从字典写入numpy数组

Question

我有一个字典文件头值（时间，帧数，年，月等）的字典，我想将其写入一个numpy数组。我目前拥有的代码如下：

    arr=np.array([(k,)+v for k,v in fileheader.iteritems()],dtype=["a3,a,i4,i4,i4,i4,f8,i4,i4,i4,i4,i4,i4,a10,a26,a33,a235,i4,i4,i4,i4,i4,i4"])

但是我得到一个错误，“只能将元组（而不是“ int”）连接到元组。

基本上，最终结果需要是存储整个文件头信息（512字节）和每个帧的数据（头和数据，每个帧49408字节）的数组。有没有更简单的方法可以做到这一点？

编辑：为了澄清（也为我自己），我需要将数据从文件的每个帧写入数组。我得到了matlab代码作为基础。这是给我的代码的粗略概念：

data.frame=zeros([512 96])
frame=uint8(fread(fid,[data.numbeams,512]),'uint8'))
data.frame=frame

如何将“框架”转换为python？

Answer 1

您最好将标头数据保留在dict中。您是否真的需要将其作为数组？（如果是这样，为什么？将标头包含在numpy数组中有一些优点，但是它比简单的更为复杂dict，并且不够灵活。）

a的一个缺点dict是其键没有可预测的顺序。如果您需要按常规顺序（类似于C结构）将标头写回到磁盘，则需要分别存储字段的顺序及其值。在这种情况下，您可以考虑使用有序的dict（collections.OrderedDict）或仅将一个简单的类放在一起以保存标头数据并在其中存储顺序。

除非有充分的理由将其放入numpy数组，否则您可能不希望这样做。

但是，结构化数组将保留标题的顺序，并使将其二进制表示形式写入磁盘更加容易，但是在其他方面则不灵活。

如果您确实想将标头设置为数组，则可以执行以下操作：

import numpy as np

# Lists can be modified, but preserve order. That's important in this case.
names = ['Name1', 'Name2', 'Name3']
# It's "S3" instead of "a3" for a string field in numpy, by the way
formats = ['S3', 'i4', 'f8']

# It's often cleaner to specify the dtype this way instead of as a giant string
dtype = dict(names=names, formats=formats)

# This won't preserve the order we're specifying things in!!
# If we iterate through it, things may be in any order.
header = dict(Name1='abc', Name2=456, Name3=3.45)

# Therefore, we'll be sure to pass things in in order...
# Also, np.array will expect a tuple instead of a list for a structured array...
values = tuple(header[name] for name in names)
header_array = np.array(values, dtype=dtype)

# We can access field in the array like this...
print header_array['Name2']

# And dump it to disk (similar to a C struct) with
header_array.tofile('test.dat')

另一方面，如果您只想访问标头中的值，则将其保留为dict。这样比较简单。

---

根据听起来您正在做的事情，我会做这样的事情。我使用numpy数组读取标头，但标头值实际上存储为类属性（以及标头数组）。

这看起来比实际要复杂。

我只是在定义两个新类，一个用于父文件，一个用于框架。您可以用更少的代码来完成相同的事情，但这为您完成更复杂的事情奠定了基础。

import numpy as np

class SonarFile(object):
    # These define the format of the file header
    header_fields = ('num_frames', 'name1', 'name2', 'name3')
    header_formats = ('i4', 'f4', 'S10', '>I4')

    def __init__(self, filename):
        self.infile = open(filename, 'r')
        dtype = dict(names=self.header_fields, formats=self.header_formats)

        # Read in the header as a numpy array (count=1 is important here!)
        self.header = np.fromfile(self.infile, dtype=dtype, count=1)

        # Store the position so we can "rewind" to the end of the header
        self.header_length = self.infile.tell()

        # You may or may not want to do this (If the field names can have
        # spaces, it's a bad idea). It will allow you to access things with
        # sonar_file.Name1 instead of sonar_file.header['Name1'], though.
        for field in self.header_fields:
            setattr(self, field, self.header[field])

    # __iter__ is a special function that defines what should happen when we  
    # try to iterate through an instance of this class.
    def __iter__(self):
        """Iterate through each frame in the dataset."""
        # Rewind to the end of the file header
        self.infile.seek(self.header_length)

        # Iterate through frames...
        for _ in range(self.num_frames):
            yield Frame(self.infile)

    def close(self):
        self.infile.close()

class Frame(object):
    header_fields = ('width', 'height', 'name')
    header_formats = ('i4', 'i4', 'S20')
    data_format = 'f4'

    def __init__(self, infile):
        dtype = dict(names=self.header_fields, formats=self.header_formats)
        self.header = np.fromfile(infile, dtype=dtype, count=1)

        # See discussion above...
        for field in self.header_fields:
            setattr(self, field, self.header[field])

        # I'm assuming that the size of the frame is in the frame header...
        ncols, nrows = self.width, self.height

        # Read the data in
        self.data = np.fromfile(infile, self.data_format, count=ncols * nrows)

        # And reshape it into a 2d array.
        # I'm assuming C-order, instead of Fortran order.
        # If it's fortran order, just do "data.reshape((ncols, nrows)).T"
        self.data = self.data.reshape((nrows, ncols))

您将使用类似于以下内容的方法：

dataset = SonarFile('input.dat')

for frame in dataset:
    im = frame.data
    # Do something...