python - Dask 数组到 zarr 形状未知

Question

我正在尝试将 dask 数组存储在 zarr 文件中。

当 dask 数组具有定义的形状时，我已经成功地做到了这一点。

import dask
import dask.array as da
import numpy as np
from tempfile import TemporaryDirectory
import zarr


np_array = np.random.randint(1, 10, size=1000)
array = da.from_array(np_array)

with TemporaryDirectory() as tmpdir:
    delayed = da.to_zarr(array, url=tmpdir,
                         compute=False, component='/data')
    dask.compute(delayed)

     z_object = zarr.open_group(tmpdir, mode='r')

     assert np.all(np_array == z_object.data[:])

但是，如果我对 dask 数组执行了任何操作，形状就会丢失，并且 zarr 会提示形状中的 Nans。

# this will fail

np_array = np.random.randint(1, 10, size=1000)
array = da.from_array(np_array)

array = array[array > 5]

with TemporaryDirectory() as tmpdir:
    delayed = da.to_zarr(array, url=tmpdir,
                         compute=False, component='/data')
    dask.compute(delayed)

    z_object = zarr.open_group(tmpdir, mode='r')

    assert np.all(np_array[np_array > 5] == z_object.data[:])

这是引发的错误:

Traceback (most recent call last):
  File "/home/peio/devel/variation/variation6/variation6/tests/test_zarr.py", line 38, in <module>
    without_shape()
  File "/home/peio/devel/variation/variation6/variation6/tests/test_zarr.py", line 29, in without_shape
    compute=False, component='/data')
  File "/home/peio/devel/variation/pyenv3/lib/python3.7/site-packages/dask/array/core.py", line 2808, in to_zarr
    **kwargs
  File "/home/peio/devel/variation/pyenv3/lib/python3.7/site-packages/zarr/creation.py", line 120, in create
    chunk_store=chunk_store, filters=filters, object_codec=object_codec)
  File "/home/peio/devel/variation/pyenv3/lib/python3.7/site-packages/zarr/storage.py", line 323, in init_array
    object_codec=object_codec)
  File "/home/peio/devel/variation/pyenv3/lib/python3.7/site-packages/zarr/storage.py", line 343, in _init_array_metadata
    shape = normalize_shape(shape) + dtype.shape
  File "/home/peio/devel/variation/pyenv3/lib/python3.7/site-packages/zarr/util.py", line 58, in normalize_shape
    shape = tuple(int(s) for s in shape)
  File "/home/peio/devel/variation/pyenv3/lib/python3.7/site-packages/zarr/util.py", line 58, in <genexpr>
    shape = tuple(int(s) for s in shape)
ValueError: cannot convert float NaN to integer

有没有办法将形状未知的 dask 数组存储到 zarr 文件中？

提前致谢!

Answer 1

Zarr 期望 block 的形状是统一的并且是事先已知的。 Dask 目前通过将数组重新分块以使其统一来促进这一点。然而 array[array > 5] 会创建一个具有未知 block 形状的 Dask 数组。因此，由于不存在所需的信息，因此无法预先将其重新分块以使其统一。也就是说，我们可以 explain this better .

可以通过使用返回已知 block 形状的 Dask 操作来解决此问题(正如 David 建议的那样)。或者，可以在存储之前确定 block 形状( at the cost of computing )。我们还可以讨论extending Zarr to handle this case ，但这是一个长期的解决方案。