diditforlulz273/fastauc

Three fast ROC AUC calculation implementations for python

fastauc

Three fast AUC calculation implementations for python:

python-based is approximately 8X faster than the default sklearn.metrics.roc_auc_score()

Python numba based is approximately 26X faster than the default sklearn.metrics.roc_auc_score()

C++/ctypes based is approximately 37X faster than the default sklearn.metrics.roc_auc_score()

10 000 AUC calculations total time:

NB: run cd fastauc && ./compile.sh to create a binary before you use a C++ version.

See demo.py for speed benchmark and usage examples.

fastauc's Issues

Moving np.argsort out of numba-compiled func.

Hi, please consider moving np.argsort out of numba-compiled func, as njitting argsort is known for slowing it down twice (still as of 2023, unfortunately). That makes Numba version faster than C version :-)

smth like

def optimized_auc(y_true: np.array, y_score: np.array) -> float:
    desc_score_indices = np.argsort(y_score,)[::-1]
    return fast_numba_auc_nonw(y_true=y_true, y_score=y_score, desc_score_indices=desc_score_indices)

My tests on 8M float32 array:

%timeit fast_auc.fast_auc(y_true=Y_val.values, y_score=probs[:, 1])
392 ms ± 2.79 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit fast_auc.fast_numba_auc(y_true=Y_val.values, y_score=probs[:, 1])
672 ms ± 2.28 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit optimized_auc(y_true=Y_val.values, y_score=probs[:, 1])
338 ms ± 3.07 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

No need to hardcode mergesort, as well.

Would you please add a license? Thank you!