如何从延迟的 dask 数组中获取特定块？

Question

假设我有以下数组：import numpy as npimport dask.array as daimport daskarr_x = list(range(0,100))arr_y = list(range(0,100))arr = np.stack([arr_x,arr_y])arr = arr.T然后我想将其转换为延迟的 dask 数组：arr = da.from_array(arr,chunks = (3,2))data = arr.to_delayed()其格式如下：[[Delayed(('array-fa3499f6a402676a68a198bef8016ec4', 0, 0))] [Delayed(('array-fa3499f6a402676a68a198bef8016ec4', 1, 0))] [Delayed(('array-fa3499f6a402676a68a198bef8016ec4', 2, 0))]... [Delayed(('array-fa3499f6a402676a68a198bef8016ec4', 31, 0))] [Delayed(('array-fa3499f6a402676a68a198bef8016ec4', 32, 0))] [Delayed(('array-fa3499f6a402676a68a198bef8016ec4', 33, 0))]]现在我想获得一个特定的块：chunk = da.from_delayed(data[1], shape=(3,2))print(chunk.compute())但是，我收到以下错误：dsk = {(name,) + (0,) * len(shape): value.key}AttributeError: 'numpy.ndarray' object has no attribute 'key'我究竟做错了什么？

Answer 1

dask.array.Array.to_delayed()返回二维数组的列表列表。您需要一直切片到延迟对象才能将其传递给 from_delayed()

In [5]: chunk = da.from_delayed(data[1][0], shape=(3,2), dtype=arr.dtype)

In [6]: chunk.compute()
Out[6]:
array([[3, 3],
       [4, 4],
       [5, 5]])

这也可以写成

In [11]: arr.blocks[1, 0].compute()
Out[11]:
array([[3, 3],
       [4, 4],
       [5, 5]])