[Numpy-discussion] numpy fileIO
Thu Oct 16 11:33:22 CDT 2008
A Thursday 16 October 2008, Prashant Saxena escrigué:
> I have never used numpy in my python applications until now. I am
> writing a python/openGL based tool to manipulate 3d geometrical
> data(meshes, curve, points etc.) I would be using numpy to store
> these objects(vertices, edges, faces, index etc.) at run time. One of
> my major concern is numpy's file IO capabilities. Geometrical objects
> would be stored in a structures, for example a logical structure to
> store a mesh goes like this:
> vertex(numpy.float16)[x, y, z]
> normal(numpy.float16)[x, y, z]
> st(2d numpy.float16)[u, v]
> There would be different structures for curve, patch, points and rest
> of the primitives. I am sure numpy users must have encounered similar
> scenerio where you need to dump this data to a file and read it back.
> In my case, a general assumption of nearly 50-150 megs of data would
> be considered as normal size. Before I go deep into coding It would
> be great if numpy user can share their expreience for the task.
> I am also open for unconventional or off the route methods, provided
> they can do the job.(C/c++, 3rd party modules etc.) Here is the
> summery of IO operations I would be working on:
> 1. Write different structures to a file.
> 2. Read data back from file.
> 3. if structure can be tagged(int or string) then read a particular
> structure using tag, from file.
Your structure seems like can be expressed through a record array .
If that is the case, just try the included tofile()/fromfile()
utilities included in numpy for saving/retrieving them.
If you need more functionality for recarray I/O, you can have a look at
pytables , which allows you to access directly the recarrays on disk
row by row (even if compressed) or use complex expressions to quickly
select interesting parts, among other niceties.
If your data structures can't be fit in recarray structures directly,
you may want to try to decompose them in smaller constructs and relate
them in file through pointers (indices to other constructs).
Hope that helps,
More information about the Numpy-discussion