numpy.memmap map to save file

Question

I'm trying to create random matrix and save it in binary file using numpy.save

Then I try to map this file using numpy.memmap, but it seems it maps it wrong.

How to fix it?

It seems it read .npy header and I need to scip some bytes from begining.

rows=6
cols=4

def create_matrix(rows,cols):
    data = (np.random.rand(rows,cols)*100).astype('uint8') #type for image [0 255] int8?
    return data

def save_matrix(filename, data):
    np.save(filename, data)

def load_matrix(filename):
    data= np.load(filename)
    return data

def test_mult_ram():
    A= create_matrix(rows,cols)
    A[1][2]= 42
    save_matrix("A.npy", A)
    A= load_matrix("A.npy")
    print A
    B= create_matrix(cols,rows)
    save_matrix("B.npy", B)
    B= load_matrix("B.npy")
    print B




fA = np.memmap('A.npy', dtype='uint8', mode='r', shape=(rows,cols))
fB = np.memmap('B.npy', dtype='uint8', mode='r', shape=(cols,rows))
print fA
print fB

UPDATE:

I just found that already np.lib.format.open_memmap function exist.

usage: a = np.lib.format.open_memmap('A.npy', dtype='uint8', mode='r+')

Answer 1

The npy format has a header that must be skipped when using np.memmap. It starts with an 6-byte magic string, '\x93NUMPY', 2 byte version number, followed by 2 bytes header length, followed by header data.

So if you open the file, find the header length, then you can compute the offset to pass to np.memmap:

def load_npy_to_memmap(filename, dtype, shape):
    # npy format is documented here
    # https://github.com/numpy/numpy/blob/master/doc/neps/npy-format.txt
    with open(filename, 'r') as f:
        # skip magic string \x93NUMPY + 2 bytes major/minor version number
        # + 2 bytes little-endian unsigned short int
        junk, header_len = struct.unpack('<8sh', f.read(10))

    data= np.memmap(filename, dtype=dtype, shape=shape, offset=6+2+2+header_len)
    return data

---

import struct
import numpy as np
np.random.seed(1)
rows = 6
cols = 4

def create_matrix(rows, cols):
    data = (np.random.rand(
        rows, cols) * 100).astype('uint8')  # type for image [0 255] int8?
    return data

def save_matrix(filename, data):
    np.save(filename, data)

def load_matrix(filename):
    data= np.load(filename)
    return data

def load_npy_to_memmap(filename, dtype, shape):
    # npy format is documented here
    # https://github.com/numpy/numpy/blob/master/doc/neps/npy-format.txt
    with open(filename, 'r') as f:
        # skip magic string \x93NUMPY + 2 bytes major/minor version number
        # + 2 bytes little-endian unsigned short int
        junk, header_len = struct.unpack('<8sh', f.read(10))

    data= np.memmap(filename, dtype=dtype, shape=shape, offset=6+2+2+header_len)
    return data

def test_mult_ram():
    A = create_matrix(rows, cols)
    A[1][2] = 42
    save_matrix("A.npy", A)
    A = load_matrix("A.npy")
    print A
    B = create_matrix(cols, rows)
    save_matrix("B.npy", B)
    B = load_matrix("B.npy")
    print B

    fA = load_npy_to_memmap('A.npy', dtype='uint8', shape=(rows, cols))
    fB = load_npy_to_memmap('B.npy', dtype='uint8', shape=(cols, rows))
    print fA
    print fB
    np.testing.assert_equal(A, fA)
    np.testing.assert_equal(B, fB)

test_mult_ram()

Answer 2

If your goal is to open arrays you saved with np.save as memmaps, then you can just use np.load with the option mmap_mode:

fA = np.load('A.npy', mmap_mode='r')
fB = np.load('B.npy', mmap_mode='r')

This way you actually benefit from the header stored in the .npy files, in the sense that it keeps track of the shape and dtype of the array.