Draft JOSS article about SciPy elementwise iterative methods #116
Conversation
paper.md
Outdated
| - `scipy.integrate.nsum` for summation of real-valued finite or infinite series, | ||
| - `scipy.optimize.elementwise` for finding roots and minimization of a single-input, single output function. | ||
|
|
||
| Although the details of the algorithms are inherently distinct, of these features rely on a common framework for elementwise iterative methods and share similar interfaces. For example, `scipy.differentiate.derivative` accepts the scalar function to be differentiated as a Python callable, the points at which to evaluate the derivative as an N-dimensional array, and any other parameters of the callable as N-dimensional arrays. `derivative` passes abscissae and parameters to the callable as N-dimensional arrays, adaptively approximating the derivative(s) to achieve specified tolerances. Furthermore, arrays need not be NumPy arrays: the same implementation works with several Python Array API Standard compatible arrays, such as PyTorch tensors and CuPy arrays. These features will dramatically improve performance of end-user applications, and together, they form the backbone of SciPy's new random variable infrastructure. |
There was a problem hiding this comment.
| Although the details of the algorithms are inherently distinct, of these features rely on a common framework for elementwise iterative methods and share similar interfaces. For example, `scipy.differentiate.derivative` accepts the scalar function to be differentiated as a Python callable, the points at which to evaluate the derivative as an N-dimensional array, and any other parameters of the callable as N-dimensional arrays. `derivative` passes abscissae and parameters to the callable as N-dimensional arrays, adaptively approximating the derivative(s) to achieve specified tolerances. Furthermore, arrays need not be NumPy arrays: the same implementation works with several Python Array API Standard compatible arrays, such as PyTorch tensors and CuPy arrays. These features will dramatically improve performance of end-user applications, and together, they form the backbone of SciPy's new random variable infrastructure. | |
| Although the details of the algorithms are inherently distinct, these features rely on a common framework for elementwise iterative methods and share similar interfaces. For example, `scipy.differentiate.derivative` accepts the scalar function to be differentiated as a Python callable, the points at which to evaluate the derivative as an N-dimensional array, and any other parameters of the callable as N-dimensional arrays. `derivative` passes abscissae and parameters to the callable as N-dimensional arrays, adaptively approximating the derivative(s) to achieve specified tolerances. Furthermore, arrays need not be NumPy arrays: the same implementation works with several Python Array API Standard compatible arrays, such as PyTorch tensors and CuPy arrays. These features will dramatically improve performance of end-user applications, and together, they form the backbone of SciPy's new random variable infrastructure. |
There was a problem hiding this comment.
Otherwise I am not sure I would go into the description of the API. The text can be difficult to read so consider adding the general API you are describing (e.g. f(a, b, c).)
There was a problem hiding this comment.
For Array API, maybe we should point to the standard as a ref.
There was a problem hiding this comment.
Removed "of" and added reference to array API standard.
I would consider removing:
For example,
scipy.differentiate.derivativeaccepts the scalar function to be differentiated as a Python callable, the points at which to evaluate the derivative as an N-dimensional array, and any other parameters of the callable as N-dimensional arrays.derivativepasses abscissae and parameters to the callable as N-dimensional arrays, adaptively approximating the derivative(s) to achieve specified tolerances.
Let's see what @steppi thinks.
paper.md
Outdated
| - Speed: the features take full advantage of vectorized user callables, avoiding slow Python loops and the excessive overhead of repeatedly calling compiled code. | ||
| - Prevalence: SciPy is one of the most popular scientific Python packages. If a scientific Python user needs these features, chances are that they already have SciPy installed, eliminating the need to find and learn a new package. | ||
| - Easy-of-use: the function API reference is thorough, and the interfaces share common features and operate smoothly with other SciPy functions. | ||
| - Dependability: as with all SciPy code, these features were designed and implemented following good software development practices. They have been carefully peer-reviewed, and extensive unit tests protect against backward incompatible changes and regressions. |
There was a problem hiding this comment.
I would add a note on the scientific accuracy. Here it might read that the code is good and correct on the software engineering side, not necessarily that the math is sound.
There was a problem hiding this comment.
These are supposed to be features that set SciPy apart from one or more of the other options. I assume all of them are mathematically correct. One could argue that ease-of-use and dependability are common to all, to some extent, so I think I'd be more likely to remove those than to add mathematical accuracy as a "unique" advantage.
mdhaber
force-pushed
the
joss_eim_draft
branch
3 times, most recently
from
8b04539 to
f509f2b
Compare
mdhaber
force-pushed
the
joss_eim_draft
branch
from
f509f2b to
e4f43bf
Compare
|
Oops. I accidentally reset to |
Draft JOSS article about SciPy elementwise iterative methods