Skip to content
This repository
Martin Spacek
file 294 lines (218 sloc) 12.163 kb
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
Title: A Simple File Format for NumPy Arrays
Version: $Revision$
Last-Modified: $Date$
Author: Robert Kern <>
Status: Draft
Type: Standards Track
Content-Type: text/plain
Created: 20-Dec-2007


    We propose a standard binary file format (NPY) for persisting
    a single arbitrary NumPy array on disk. The format stores all of
    the shape and dtype information necessary to reconstruct the array
    correctly even on another machine with a different architecture.
    The format is designed to be as simple as possible while achieving
    its limited goals. The implementation is intended to be pure
    Python and distributed as part of the main numpy package.


    A lightweight, omnipresent system for saving NumPy arrays to disk
    is a frequent need. Python in general has pickle [1] for saving
    most Python objects to disk. This often works well enough with
    NumPy arrays for many purposes, but it has a few drawbacks:

    - Dumping or loading a pickle file require the duplication of the
      data in memory. For large arrays, this can be a showstopper.

    - The array data is not directly accessible through
      memory-mapping. Now that numpy has that capability, it has
      proved very useful for loading large amounts of data (or more to
      the point: avoiding loading large amounts of data when you only
      need a small part).

    Both of these problems can be addressed by dumping the raw bytes
    to disk using ndarray.tofile() and numpy.fromfile(). However,
    these have their own problems:

    - The data which is written has no information about the shape or
      dtype of the array.

    - It is incapable of handling object arrays.

    The NPY file format is an evolutionary advance over these two
    approaches. Its design is mostly limited to solving the problems
    with pickles and tofile()/fromfile(). It does not intend to solve
    more complicated problems for which more complicated formats like
    HDF5 [2] are a better solution.

Use Cases

    - Neville Newbie has just started to pick up Python and NumPy. He
      has not installed many packages, yet, nor learned the standard
      library, but he has been playing with NumPy at the interactive
      prompt to do small tasks. He gets a result that he wants to

    - Annie Analyst has been using large nested record arrays to
      represent her statistical data. She wants to convince her
      R-using colleague, David Doubter, that Python and NumPy are
      awesome by sending him her analysis code and data. She needs
      the data to load at interactive speeds. Since David does not
      use Python usually, needing to install large packages would turn
      him off.

    - Simon Seismologist is developing new seismic processing tools.
      One of his algorithms requires large amounts of intermediate
      data to be written to disk. The data does not really fit into
      the industry-standard SEG-Y schema, but he already has a nice
      record-array dtype for using it internally.

    - Polly Parallel wants to split up a computation on her multicore
      machine as simply as possible. Parts of the computation can be
      split up among different processes without any communication
      between processes; they just need to fill in the appropriate
      portion of a large array with their results. Having several
      child processes memory-mapping a common array is a good way to
      achieve this.


    The format MUST be able to:

    - Represent all NumPy arrays including nested record
      arrays and object arrays.

    - Represent the data in its native binary form.

    - Be contained in a single file.

    - Support Fortran-contiguous arrays directly.

    - Store all of the necessary information to reconstruct the array
      including shape and dtype on a machine of a different
      architecture. Both little-endian and big-endian arrays must be
      supported and a file with little-endian numbers will yield
      a little-endian array on any machine reading the file. The
      types must be described in terms of their actual sizes. For
      example, if a machine with a 64-bit C "long int" writes out an
      array with "long ints", a reading machine with 32-bit C "long
      ints" will yield an array with 64-bit integers.

    - Be reverse engineered. Datasets often live longer than the
      programs that created them. A competent developer should be
      able to create a solution in his preferred programming language to
      read most NPY files that he has been given without much

    - Allow memory-mapping of the data.

    - Be read from a filelike stream object instead of an actual file.
      This allows the implementation to be tested easily and makes the
      system more flexible. NPY files can be stored in ZIP files and
      easily read from a ZipFile object.

    - Store object arrays. Since general Python objects are
      complicated and can only be reliably serialized by pickle (if at
      all), many of the other requirements are waived for files
      containing object arrays. Files with object arrays do not have
      to be mmapable since that would be technically impossible. We
      cannot expect the pickle format to be reverse engineered without
      knowledge of pickle. However, one should at least be able to
      read and write object arrays with the same generic interface as
      other arrays.

    - Be read and written using APIs provided in the numpy package
      itself without any other libraries. The implementation inside
      numpy may be in C if necessary.

    The format explicitly *does not* need to:

    - Support multiple arrays in a file. Since we require filelike
      objects to be supported, one could use the API to build an ad
      hoc format that supported multiple arrays. However, solving the
      general problem and use cases is beyond the scope of the format
      and the API for numpy.

    - Fully handle arbitrary subclasses of numpy.ndarray. Subclasses
      will be accepted for writing, but only the array data will be
      written out. A regular numpy.ndarray object will be created
      upon reading the file. The API can be used to build a format
      for a particular subclass, but that is out of scope for the
      general NPY format.

Format Specification: Version 1.0

    The first 6 bytes are a magic string: exactly "\x93NUMPY".

    The next 1 byte is an unsigned byte: the major version number of
    the file format, e.g. \x01.

    The next 1 byte is an unsigned byte: the minor version number of
    the file format, e.g. \x00. Note: the version of the file format
    is not tied to the version of the numpy package.

    The next 2 bytes form a little-endian unsigned short int: the
    length of the header data HEADER_LEN.

    The next HEADER_LEN bytes form the header data describing the
    array's format. It is an ASCII string which contains a Python
    literal expression of a dictionary. It is terminated by a newline
    ('\n') and padded with spaces ('\x20') to make the total length of
    the magic string + 4 + HEADER_LEN be evenly divisible by 16 for
    alignment purposes.

    The dictionary contains three keys:

        "descr" : dtype.descr
            An object that can be passed as an argument to the
            numpy.dtype() constructor to create the array's dtype.

        "fortran_order" : bool
            Whether the array data is Fortran-contiguous or not.
            Since Fortran-contiguous arrays are a common form of
            non-C-contiguity, we allow them to be written directly to
            disk for efficiency.

        "shape" : tuple of int
            The shape of the array.

    For repeatability and readability, this dictionary is formatted
    using pprint.pformat() so the keys are in alphabetic order.

    Following the header comes the array data. If the dtype contains
    Python objects (i.e. dtype.hasobject is True), then the data is
    a Python pickle of the array. Otherwise the data is the
    contiguous (either C- or Fortran-, depending on fortran_order)
    bytes of the array. Consumers can figure out the number of bytes
    by multiplying the number of elements given by the shape (noting
    that shape=() means there is 1 element) by dtype.itemsize.


    We recommend using the ".npy" extension for files following this
    format. This is by no means a requirement; applications may wish
    to use this file format but use an extension specific to the
    application. In the absence of an obvious alternative, however,
    we suggest using ".npy".

    For a simple way to combine multiple arrays into a single file,
    one can use ZipFile to contain multiple ".npy" files. We
    recommend using the file extension ".npz" for these archives.


    The author believes that this system (or one along these lines) is
    about the simplest system that satisfies all of the requirements.
    However, one must always be wary of introducing a new binary
    format to the world.

    HDF5 [2] is a very flexible format that should be able to
    represent all of NumPy's arrays in some fashion. It is probably
    the only widely-used format that can faithfully represent all of
    NumPy's array features. It has seen substantial adoption by the
    scientific community in general and the NumPy community in
    particular. It is an excellent solution for a wide variety of
    array storage problems with or without NumPy.

    HDF5 is a complicated format that more or less implements
    a hierarchical filesystem-in-a-file. This fact makes satisfying
    some of the Requirements difficult. To the author's knowledge, as
    of this writing, there is no application or library that reads or
    writes even a subset of HDF5 files that does not use the canonical
    libhdf5 implementation. This implementation is a large library
    that is not always easy to build. It would be infeasible to
    include it in numpy.

    It might be feasible to target an extremely limited subset of
    HDF5. Namely, there would be only one object in it: the array.
    Using contiguous storage for the data, one should be able to
    implement just enough of the format to provide the same metadata
    that the proposed format does. One could still meet all of the
    technical requirements like mmapability.

    We would accrue a substantial benefit by being able to generate
    files that could be read by other HDF5 software. Furthermore, by
    providing the first non-libhdf5 implementation of HDF5, we would
    be able to encourage more adoption of simple HDF5 in applications
    where it was previously infeasible because of the size of the
    library. The basic work may encourage similar dead-simple
    implementations in other languages and further expand the

    The remaining concern is about reverse engineerability of the
    format. Even the simple subset of HDF5 would be very difficult to
    reverse engineer given just a file by itself. However, given the
    prominence of HDF5, this might not be a substantial concern.

    In conclusion, we are going forward with the design laid out in
    this document. If someone writes code to handle the simple subset
    of HDF5 that would be useful to us, we may consider a revision of
    the file format.


    The current implementation is included in the 1.0.5 release of numpy.

    Specifically, the file in this directory implements the
    format as described here.





    This document has been placed in the public domain.

Local Variables:
mode: indented-text
indent-tabs-mode: nil
sentence-end-double-space: t
fill-column: 70
coding: utf-8
Something went wrong with that request. Please try again.