在 python pickle 文件中保留 numpy 步幅

Question

有没有办法告诉numpy在写入python泡菜文件时保持非标准步伐？

>>> # Create an array with non-standard striding
>>> x = numpy.arange(2*3*4, dtype='uint8').reshape((2,3,4)).transpose(0,2,1)
>>> x.strides
(12, 1, 4)
>>> # The pickling process converts it to a c-contiguous array.
>>> # Often, this is a good thing, but for some applications, the
>>> # non-standard striding is intentional and important to preserve.
>>> pickled = cPickle.dumps(x, protocol=cPickle.HIGHEST_PROTOCOL)
>>> cPickle.loads(pickled).strides
(12, 3, 1)
>>> # This is indeed happening during serialization, not deserialization
>>> pickletools.dis(pickled)
...
151: S        STRING     'x00x04x08x01x05tx02x06nx03x07x0bx0cx10x14rx11x15x0ex12x16x0fx13x17'
...

注意：numpy 足够聪明，可以保留 c 连续或 fortran-contiguous，但它不会保留酸洗和脱酸的所有非标准步幅模式。

Answer 1

我能想到的唯一方法是自己做：

# ---------------------------------------------
import numpy
from numpy.lib.stride_tricks import as_strided
import cPickle
def dumps(x, protocol=cPickle.HIGHEST_PROTOCOL):
    # flatten that keep real data order
    y = as_strided(x, shape=(x.size,), strides=(min(x.strides),))
    return cPickle.dumps([y,x.shape,x.strides],protocol=protocol)
def loads(pickled):
    y,shape,strides = cPickle.loads(pickled)
    return as_strided(y,shape=shape,strides=strides)
if __name__=='__main__':
    x = numpy.arange(2*3*4, dtype='uint8').reshape((2,3,4)).transpose(0,2,1)
    pickled = dumps(x)
    y = loads(pickled)
    print 'x strides =', x.strides
    print 'y strides =', y.strides
    print 'x==y:', (x==y).all()
# ---------------------------------------------

输出：

x strides = (12, 1, 4)
y strides = (12, 1, 4)
x==y: True