python - 将numpy数组保存到文件（尽可能小的大小）

Question

Right now I have a python program building a fairly large 2D numpy array and saving it as a tab delimited text file using numpy.savetxt. The numpy array contains only floats. I then read the file in one row at a time in a separate C++ program.

What I would like to do is find a way to accomplish this same task, changing my code as little as possible such that I can decrease the size of the file I am passing between the two programs.

I found that I can use numpy.savetxt to save to a compressed .gz file instead of a text file. This lowers the file size from ~2MB to ~100kB.

Is there a better way to do this? Could I, perhaps, write the numpy array in binary to the file to save space? If so, how would I do this so that I can still read it into the C++ program?

Thank you for the help. I appreciate any guidance I can get.

EDIT:

There are a lot of zeros (probably 70% of the values in the numpy array are 0.0000) I am not sure of how I can somehow exploit this though and generate a tiny file that my c++ program can read in

Answer 1

Unless you are sure you don't need to worry about endianness and such, best use numpy.savez , as explained in @unutbu's answer and @jorgeca's comment here: numpy's tostring/fromstring --- what do I need to specify to restore the array .

If the resulting size is not small enough, there's always zlib (on python's side: import zlib , on the C++ side, I'm sure an implementation exists).

An alternative would be to use hdf5 format: while it does not necessarily reduce the on-disk file size, it does make saving/loading faster (this is what the format was designed for, large data arrays). There are both python and C++ readers/writers for hdf5 .

Answer 2

Since you have a lot of zeroes, you could only write out the non-zero elements in the form (index, number).

Suppose you have an array with a small amount of nonzero numbers:

In [5]: a = np.zeros((10, 10))

In [6]: a
Out[6]: 
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]])

In [7]: a[3,1] = 2.0

In [8]: a[7,4] = 17.0

In [9]: a[9,0] = 1.5

First, isolate the interesting numbers and their indices:

In [11]: x, y = a.nonzero()

In [12]: zip(x,y)
Out[12]: [(3, 1), (7, 4), (9, 0)]

In [13]: nonzero = zip(x,y)

Now you only have a small number of data elements left. The easiest thing is to write them to a text file:

In [17]: with open('numbers.txt', 'w+') as outf:
   ....:     for r, k in nonzero:
   ....:         outf.write('{:d} {:d} {:g}\n'.format(r, k, a[r,k]))
   ....:         

In [18]: cat numbers.txt
3 1 2
7 4 17
9 0 1.5

This also gives you an opportunity to eyeball the data. In your C++ program you can read this data with fscanf .

But you can reduce the size even more by writing binary data using struct :

In [17]: import struct

In [19]: c = struct.Struct('=IId')

In [20]: with open('numbers.bin', 'w+') as outf:
   ....:     for r, k in nonzero:
   ....:         outf.write(c.pack(r, k, a[r,k]))

The argument to the Struct constructor means; use native date format '='. The first and second data elements are unsigned integers 'I', the third element is a double 'd'.

In your C++ program this data is probably best read as binary data into a packed struct .

EDIT : Answer updated for a 2D array.

Answer 3

numpy.ndarray.tofile and numpy.fromfile are useful for direct binary output/input from python. std::ostream::write std::istream::read are useful for binary output/input in c++.

You should be careful about endianess if the data are transferred from one machine to another.

Answer 4

Use the an hdf5 file, they are really simple to use through h5py and you can use set compression a flag. Note that hdf5 has also a c++ interface.

Answer 5

如果您不介意安装其他软件包（对于python和c++ ），您可以使用[BSON][1] （二进制JSON）。