根据denfromufa's link,我认为Robert McLeod提供了最好的解决方案。他还指出使用np.frombuffer
的缺点:
虽然可以使用np.frombuffer进行零拷贝,但这样会导致Python垃圾回收器和C#垃圾回收器都管理内存,使得内存混乱。
以下是Robert McLeod在Github问题中的代码片段:
import numpy as np
import ctypes
import clr, System
from System import Array, Int32
from System.Runtime.InteropServices import GCHandle, GCHandleType
_MAP_NP_NET = {
np.dtype('float32'): System.Single,
np.dtype('float64'): System.Double,
np.dtype('int8') : System.SByte,
np.dtype('int16') : System.Int16,
np.dtype('int32') : System.Int32,
np.dtype('int64') : System.Int64,
np.dtype('uint8') : System.Byte,
np.dtype('uint16') : System.UInt16,
np.dtype('uint32') : System.UInt32,
np.dtype('uint64') : System.UInt64,
np.dtype('bool') : System.Boolean,
}
_MAP_NET_NP = {
'Single' : np.dtype('float32'),
'Double' : np.dtype('float64'),
'SByte' : np.dtype('int8'),
'Int16' : np.dtype('int16'),
'Int32' : np.dtype('int32'),
'Int64' : np.dtype('int64'),
'Byte' : np.dtype('uint8'),
'UInt16' : np.dtype('uint16'),
'UInt32' : np.dtype('uint32'),
'UInt64' : np.dtype('uint64'),
'Boolean': np.dtype('bool'),
}
def asNumpyArray(netArray):
'''
Given a CLR `System.Array` returns a `numpy.ndarray`. See _MAP_NET_NP for
the mapping of CLR types to Numpy dtypes.
'''
dims = np.empty(netArray.Rank, dtype=int)
for I in range(netArray.Rank):
dims[I] = netArray.GetLength(I)
netType = netArray.GetType().GetElementType().Name
try:
npArray = np.empty(dims, order='C', dtype=_MAP_NET_NP[netType])
except KeyError:
raise NotImplementedError("asNumpyArray does not yet support System type {}".format(netType) )
try:
sourceHandle = GCHandle.Alloc(netArray, GCHandleType.Pinned)
sourcePtr = sourceHandle.AddrOfPinnedObject().ToInt64()
destPtr = npArray.__array_interface__['data'][0]
ctypes.memmove(destPtr, sourcePtr, npArray.nbytes)
finally:
if sourceHandle.IsAllocated: sourceHandle.Free()
return npArray
def asNetArray(npArray):
'''
Given a `numpy.ndarray` returns a CLR `System.Array`. See _MAP_NP_NET for
the mapping of Numpy dtypes to CLR types.
Note: `complex64` and `complex128` arrays are converted to `float32`
and `float64` arrays respectively with shape [m,n,...] -> [m,n,...,2]
'''
dims = npArray.shape
dtype = npArray.dtype
if dtype == np.complex64:
dtype = np.dtype('float32')
dims.append(2)
npArray = npArray.view(np.float32).reshape(dims)
elif dtype == np.complex128:
dtype = np.dtype('float64')
dims.append(2)
npArray = npArray.view(np.float64).reshape(dims)
netDims = Array.CreateInstance(Int32, npArray.ndim)
for I in range(npArray.ndim):
netDims[I] = Int32(dims[I])
if not npArray.flags.c_contiguous:
npArray = npArray.copy(order='C')
assert npArray.flags.c_contiguous
try:
netArray = Array.CreateInstance(_MAP_NP_NET[dtype], netDims)
except KeyError:
raise NotImplementedError("asNetArray does not yet support dtype {}".format(dtype))
try:
destHandle = GCHandle.Alloc(netArray, GCHandleType.Pinned)
sourcePtr = npArray.__array_interface__['data'][0]
destPtr = destHandle.AddrOfPinnedObject().ToInt64()
ctypes.memmove(destPtr, sourcePtr, npArray.nbytes)
finally:
if destHandle.IsAllocated: destHandle.Free()
return netArray
if __name__ == '__main__':
from time import perf_counter
import matplotlib.pyplot as plt
import psutil
tries = 1000
foo = np.full([1024,1024], 2.5, dtype='float32')
netMem = np.zeros(tries)
t_asNet = np.zeros(tries)
netFoo = asNetArray( foo )
for I in range(tries):
t0 = perf_counter()
netFoo = asNetArray( foo )
t_asNet[I] = perf_counter() - t0
netMem[I] = psutil.virtual_memory().free / 2.0**20
t_asNumpy = np.zeros(tries)
numpyMem = np.zeros(tries)
unNetFoo = asNumpyArray( netFoo )
for I in range(tries):
t0 = perf_counter()
unNetFoo = asNumpyArray( netFoo )
t_asNumpy[I] = perf_counter() - t0
numpyMem[I] = psutil.virtual_memory().free / 2.0**20
t_asNet *= 1000
t_asNumpy *= 1000
np.testing.assert_array_almost_equal( unNetFoo, foo )
print( "Numpy to .NET converted {} bytes in {:.3f} +/- {:.3f} ms (mean: {:.1f} ns/ele)".format( \
foo.nbytes, t_asNet.mean(), t_asNet.std(), t_asNet.mean()/foo.size*1e6 ) )
print( ".NET to Numpy converted {} bytes in {:.3f} +/- {:.3f} ms (mean: {:.1f} ns/ele)".format( \
foo.nbytes, t_asNumpy.mean(), t_asNumpy.std(), t_asNumpy.mean()/foo.size*1e6 ) )
plt.figure()
plt.plot(np.arange(tries), netMem, '-', label='asNetArray')
plt.plot(np.arange(tries), numpyMem, '-', label='asNumpyArray')
plt.legend(loc='best')
plt.ylabel('Free memory (MB)')
plt.xlabel('Iteration')
plt.show(block=True)
值得一提的是,pythonnet有一个新的实验性功能,看起来很有前途:
编解码器。只有在构建源代码并成功理解文档时才相关:
numpy.array(TwoDArray)
吗? - Nils Wernerarray(<System.Single[,] object at 0x0000000011501438>, dtype=object)
。 - rbp109