Saving a scipy.sparse matrix directly as a regular txt file

I have a scipy.sparse matrix (csr_matrix()). But I need to save it to a file not in the .npz format but as a regular .txt or .csv file. My problem is that I don't have enough memory to convert the sparse matrix into a regular np.array() and then save it to a file. Is there a way to have the data as a sparse matrix in memory but save it directly as a regular matrix in the form:

0 0 0
0 1 0
1 0 1

to the disk? Or is there a way to "unzip" a .npz file without loading it into memory inside Python? (like for example gunzip or unzip in Bash).

Solution

Answer to new question:

import numpy as np
from scipy import sparse, io
A = sparse.eye(5, format='csr') * np.pi
np.set_printoptions(precision=16, linewidth=1000)
with open('matrix.txt', 'a') as f:
    for row in A:
        f.write(str(row.toarray()[0]))
        f.write('\n')

# [3.141592653589793 0.                0.                0.                0.               ]
# [0.                3.141592653589793 0.                0.                0.               ]
# [0.                0.                3.141592653589793 0.                0.               ]
# [0.                0.                0.                3.141592653589793 0.               ]
# [0.                0.                0.                0.                3.141592653589793]

And with begin/end brackets:

import numpy as np
from scipy import sparse, io
A = sparse.eye(5, format='csr') * np.pi
np.set_printoptions(precision=16, linewidth=1000)
with open('matrix.txt', 'a') as f:
    for i, row in enumerate(A):
        f.write('[' if (i == 0) else ' ')
        f.write(str(row.toarray()[0]))
        f.write(']' if (i == A.shape[0] - 1) else '\n')

# [[3.141592653589793 0.                0.                0.                0.               ]
#  [0.                3.141592653589793 0.                0.                0.               ]
#  [0.                0.                3.141592653589793 0.                0.               ]
#  [0.                0.                0.                3.141592653589793 0.               ]
#  [0.                0.                0.                0.                3.141592653589793]]

You may have to fiddle with set_printoptions depending on your data.

Answer to original question, which did not require that the matrix be written as dense.

Harwell-Boeing format is plain text:

import numpy as np
from scipy import sparse, io
A = sparse.eye(3, format='csr') * np.pi

# Default title                                                           0       
#              3             1             1             1
# RUA                        3             3             3             0
# (40I2)          (40I2)          (3E25.16)           
#  1 2 3 4
#  1 2 3
#   3.1415926535897931E+00  3.1415926535897931E+00  3.1415926535897931E+00

io.hb_write('matrix.txt', A)  # saves as matrix.txt
A2 = io.hb_read('matrix.txt')
assert not (A2 != A).nnz  # efficient check for equality

So is Matrix Market:

io.mmwrite('matrix', A)  # saves as matrix.mtx

# %%MatrixMarket matrix coordinate real symmetric
# %
# 3 3 3
# 1 1 3.141592653589793e+00
# 2 2 3.141592653589793e+00
# 3 3 3.141592653589793e+00

A2 = io.mmread('matrix')
assert not (A2 != A).nnz

If you want an even simpler format, although it involves more code:

import numpy as np
from scipy import sparse
A = sparse.eye(10, format='csr')*np.pi

np.savetxt('data.txt', A.data)
np.savetxt('indices.txt', A.indices, fmt='%i')
np.savetxt('indptr.txt', A.indptr, fmt='%i')

To load:

data = np.loadtxt('data.txt')
indices = np.loadtxt('indices.txt', dtype=np.int32)
indptr = np.loadtxt('indptr.txt', dtype=np.int32)

A2 = sparse.csr_matrix((data, indices, indptr))
assert not (A2 != A).nnz

But the important idea is that all you need to save are the data, indices, and indptr attributes of the csr_matrix.