add a non-circular shift function that is similar to [`numpy.roll`](https://numpy.org/doc/stable/reference/generated/numpy.roll.html) but, instead of wrapping elements around, replaces unoccupied entries with a specified fill value. Examples: ``` >>> import numpy as np >>> a = np.arange(1, 6) >>> a array([1, 2, 3, 4, 5]) >>> np.shifted(a, 2) array([0, 0, 1, 2, 3]) >>> np.shifted(a, -2) array([3, 4, 5, 0, 0]) >>> a = np.arange(1, 10).reshape((3, 3)) >>> a array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) >>> np.shifted(a, (1, -2), (0, 1)) array([[0, 0, 0], [3, 0, 0], [6, 0, 0]]) ``` Prior discussion: https://github.com/numpy/numpy/issues/26220 Pull request: https://github.com/numpy/numpy/pull/29389 ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3//lists/numpy-discussion.python.org Member address: arch...@mail-archive.com
