python - 查询保存为 npz 的 NumPy 数组的 NumPy 数组很慢

Question

我生成一个npz文件如下:

import numpy as np
import os

# Generate npz file
dataset_text_filepath = 'test_np_load.npz'
texts = []
for text_number in range(30000): 
    texts.append(np.random.random_integers(0, 20000, 
                 size = np.random.random_integers(0, 100)))
texts = np.array(texts)
np.savez(dataset_text_filepath, texts=texts)

这给了我这个 ~7MiB npz 文件(基本上只有 1 个变量 texts，它是 Numpy 数组的 NumPy 数组):

我使用 numpy.load() 加载:

# Load data
dataset = np.load(dataset_text_filepath)

如果我这样查询，需要几分钟:

# Querying data: the slow way
for i in range(20):
    print('Run {0}'.format(i))
    random_indices = np.random.randint(0, len(dataset['texts']), size=10)
    dataset['texts'][random_indices]

而如果我如下查询，它只需要不到 5 秒:

# Querying data: the fast way
data_texts = dataset['texts']
for i in range(20):
    print('Run {0}'.format(i))
    random_indices = np.random.randint(0, len(data_texts), size=10)
    data_texts[random_indices]

为什么第二种方法比第一种方法快这么多？

Answer 1

dataset['texts'] 在每次使用时读取文件。 load npz 只返回一个文件加载器，而不是实际数据。它是一个“惰性加载程序”，仅在访问时加载特定数组。 load 文档可能更清晰，但他们说:

- If the file is a ``.npz`` file, the returned value supports the context
  manager protocol in a similar fashion to the open function::

    with load('foo.npz') as data:
        a = data['a']

  The underlying file descriptor is closed when exiting the 'with' block.

从savez:

 When opening the saved ``.npz`` file with `load` a `NpzFile` object is
returned. This is a dictionary-like object which can be queried for
its list of arrays (with the ``.files`` attribute), and for the arrays
themselves.

help(np.lib.npyio.NpzFile)中有更多详细信息