A Simple File Format for NumPy Arrays

Author: Robert Kern <robert.kern@gmail.com> Status: Draft Created: 20-Dec-2007

Abstract

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.

Rationale

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 save.
  • 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.

Requirements

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 documentation.
  • 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.

Format Specification: Version 2.0

The version 1.0 format only allowed the array header to have a total size of 65535 bytes. This can be exceeded by structured arrays with a large number of columns. The version 2.0 format extends the header size to 4 GiB. numpy.save will automatically save in 2.0 format if the data requires it, else it will always use the more compatible 1.0 format.

The description of the fourth element of the header therefore has become:

The next 4 bytes form a little-endian unsigned int: the length of the header data HEADER_LEN.

Conventions

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.

Alternatives

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 community.

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.

Implementation

The version 1.0 implementation was first included in the 1.0.5 release of numpy, and remains available. The version 2.0 implementation was first included in the 1.9.0 release of numpy.

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