Create Numpy .npy
files by appending on the growth axis (0 for C order, -1
for Fortran order). It behaves like numpy.concatenate
with the difference
that the result is stored out-of-memory in a .npy
file and can be reused for
further appending. After creation, the file can then be read with memory
mapping (e.g. by adding mmap_mode="r"
) which altogether allows to create and
read files (optionally) larger than the machine's main memory.
Some possible applications:
- efficiently create large
.npy
(optionally database-like) files- Handling of offsets not included, can be done in an extra array
- Large legacy files can be made appendable by calling
ensure_appendable
- can (optionally) be performed in-place to minimize disk space usage
- create binary log files (optionally on low-memory embedded devices)
- Check the option
rewrite_header_on_append=False
for extra efficiency - Binary log files can be accessed very efficiently without parsing
- Incomplete files can be recovered efficiently by calling
recover
- Check the option
Another feature of this library is the (above mentioned) recover
function,
which makes incomplete .npy
files readable by numpy.load
again, no matter
whether they should be appended to or not.
Incomplete files can be the result of broken downloads or unfinished writes.
Recovery works by rewriting the header and inferring the growth axis (see
above) by the file size. As the data length may not be evenly divisible by the
non-append-axis shape, incomplete entries can either be ignored
(zerofill_incomplete=False
), which probably makes sense in most scenarios.
Alternatively, to squeeze out the as much information from the file as
possible, zerofill_incomplete=True
can be used, which fills the incomplete
last append axis item with zeros.
Raises ValueError
instead of TypeError
since version 0.9.14 to be more
consistent with Numpy.
NpyAppendArray can be used in multithreaded environments.
conda install -c conda-forge npy-append-array
or
pip install npy-append-array
from npy_append_array import NpyAppendArray
import numpy as np
arr1 = np.array([[1,2],[3,4]])
arr2 = np.array([[1,2],[3,4],[5,6]])
filename = 'out.npy'
with NpyAppendArray(filename, delete_if_exists=True) as npaa:
npaa.append(arr1)
npaa.append(arr2)
npaa.append(arr2)
data = np.load(filename, mmap_mode="r")
print(data)
Concurrency can be achieved by multithreading: A single NpyAppendArray
object (per file) needs to be created. Then, append
can be called from
multiple threads and locks will ensure that file writes do not happen in
parallel. When using with a with
statement, make sure the join
happens
within it, compare test.py
.
Multithreaded writes are not the pinnacle of what is technically possible with
modern operating systems. It would be highly desirable to use async
file
writes. However, although modules like aiofile
exist, this is currently not
supported natively by Python or Numpy, compare
NpyAppendArray contains a modified, partial version of format.py
from the
Numpy package. It ensures that array headers are created with 21
(=len(str(8*2**64-1))
) bytes of spare space. This allows to fit an array of
maxed out dimensions (for a 64 bit machine) without increasing the array
header size. This allows to simply rewrite the header as we append data to the
end of the .npy
file.
Tested with Ubuntu Linux, macOS and Windows.