花式索引比 numpy.take 快得多？

Question

我在许多不同的地方读到numpy.take 是一种比花哨索引更快的替代方案，例如here 和here。

但是，我发现情况并非如此……完全没有。以下是我在调试期间查看代码时的示例：

knn_idx
Out[2]: 
array([ 3290,  5847,  7682,  6957, 22660,  5482, 22661, 10965,     7,
        1477,  7681,     3, 17541, 15717,  9139,  1475, 14251,  4400,
        7680,  9140,  4758, 22289,  7679,  8407, 20101, 15718, 15716,
        8405, 15710, 20829, 22662], dtype=uint32)
%timeit X.take(knn_idx, axis=0)
100 loops, best of 3: 3.14 ms per loop
%timeit X[knn_idx]
The slowest run took 60.61 times longer than the fastest. This could mean that an intermediate result is being cached.
100000 loops, best of 3: 5.48 µs per loop
X.shape
Out[5]: 
(23011, 30)
X.dtype
Out[6]: 
dtype('float64')

这表明花式索引要快得多！使用numpy.arange 生成索引我得到了类似的结果：

idx = np.arange(0, len(X), 100)
%timeit X.take(idx, axis=0)
100 loops, best of 3: 3.04 ms per loop
%timeit X[idx]
The slowest run took 9.41 times longer than the fastest. This could mean that an intermediate result is being cached.
10000 loops, best of 3: 20.7 µs per loop

为什么花式索引比现在使用numpy.take 快得多？我遇到了某种极端情况吗？

我通过 Anaconda 使用 Python 3.6，如果相关，这是我的 numpy 信息：

np.__version__
Out[11]: 
'1.11.3'
np.__config__.show()
blas_mkl_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\include']
blas_opt_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\include']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\include']
lapack_opt_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Users/pbreach/Continuum/Anaconda3\\Library\\include']

Answer 1

在我的测试中，take 稍微快一些；但是由于时间短和“缓存”警告，我并没有在差异中投入大量库存：

In [192]: timeit X.take(idx2, axis=0).shape
The slowest run took 23.29 times longer than the fastest. This could mean that an intermediate result is being cached.
100000 loops, best of 3: 3.66 µs per loop
In [193]: timeit X[idx2,:].shape
The slowest run took 16.75 times longer than the fastest. This could mean that an intermediate result is being cached.
100000 loops, best of 3: 5.58 µs per loop

但是你的索引数组是uint32。这对索引很有效，但是 take 给了我一个转换错误；所以我的idx2 是astype(int)。

使用 arange idx，时间分别为 11.5 µs、16 µs。

请注意，我正在使用 .shape 进行计时；我不完全确定这会有所不同。

我不知道您为什么会收到ms 次。感觉更像是时间问题，而不是take 的实际差异。

我不认为图书馆、BLAS 等会有所作为。底层任务基本相同——单步执行数据缓冲区并复制选定的字节。没有复杂的计算可以解决。但是我没有研究过take的C代码。

Numpy 版本“1.12.0”，Linux，4gb 翻新桌面。