Pairwise distances: first implementation #2
Conversation
mbatoul
changed the title
mbatoul
changed the title
There was a problem hiding this comment.
That's a nice first start! 👏
The generated .c file is still present. You might want to erase it from your local repo cache using something like:
git rm --cached **/*.cI left some comment, but in overall it's good: the Python API is concise and the implementation is done in cdef functions which should map to C code with no much interaction with the CPython interpreter.
Looking forward to your future iterations. :)
src/pairwise_dist/_pairwise_dist.pyx
Outdated
|
|
||
| for i in prange(n_rows_X_a, nogil=True): | ||
| for j in range(n_rows_X_b): | ||
| _compute_dist(X_a[i], X_b[j], n_features, &distances[i, j]) |
There was a problem hiding this comment.
in this implementation passing a pointer for distances[i, j] seems unecessary. You could just do
distances[i, j] = _compute(dist(X_a[i], X_b[j], n_features))
There was a problem hiding this comment.
See this issue scikit-learn/scikit-learn#17299. It looks like a good reproducer for it. Is it faster to do as explained in the issue ?
There was a problem hiding this comment.
It seems faster indeed! Some quick benchmarks:
We will see how it improves performance compared to scipy and scikit-learn.
No description provided.